2 * Copyright (c) 2009-2017 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "classifier.h"
19 #include "classifier-private.h"
21 #include <sys/types.h>
22 #include <netinet/in.h>
23 #include "byte-order.h"
24 #include "openvswitch/dynamic-string.h"
31 /* A collection of "struct cls_conjunction"s currently embedded into a
33 struct cls_conjunction_set
{
34 /* Link back to the cls_match.
36 * cls_conjunction_set is mostly used during classifier lookup, and, in
37 * turn, during classifier lookup the most used member of
38 * cls_conjunction_set is the rule's priority, so we cache it here for fast
40 struct cls_match
*match
;
41 int priority
; /* Cached copy of match->priority. */
43 /* Conjunction information.
45 * 'min_n_clauses' allows some optimization during classifier lookup. */
46 unsigned int n
; /* Number of elements in 'conj'. */
47 unsigned int min_n_clauses
; /* Smallest 'n' among elements of 'conj'. */
48 struct cls_conjunction conj
[];
51 /* Ports trie depends on both ports sharing the same ovs_be32. */
52 #define TP_PORTS_OFS32 (offsetof(struct flow, tp_src) / 4)
53 BUILD_ASSERT_DECL(TP_PORTS_OFS32
== offsetof(struct flow
, tp_dst
) / 4);
54 BUILD_ASSERT_DECL(TP_PORTS_OFS32
% 2 == 0);
55 #define TP_PORTS_OFS64 (TP_PORTS_OFS32 / 2)
58 cls_conjunction_set_size(size_t n
)
60 return (sizeof(struct cls_conjunction_set
)
61 + n
* sizeof(struct cls_conjunction
));
64 static struct cls_conjunction_set
*
65 cls_conjunction_set_alloc(struct cls_match
*match
,
66 const struct cls_conjunction conj
[], size_t n
)
69 size_t min_n_clauses
= conj
[0].n_clauses
;
70 for (size_t i
= 1; i
< n
; i
++) {
71 min_n_clauses
= MIN(min_n_clauses
, conj
[i
].n_clauses
);
74 struct cls_conjunction_set
*set
= xmalloc(cls_conjunction_set_size(n
));
76 set
->priority
= match
->priority
;
78 set
->min_n_clauses
= min_n_clauses
;
79 memcpy(set
->conj
, conj
, n
* sizeof *conj
);
86 static struct cls_match
*
87 cls_match_alloc(const struct cls_rule
*rule
, ovs_version_t version
,
88 const struct cls_conjunction conj
[], size_t n
)
90 size_t count
= miniflow_n_values(rule
->match
.flow
);
92 struct cls_match
*cls_match
93 = xmalloc(sizeof *cls_match
+ MINIFLOW_VALUES_SIZE(count
));
95 ovsrcu_init(&cls_match
->next
, NULL
);
96 *CONST_CAST(const struct cls_rule
**, &cls_match
->cls_rule
) = rule
;
97 *CONST_CAST(int *, &cls_match
->priority
) = rule
->priority
;
98 /* Make rule initially invisible. */
99 cls_match
->versions
= VERSIONS_INITIALIZER(version
, version
);
100 miniflow_clone(CONST_CAST(struct miniflow
*, &cls_match
->flow
),
101 rule
->match
.flow
, count
);
102 ovsrcu_set_hidden(&cls_match
->conj_set
,
103 cls_conjunction_set_alloc(cls_match
, conj
, n
));
108 static struct cls_subtable
*find_subtable(const struct classifier
*cls
,
109 const struct minimask
*);
110 static struct cls_subtable
*insert_subtable(struct classifier
*cls
,
111 const struct minimask
*);
112 static void destroy_subtable(struct classifier
*cls
, struct cls_subtable
*);
114 static const struct cls_match
*find_match_wc(const struct cls_subtable
*,
115 ovs_version_t version
,
118 unsigned int n_tries
,
119 struct flow_wildcards
*);
120 static struct cls_match
*find_equal(const struct cls_subtable
*,
121 const struct miniflow
*, uint32_t hash
);
123 /* Return the next visible (lower-priority) rule in the list. Multiple
124 * identical rules with the same priority may exist transitionally, but when
125 * versioning is used at most one of them is ever visible for lookups on any
126 * given 'version'. */
127 static inline const struct cls_match
*
128 next_visible_rule_in_list(const struct cls_match
*rule
, ovs_version_t version
)
131 rule
= cls_match_next(rule
);
132 } while (rule
&& !cls_match_visible_in_version(rule
, version
));
137 /* Type with maximum supported prefix length. */
139 struct in6_addr ipv6
; /* For sizing. */
140 ovs_be32 be32
; /* For access. */
143 static unsigned int minimask_get_prefix_len(const struct minimask
*,
144 const struct mf_field
*);
145 static void trie_init(struct classifier
*cls
, int trie_idx
,
146 const struct mf_field
*);
147 static unsigned int trie_lookup(const struct cls_trie
*, const struct flow
*,
148 union trie_prefix
*plens
);
149 static unsigned int trie_lookup_value(const rcu_trie_ptr
*,
150 const ovs_be32 value
[], ovs_be32 plens
[],
151 unsigned int value_bits
);
152 static void trie_destroy(rcu_trie_ptr
*);
153 static void trie_insert(struct cls_trie
*, const struct cls_rule
*, int mlen
);
154 static void trie_insert_prefix(rcu_trie_ptr
*, const ovs_be32
*prefix
,
156 static void trie_remove(struct cls_trie
*, const struct cls_rule
*, int mlen
);
157 static void trie_remove_prefix(rcu_trie_ptr
*, const ovs_be32
*prefix
,
159 static void mask_set_prefix_bits(struct flow_wildcards
*, uint8_t be32ofs
,
160 unsigned int n_bits
);
161 static bool mask_prefix_bits_set(const struct flow_wildcards
*,
162 uint8_t be32ofs
, unsigned int n_bits
);
167 cls_rule_init__(struct cls_rule
*rule
, unsigned int priority
)
169 rculist_init(&rule
->node
);
170 *CONST_CAST(int *, &rule
->priority
) = priority
;
171 ovsrcu_init(&rule
->cls_match
, NULL
);
174 /* Initializes 'rule' to match packets specified by 'match' at the given
175 * 'priority'. 'match' must satisfy the invariant described in the comment at
176 * the definition of struct match.
178 * The caller must eventually destroy 'rule' with cls_rule_destroy().
180 * Clients should not use priority INT_MIN. (OpenFlow uses priorities between
181 * 0 and UINT16_MAX, inclusive.) */
183 cls_rule_init(struct cls_rule
*rule
, const struct match
*match
, int priority
)
185 cls_rule_init__(rule
, priority
);
186 minimatch_init(CONST_CAST(struct minimatch
*, &rule
->match
), match
);
189 /* Same as cls_rule_init() for initialization from a "struct minimatch". */
191 cls_rule_init_from_minimatch(struct cls_rule
*rule
,
192 const struct minimatch
*match
, int priority
)
194 cls_rule_init__(rule
, priority
);
195 minimatch_clone(CONST_CAST(struct minimatch
*, &rule
->match
), match
);
198 /* Initializes 'dst' as a copy of 'src'.
200 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
202 cls_rule_clone(struct cls_rule
*dst
, const struct cls_rule
*src
)
204 cls_rule_init__(dst
, src
->priority
);
205 minimatch_clone(CONST_CAST(struct minimatch
*, &dst
->match
), &src
->match
);
208 /* Initializes 'dst' with the data in 'src', destroying 'src'.
210 * 'src' must be a cls_rule NOT in a classifier.
212 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
214 cls_rule_move(struct cls_rule
*dst
, struct cls_rule
*src
)
216 cls_rule_init__(dst
, src
->priority
);
217 minimatch_move(CONST_CAST(struct minimatch
*, &dst
->match
),
218 CONST_CAST(struct minimatch
*, &src
->match
));
221 /* Frees memory referenced by 'rule'. Doesn't free 'rule' itself (it's
222 * normally embedded into a larger structure).
224 * ('rule' must not currently be in a classifier.) */
226 cls_rule_destroy(struct cls_rule
*rule
)
227 OVS_NO_THREAD_SAFETY_ANALYSIS
229 /* Must not be in a classifier. */
230 ovs_assert(!get_cls_match_protected(rule
));
232 /* Check that the rule has been properly removed from the classifier. */
233 ovs_assert(rule
->node
.prev
== RCULIST_POISON
234 || rculist_is_empty(&rule
->node
));
235 rculist_poison__(&rule
->node
); /* Poisons also the next pointer. */
237 minimatch_destroy(CONST_CAST(struct minimatch
*, &rule
->match
));
240 /* This may only be called by the exclusive writer. */
242 cls_rule_set_conjunctions(struct cls_rule
*cr
,
243 const struct cls_conjunction
*conj
, size_t n
)
245 struct cls_match
*match
= get_cls_match_protected(cr
);
246 struct cls_conjunction_set
*old
247 = ovsrcu_get_protected(struct cls_conjunction_set
*, &match
->conj_set
);
248 struct cls_conjunction
*old_conj
= old
? old
->conj
: NULL
;
249 unsigned int old_n
= old
? old
->n
: 0;
251 if (old_n
!= n
|| (n
&& memcmp(old_conj
, conj
, n
* sizeof *conj
))) {
253 ovsrcu_postpone(free
, old
);
255 ovsrcu_set(&match
->conj_set
,
256 cls_conjunction_set_alloc(match
, conj
, n
));
261 /* Returns true if 'a' and 'b' match the same packets at the same priority,
262 * false if they differ in some way. */
264 cls_rule_equal(const struct cls_rule
*a
, const struct cls_rule
*b
)
266 return a
->priority
== b
->priority
&& minimatch_equal(&a
->match
, &b
->match
);
269 /* Appends a string describing 'rule' to 's'. */
271 cls_rule_format(const struct cls_rule
*rule
, const struct tun_table
*tun_table
,
272 const struct ofputil_port_map
*port_map
, struct ds
*s
)
274 minimatch_format(&rule
->match
, tun_table
, port_map
, s
, rule
->priority
);
277 /* Returns true if 'rule' matches every packet, false otherwise. */
279 cls_rule_is_catchall(const struct cls_rule
*rule
)
281 return minimask_is_catchall(rule
->match
.mask
);
284 /* Makes 'rule' invisible in 'remove_version'. Once that version is used in
285 * lookups, the caller should remove 'rule' via ovsrcu_postpone().
287 * 'rule' must be in a classifier.
288 * This may only be called by the exclusive writer. */
290 cls_rule_make_invisible_in_version(const struct cls_rule
*rule
,
291 ovs_version_t remove_version
)
293 struct cls_match
*cls_match
= get_cls_match_protected(rule
);
295 ovs_assert(remove_version
>= cls_match
->versions
.add_version
);
297 cls_match_set_remove_version(cls_match
, remove_version
);
300 /* This undoes the change made by cls_rule_make_invisible_in_version().
302 * 'rule' must be in a classifier.
303 * This may only be called by the exclusive writer. */
305 cls_rule_restore_visibility(const struct cls_rule
*rule
)
307 cls_match_set_remove_version(get_cls_match_protected(rule
),
308 OVS_VERSION_NOT_REMOVED
);
311 /* Return true if 'rule' is visible in 'version'.
313 * 'rule' must be in a classifier. */
315 cls_rule_visible_in_version(const struct cls_rule
*rule
, ovs_version_t version
)
317 struct cls_match
*cls_match
= get_cls_match(rule
);
319 return cls_match
&& cls_match_visible_in_version(cls_match
, version
);
322 /* Initializes 'cls' as a classifier that initially contains no classification
325 classifier_init(struct classifier
*cls
, const uint8_t *flow_segments
)
328 cmap_init(&cls
->subtables_map
);
329 pvector_init(&cls
->subtables
);
330 cls
->n_flow_segments
= 0;
332 while (cls
->n_flow_segments
< CLS_MAX_INDICES
333 && *flow_segments
< FLOW_U64S
) {
334 cls
->flow_segments
[cls
->n_flow_segments
++] = *flow_segments
++;
338 for (int i
= 0; i
< CLS_MAX_TRIES
; i
++) {
339 trie_init(cls
, i
, NULL
);
344 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
345 * caller's responsibility.
346 * May only be called after all the readers have been terminated. */
348 classifier_destroy(struct classifier
*cls
)
351 struct cls_subtable
*subtable
;
354 for (i
= 0; i
< cls
->n_tries
; i
++) {
355 trie_destroy(&cls
->tries
[i
].root
);
358 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
359 destroy_subtable(cls
, subtable
);
361 cmap_destroy(&cls
->subtables_map
);
363 pvector_destroy(&cls
->subtables
);
367 /* Set the fields for which prefix lookup should be performed. */
369 classifier_set_prefix_fields(struct classifier
*cls
,
370 const enum mf_field_id
*trie_fields
,
371 unsigned int n_fields
)
373 const struct mf_field
* new_fields
[CLS_MAX_TRIES
];
374 struct mf_bitmap fields
= MF_BITMAP_INITIALIZER
;
376 bool changed
= false;
378 for (i
= 0; i
< n_fields
&& n_tries
< CLS_MAX_TRIES
; i
++) {
379 const struct mf_field
*field
= mf_from_id(trie_fields
[i
]);
380 if (field
->flow_be32ofs
< 0 || field
->n_bits
% 32) {
381 /* Incompatible field. This is the only place where we
382 * enforce these requirements, but the rest of the trie code
383 * depends on the flow_be32ofs to be non-negative and the
384 * field length to be a multiple of 32 bits. */
388 if (bitmap_is_set(fields
.bm
, trie_fields
[i
])) {
389 /* Duplicate field, there is no need to build more than
390 * one index for any one field. */
393 bitmap_set1(fields
.bm
, trie_fields
[i
]);
395 new_fields
[n_tries
] = NULL
;
396 const struct mf_field
*cls_field
397 = ovsrcu_get(struct mf_field
*, &cls
->tries
[n_tries
].field
);
398 if (n_tries
>= cls
->n_tries
|| field
!= cls_field
) {
399 new_fields
[n_tries
] = field
;
405 if (changed
|| n_tries
< cls
->n_tries
) {
406 struct cls_subtable
*subtable
;
408 /* Trie configuration needs to change. Disable trie lookups
409 * for the tries that are changing and wait all the current readers
410 * with the old configuration to be done. */
412 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
413 for (i
= 0; i
< cls
->n_tries
; i
++) {
414 if ((i
< n_tries
&& new_fields
[i
]) || i
>= n_tries
) {
415 if (subtable
->trie_plen
[i
]) {
416 subtable
->trie_plen
[i
] = 0;
422 /* Synchronize if any readers were using tries. The readers may
423 * temporarily function without the trie lookup based optimizations. */
425 /* ovsrcu_synchronize() functions as a memory barrier, so it does
426 * not matter that subtable->trie_plen is not atomic. */
427 ovsrcu_synchronize();
430 /* Now set up the tries. */
431 for (i
= 0; i
< n_tries
; i
++) {
433 trie_init(cls
, i
, new_fields
[i
]);
436 /* Destroy the rest, if any. */
437 for (; i
< cls
->n_tries
; i
++) {
438 trie_init(cls
, i
, NULL
);
441 cls
->n_tries
= n_tries
;
445 return false; /* No change. */
449 trie_init(struct classifier
*cls
, int trie_idx
, const struct mf_field
*field
)
451 struct cls_trie
*trie
= &cls
->tries
[trie_idx
];
452 struct cls_subtable
*subtable
;
454 if (trie_idx
< cls
->n_tries
) {
455 trie_destroy(&trie
->root
);
457 ovsrcu_set_hidden(&trie
->root
, NULL
);
459 ovsrcu_set_hidden(&trie
->field
, CONST_CAST(struct mf_field
*, field
));
461 /* Add existing rules to the new trie. */
462 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
465 plen
= field
? minimask_get_prefix_len(&subtable
->mask
, field
) : 0;
467 struct cls_match
*head
;
469 CMAP_FOR_EACH (head
, cmap_node
, &subtable
->rules
) {
470 trie_insert(trie
, head
->cls_rule
, plen
);
473 /* Initialize subtable's prefix length on this field. This will
474 * allow readers to use the trie. */
475 atomic_thread_fence(memory_order_release
);
476 subtable
->trie_plen
[trie_idx
] = plen
;
480 /* Returns true if 'cls' contains no classification rules, false otherwise.
481 * Checking the cmap requires no locking. */
483 classifier_is_empty(const struct classifier
*cls
)
485 return cmap_is_empty(&cls
->subtables_map
);
488 /* Returns the number of rules in 'cls'. */
490 classifier_count(const struct classifier
*cls
)
492 /* n_rules is an int, so in the presence of concurrent writers this will
493 * return either the old or a new value. */
497 static inline ovs_be32
minimatch_get_ports(const struct minimatch
*match
)
499 /* Could optimize to use the same map if needed for fast path. */
500 return (miniflow_get_ports(match
->flow
)
501 & miniflow_get_ports(&match
->mask
->masks
));
504 /* Inserts 'rule' into 'cls' in 'version'. Until 'rule' is removed from 'cls',
505 * the caller must not modify or free it.
507 * If 'cls' already contains an identical rule (including wildcards, values of
508 * fixed fields, and priority) that is visible in 'version', replaces the old
509 * rule by 'rule' and returns the rule that was replaced. The caller takes
510 * ownership of the returned rule and is thus responsible for destroying it
511 * with cls_rule_destroy(), after RCU grace period has passed (see
512 * ovsrcu_postpone()).
514 * Returns NULL if 'cls' does not contain a rule with an identical key, after
515 * inserting the new rule. In this case, no rules are displaced by the new
516 * rule, even rules that cannot have any effect because the new rule matches a
517 * superset of their flows and has higher priority.
519 const struct cls_rule
*
520 classifier_replace(struct classifier
*cls
, const struct cls_rule
*rule
,
521 ovs_version_t version
,
522 const struct cls_conjunction
*conjs
, size_t n_conjs
)
524 struct cls_match
*new;
525 struct cls_subtable
*subtable
;
526 uint32_t ihash
[CLS_MAX_INDICES
];
527 struct cls_match
*head
;
528 unsigned int mask_offset
;
534 /* 'new' is initially invisible to lookups. */
535 new = cls_match_alloc(rule
, version
, conjs
, n_conjs
);
536 ovsrcu_set(&CONST_CAST(struct cls_rule
*, rule
)->cls_match
, new);
538 subtable
= find_subtable(cls
, rule
->match
.mask
);
540 subtable
= insert_subtable(cls
, rule
->match
.mask
);
543 /* Compute hashes in segments. */
546 for (i
= 0; i
< subtable
->n_indices
; i
++) {
547 ihash
[i
] = minimatch_hash_range(&rule
->match
, subtable
->index_maps
[i
],
548 &mask_offset
, &basis
);
550 hash
= minimatch_hash_range(&rule
->match
, subtable
->index_maps
[i
],
551 &mask_offset
, &basis
);
553 head
= find_equal(subtable
, rule
->match
.flow
, hash
);
555 /* Add rule to tries.
557 * Concurrent readers might miss seeing the rule until this update,
558 * which might require being fixed up by revalidation later. */
559 for (i
= 0; i
< cls
->n_tries
; i
++) {
560 if (subtable
->trie_plen
[i
]) {
561 trie_insert(&cls
->tries
[i
], rule
, subtable
->trie_plen
[i
]);
565 /* Add rule to ports trie. */
566 if (subtable
->ports_mask_len
) {
567 /* We mask the value to be inserted to always have the wildcarded
568 * bits in known (zero) state, so we can include them in comparison
569 * and they will always match (== their original value does not
571 ovs_be32 masked_ports
= minimatch_get_ports(&rule
->match
);
573 trie_insert_prefix(&subtable
->ports_trie
, &masked_ports
,
574 subtable
->ports_mask_len
);
577 /* Add new node to segment indices. */
578 for (i
= 0; i
< subtable
->n_indices
; i
++) {
579 ccmap_inc(&subtable
->indices
[i
], ihash
[i
]);
581 n_rules
= cmap_insert(&subtable
->rules
, &new->cmap_node
, hash
);
582 } else { /* Equal rules exist in the classifier already. */
583 struct cls_match
*prev
, *iter
;
585 /* Scan the list for the insertion point that will keep the list in
586 * order of decreasing priority. Insert after rules marked invisible
587 * in any version of the same priority. */
588 FOR_EACH_RULE_IN_LIST_PROTECTED (iter
, prev
, head
) {
589 if (rule
->priority
> iter
->priority
590 || (rule
->priority
== iter
->priority
591 && !cls_match_is_eventually_invisible(iter
))) {
596 /* Replace 'iter' with 'new' or insert 'new' between 'prev' and
599 struct cls_rule
*old
;
601 if (rule
->priority
== iter
->priority
) {
602 cls_match_replace(prev
, iter
, new);
603 old
= CONST_CAST(struct cls_rule
*, iter
->cls_rule
);
605 cls_match_insert(prev
, iter
, new);
609 /* Replace the existing head in data structures, if rule is the new
612 cmap_replace(&subtable
->rules
, &head
->cmap_node
,
613 &new->cmap_node
, hash
);
617 struct cls_conjunction_set
*conj_set
;
619 conj_set
= ovsrcu_get_protected(struct cls_conjunction_set
*,
622 ovsrcu_postpone(free
, conj_set
);
625 ovsrcu_set(&old
->cls_match
, NULL
); /* Marks old rule as removed
626 * from the classifier. */
627 ovsrcu_postpone(cls_match_free_cb
, iter
);
629 /* No change in subtable's max priority or max count. */
631 /* Make 'new' visible to lookups in the appropriate version. */
632 cls_match_set_remove_version(new, OVS_VERSION_NOT_REMOVED
);
634 /* Make rule visible to iterators (immediately). */
635 rculist_replace(CONST_CAST(struct rculist
*, &rule
->node
),
638 /* Return displaced rule. Caller is responsible for keeping it
639 * around until all threads quiesce. */
643 /* 'new' is new node after 'prev' */
644 cls_match_insert(prev
, iter
, new);
648 /* Make 'new' visible to lookups in the appropriate version. */
649 cls_match_set_remove_version(new, OVS_VERSION_NOT_REMOVED
);
651 /* Make rule visible to iterators (immediately). */
652 rculist_push_back(&subtable
->rules_list
,
653 CONST_CAST(struct rculist
*, &rule
->node
));
655 /* Rule was added, not replaced. Update 'subtable's 'max_priority' and
656 * 'max_count', if necessary.
658 * The rule was already inserted, but concurrent readers may not see the
659 * rule yet as the subtables vector is not updated yet. This will have to
660 * be fixed by revalidation later. */
662 subtable
->max_priority
= rule
->priority
;
663 subtable
->max_count
= 1;
664 pvector_insert(&cls
->subtables
, subtable
, rule
->priority
);
665 } else if (rule
->priority
== subtable
->max_priority
) {
666 ++subtable
->max_count
;
667 } else if (rule
->priority
> subtable
->max_priority
) {
668 subtable
->max_priority
= rule
->priority
;
669 subtable
->max_count
= 1;
670 pvector_change_priority(&cls
->subtables
, subtable
, rule
->priority
);
673 /* Nothing was replaced. */
677 pvector_publish(&cls
->subtables
);
683 /* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller
684 * must not modify or free it.
686 * 'cls' must not contain an identical rule (including wildcards, values of
687 * fixed fields, and priority). Use classifier_find_rule_exactly() to find
690 classifier_insert(struct classifier
*cls
, const struct cls_rule
*rule
,
691 ovs_version_t version
, const struct cls_conjunction conj
[],
694 const struct cls_rule
*displaced_rule
695 = classifier_replace(cls
, rule
, version
, conj
, n_conj
);
696 ovs_assert(!displaced_rule
);
699 /* If 'rule' is in 'cls', removes 'rule' from 'cls' and returns true. It is
700 * the caller's responsibility to destroy 'rule' with cls_rule_destroy(),
701 * freeing the memory block in which 'rule' resides, etc., as necessary.
703 * If 'rule' is not in any classifier, returns false without making any
706 * 'rule' must not be in some classifier other than 'cls'.
709 classifier_remove(struct classifier
*cls
, const struct cls_rule
*cls_rule
)
711 struct cls_match
*rule
, *prev
, *next
, *head
;
712 struct cls_conjunction_set
*conj_set
;
713 struct cls_subtable
*subtable
;
714 uint32_t basis
= 0, hash
, ihash
[CLS_MAX_INDICES
];
715 unsigned int mask_offset
;
719 rule
= get_cls_match_protected(cls_rule
);
723 /* Mark as removed. */
724 ovsrcu_set(&CONST_CAST(struct cls_rule
*, cls_rule
)->cls_match
, NULL
);
726 /* Remove 'cls_rule' from the subtable's rules list. */
727 rculist_remove(CONST_CAST(struct rculist
*, &cls_rule
->node
));
729 subtable
= find_subtable(cls
, cls_rule
->match
.mask
);
730 ovs_assert(subtable
);
733 for (i
= 0; i
< subtable
->n_indices
; i
++) {
734 ihash
[i
] = minimatch_hash_range(&cls_rule
->match
,
735 subtable
->index_maps
[i
],
736 &mask_offset
, &basis
);
738 hash
= minimatch_hash_range(&cls_rule
->match
, subtable
->index_maps
[i
],
739 &mask_offset
, &basis
);
741 head
= find_equal(subtable
, cls_rule
->match
.flow
, hash
);
743 /* Check if the rule is not the head rule. */
745 struct cls_match
*iter
;
747 /* Not the head rule, but potentially one with the same priority. */
748 /* Remove from the list of equal rules. */
749 FOR_EACH_RULE_IN_LIST_PROTECTED (iter
, prev
, head
) {
754 ovs_assert(iter
== rule
);
756 cls_match_remove(prev
, rule
);
761 /* 'rule' is the head rule. Check if there is another rule to
762 * replace 'rule' in the data structures. */
763 next
= cls_match_next_protected(rule
);
765 cmap_replace(&subtable
->rules
, &rule
->cmap_node
, &next
->cmap_node
,
770 /* 'rule' is last of the kind in the classifier, must remove from all the
771 * data structures. */
773 if (subtable
->ports_mask_len
) {
774 ovs_be32 masked_ports
= minimatch_get_ports(&cls_rule
->match
);
776 trie_remove_prefix(&subtable
->ports_trie
,
777 &masked_ports
, subtable
->ports_mask_len
);
779 for (i
= 0; i
< cls
->n_tries
; i
++) {
780 if (subtable
->trie_plen
[i
]) {
781 trie_remove(&cls
->tries
[i
], cls_rule
, subtable
->trie_plen
[i
]);
785 /* Remove rule node from indices. */
786 for (i
= 0; i
< subtable
->n_indices
; i
++) {
787 ccmap_dec(&subtable
->indices
[i
], ihash
[i
]);
789 n_rules
= cmap_remove(&subtable
->rules
, &rule
->cmap_node
, hash
);
792 destroy_subtable(cls
, subtable
);
795 if (subtable
->max_priority
== rule
->priority
796 && --subtable
->max_count
== 0) {
797 /* Find the new 'max_priority' and 'max_count'. */
798 int max_priority
= INT_MIN
;
799 CMAP_FOR_EACH (head
, cmap_node
, &subtable
->rules
) {
800 if (head
->priority
> max_priority
) {
801 max_priority
= head
->priority
;
802 subtable
->max_count
= 1;
803 } else if (head
->priority
== max_priority
) {
804 ++subtable
->max_count
;
807 subtable
->max_priority
= max_priority
;
808 pvector_change_priority(&cls
->subtables
, subtable
, max_priority
);
813 pvector_publish(&cls
->subtables
);
817 conj_set
= ovsrcu_get_protected(struct cls_conjunction_set
*,
820 ovsrcu_postpone(free
, conj_set
);
822 ovsrcu_postpone(cls_match_free_cb
, rule
);
829 classifier_remove_assert(struct classifier
*cls
,
830 const struct cls_rule
*cls_rule
)
832 ovs_assert(classifier_remove(cls
, cls_rule
));
835 /* Prefix tree context. Valid when 'lookup_done' is true. Can skip all
836 * subtables which have a prefix match on the trie field, but whose prefix
837 * length is not indicated in 'match_plens'. For example, a subtable that
838 * has a 8-bit trie field prefix match can be skipped if
839 * !be_get_bit_at(&match_plens, 8 - 1). If skipped, 'maskbits' prefix bits
840 * must be unwildcarded to make datapath flow only match packets it should. */
842 const struct cls_trie
*trie
;
843 bool lookup_done
; /* Status of the lookup. */
844 unsigned int maskbits
; /* Prefix length needed to avoid false matches. */
845 union trie_prefix match_plens
; /* Bitmask of prefix lengths with possible
850 trie_ctx_init(struct trie_ctx
*ctx
, const struct cls_trie
*trie
)
853 ctx
->lookup_done
= false;
856 struct conjunctive_match
{
857 struct hmap_node hmap_node
;
862 static struct conjunctive_match
*
863 find_conjunctive_match__(struct hmap
*matches
, uint64_t id
, uint32_t hash
)
865 struct conjunctive_match
*m
;
867 HMAP_FOR_EACH_IN_BUCKET (m
, hmap_node
, hash
, matches
) {
876 find_conjunctive_match(const struct cls_conjunction_set
*set
,
877 unsigned int max_n_clauses
, struct hmap
*matches
,
878 struct conjunctive_match
*cm_stubs
, size_t n_cm_stubs
,
881 const struct cls_conjunction
*c
;
883 if (max_n_clauses
< set
->min_n_clauses
) {
887 for (c
= set
->conj
; c
< &set
->conj
[set
->n
]; c
++) {
888 struct conjunctive_match
*cm
;
891 if (c
->n_clauses
> max_n_clauses
) {
895 hash
= hash_int(c
->id
, 0);
896 cm
= find_conjunctive_match__(matches
, c
->id
, hash
);
898 size_t n
= hmap_count(matches
);
900 cm
= n
< n_cm_stubs
? &cm_stubs
[n
] : xmalloc(sizeof *cm
);
901 hmap_insert(matches
, &cm
->hmap_node
, hash
);
903 cm
->clauses
= UINT64_MAX
<< (c
->n_clauses
& 63);
905 cm
->clauses
|= UINT64_C(1) << c
->clause
;
906 if (cm
->clauses
== UINT64_MAX
) {
915 free_conjunctive_matches(struct hmap
*matches
,
916 struct conjunctive_match
*cm_stubs
, size_t n_cm_stubs
)
918 if (hmap_count(matches
) > n_cm_stubs
) {
919 struct conjunctive_match
*cm
, *next
;
921 HMAP_FOR_EACH_SAFE (cm
, next
, hmap_node
, matches
) {
922 if (!(cm
>= cm_stubs
&& cm
< &cm_stubs
[n_cm_stubs
])) {
927 hmap_destroy(matches
);
930 /* Like classifier_lookup(), except that support for conjunctive matches can be
931 * configured with 'allow_conjunctive_matches'. That feature is not exposed
932 * externally because turning off conjunctive matches is only useful to avoid
933 * recursion within this function itself.
935 * 'flow' is non-const to allow for temporary modifications during the lookup.
936 * Any changes are restored before returning. */
937 static const struct cls_rule
*
938 classifier_lookup__(const struct classifier
*cls
, ovs_version_t version
,
939 struct flow
*flow
, struct flow_wildcards
*wc
,
940 bool allow_conjunctive_matches
)
942 struct trie_ctx trie_ctx
[CLS_MAX_TRIES
];
943 const struct cls_match
*match
;
944 /* Highest-priority flow in 'cls' that certainly matches 'flow'. */
945 const struct cls_match
*hard
= NULL
;
946 int hard_pri
= INT_MIN
; /* hard ? hard->priority : INT_MIN. */
948 /* Highest-priority conjunctive flows in 'cls' matching 'flow'. Since
949 * these are (components of) conjunctive flows, we can only know whether
950 * the full conjunctive flow matches after seeing multiple of them. Thus,
951 * we refer to these as "soft matches". */
952 struct cls_conjunction_set
*soft_stub
[64];
953 struct cls_conjunction_set
**soft
= soft_stub
;
954 size_t n_soft
= 0, allocated_soft
= ARRAY_SIZE(soft_stub
);
955 int soft_pri
= INT_MIN
; /* n_soft ? MAX(soft[*]->priority) : INT_MIN. */
957 /* Synchronize for cls->n_tries and subtable->trie_plen. They can change
958 * when table configuration changes, which happens typically only on
960 atomic_thread_fence(memory_order_acquire
);
962 /* Initialize trie contexts for find_match_wc(). */
963 for (int i
= 0; i
< cls
->n_tries
; i
++) {
964 trie_ctx_init(&trie_ctx
[i
], &cls
->tries
[i
]);
968 struct cls_subtable
*subtable
;
969 PVECTOR_FOR_EACH_PRIORITY (subtable
, hard_pri
+ 1, 2, sizeof *subtable
,
971 struct cls_conjunction_set
*conj_set
;
973 /* Skip subtables with no match, or where the match is lower-priority
974 * than some certain match we've already found. */
975 match
= find_match_wc(subtable
, version
, flow
, trie_ctx
, cls
->n_tries
,
977 if (!match
|| match
->priority
<= hard_pri
) {
981 conj_set
= ovsrcu_get(struct cls_conjunction_set
*, &match
->conj_set
);
983 /* 'match' isn't part of a conjunctive match. It's the best
984 * certain match we've got so far, since we know that it's
985 * higher-priority than hard_pri.
987 * (There might be a higher-priority conjunctive match. We can't
990 hard_pri
= hard
->priority
;
991 } else if (allow_conjunctive_matches
) {
992 /* 'match' is part of a conjunctive match. Add it to the list. */
993 if (OVS_UNLIKELY(n_soft
>= allocated_soft
)) {
994 struct cls_conjunction_set
**old_soft
= soft
;
997 soft
= xmalloc(allocated_soft
* sizeof *soft
);
998 memcpy(soft
, old_soft
, n_soft
* sizeof *soft
);
999 if (old_soft
!= soft_stub
) {
1003 soft
[n_soft
++] = conj_set
;
1005 /* Keep track of the highest-priority soft match. */
1006 if (soft_pri
< match
->priority
) {
1007 soft_pri
= match
->priority
;
1012 /* In the common case, at this point we have no soft matches and we can
1013 * return immediately. (We do the same thing if we have potential soft
1014 * matches but none of them are higher-priority than our hard match.) */
1015 if (hard_pri
>= soft_pri
) {
1016 if (soft
!= soft_stub
) {
1019 return hard
? hard
->cls_rule
: NULL
;
1022 /* At this point, we have some soft matches. We might also have a hard
1023 * match; if so, its priority is lower than the highest-priority soft
1028 * Check whether soft matches are real matches. */
1030 /* Delete soft matches that are null. This only happens in second and
1031 * subsequent iterations of the soft match loop, when we drop back from
1032 * a high-priority soft match to a lower-priority one.
1034 * Also, delete soft matches whose priority is less than or equal to
1035 * the hard match's priority. In the first iteration of the soft
1036 * match, these can be in 'soft' because the earlier main loop found
1037 * the soft match before the hard match. In second and later iteration
1038 * of the soft match loop, these can be in 'soft' because we dropped
1039 * back from a high-priority soft match to a lower-priority soft match.
1041 * It is tempting to delete soft matches that cannot be satisfied
1042 * because there are fewer soft matches than required to satisfy any of
1043 * their conjunctions, but we cannot do that because there might be
1044 * lower priority soft or hard matches with otherwise identical
1045 * matches. (We could special case those here, but there's no
1046 * need--we'll do so at the bottom of the soft match loop anyway and
1047 * this duplicates less code.)
1049 * It's also tempting to break out of the soft match loop if 'n_soft ==
1050 * 1' but that would also miss lower-priority hard matches. We could
1051 * special case that also but again there's no need. */
1052 for (int i
= 0; i
< n_soft
; ) {
1053 if (!soft
[i
] || soft
[i
]->priority
<= hard_pri
) {
1054 soft
[i
] = soft
[--n_soft
];
1063 /* Find the highest priority among the soft matches. (We know this
1064 * must be higher than the hard match's priority; otherwise we would
1065 * have deleted all of the soft matches in the previous loop.) Count
1066 * the number of soft matches that have that priority. */
1069 for (int i
= 0; i
< n_soft
; i
++) {
1070 if (soft
[i
]->priority
> soft_pri
) {
1071 soft_pri
= soft
[i
]->priority
;
1073 } else if (soft
[i
]->priority
== soft_pri
) {
1077 ovs_assert(soft_pri
> hard_pri
);
1079 /* Look for a real match among the highest-priority soft matches.
1081 * It's unusual to have many conjunctive matches, so we use stubs to
1082 * avoid calling malloc() in the common case. An hmap has a built-in
1083 * stub for up to 2 hmap_nodes; possibly, we would benefit a variant
1084 * with a bigger stub. */
1085 struct conjunctive_match cm_stubs
[16];
1086 struct hmap matches
;
1088 hmap_init(&matches
);
1089 for (int i
= 0; i
< n_soft
; i
++) {
1092 if (soft
[i
]->priority
== soft_pri
1093 && find_conjunctive_match(soft
[i
], n_soft_pri
, &matches
,
1094 cm_stubs
, ARRAY_SIZE(cm_stubs
),
1096 uint32_t saved_conj_id
= flow
->conj_id
;
1097 const struct cls_rule
*rule
;
1100 rule
= classifier_lookup__(cls
, version
, flow
, wc
, false);
1101 flow
->conj_id
= saved_conj_id
;
1104 free_conjunctive_matches(&matches
,
1105 cm_stubs
, ARRAY_SIZE(cm_stubs
));
1106 if (soft
!= soft_stub
) {
1113 free_conjunctive_matches(&matches
, cm_stubs
, ARRAY_SIZE(cm_stubs
));
1115 /* There's no real match among the highest-priority soft matches.
1116 * However, if any of those soft matches has a lower-priority but
1117 * otherwise identical flow match, then we need to consider those for
1118 * soft or hard matches.
1120 * The next iteration of the soft match loop will delete any null
1121 * pointers we put into 'soft' (and some others too). */
1122 for (int i
= 0; i
< n_soft
; i
++) {
1123 if (soft
[i
]->priority
!= soft_pri
) {
1127 /* Find next-lower-priority flow with identical flow match. */
1128 match
= next_visible_rule_in_list(soft
[i
]->match
, version
);
1130 soft
[i
] = ovsrcu_get(struct cls_conjunction_set
*,
1133 /* The flow is a hard match; don't treat as a soft
1135 if (match
->priority
> hard_pri
) {
1137 hard_pri
= hard
->priority
;
1141 /* No such lower-priority flow (probably the common case). */
1147 if (soft
!= soft_stub
) {
1150 return hard
? hard
->cls_rule
: NULL
;
1153 /* Finds and returns the highest-priority rule in 'cls' that matches 'flow' and
1154 * that is visible in 'version'. Returns a null pointer if no rules in 'cls'
1155 * match 'flow'. If multiple rules of equal priority match 'flow', returns one
1158 * If a rule is found and 'wc' is non-null, bitwise-OR's 'wc' with the
1159 * set of bits that were significant in the lookup. At some point
1160 * earlier, 'wc' should have been initialized (e.g., by
1161 * flow_wildcards_init_catchall()).
1163 * 'flow' is non-const to allow for temporary modifications during the lookup.
1164 * Any changes are restored before returning. */
1165 const struct cls_rule
*
1166 classifier_lookup(const struct classifier
*cls
, ovs_version_t version
,
1167 struct flow
*flow
, struct flow_wildcards
*wc
)
1169 return classifier_lookup__(cls
, version
, flow
, wc
, true);
1172 /* Finds and returns a rule in 'cls' with exactly the same priority and
1173 * matching criteria as 'target', and that is visible in 'version'.
1174 * Only one such rule may ever exist. Returns a null pointer if 'cls' doesn't
1175 * contain an exact match. */
1176 const struct cls_rule
*
1177 classifier_find_rule_exactly(const struct classifier
*cls
,
1178 const struct cls_rule
*target
,
1179 ovs_version_t version
)
1181 const struct cls_match
*head
, *rule
;
1182 const struct cls_subtable
*subtable
;
1184 subtable
= find_subtable(cls
, target
->match
.mask
);
1189 head
= find_equal(subtable
, target
->match
.flow
,
1190 miniflow_hash_in_minimask(target
->match
.flow
,
1191 target
->match
.mask
, 0));
1195 CLS_MATCH_FOR_EACH (rule
, head
) {
1196 if (rule
->priority
< target
->priority
) {
1197 break; /* Not found. */
1199 if (rule
->priority
== target
->priority
1200 && cls_match_visible_in_version(rule
, version
)) {
1201 return rule
->cls_rule
;
1207 /* Finds and returns a rule in 'cls' with priority 'priority' and exactly the
1208 * same matching criteria as 'target', and that is visible in 'version'.
1209 * Returns a null pointer if 'cls' doesn't contain an exact match visible in
1211 const struct cls_rule
*
1212 classifier_find_match_exactly(const struct classifier
*cls
,
1213 const struct match
*target
, int priority
,
1214 ovs_version_t version
)
1216 const struct cls_rule
*retval
;
1219 cls_rule_init(&cr
, target
, priority
);
1220 retval
= classifier_find_rule_exactly(cls
, &cr
, version
);
1221 cls_rule_destroy(&cr
);
1226 /* Finds and returns a rule in 'cls' with priority 'priority' and exactly the
1227 * same matching criteria as 'target', and that is visible in 'version'.
1228 * Returns a null pointer if 'cls' doesn't contain an exact match visible in
1230 const struct cls_rule
*
1231 classifier_find_minimatch_exactly(const struct classifier
*cls
,
1232 const struct minimatch
*target
, int priority
,
1233 ovs_version_t version
)
1235 const struct cls_rule
*retval
;
1238 cls_rule_init_from_minimatch(&cr
, target
, priority
);
1239 retval
= classifier_find_rule_exactly(cls
, &cr
, version
);
1240 cls_rule_destroy(&cr
);
1245 /* Checks if 'target' would overlap any other rule in 'cls' in 'version'. Two
1246 * rules are considered to overlap if both rules have the same priority and a
1247 * packet could match both, and if both rules are visible in the same version.
1249 * A trivial example of overlapping rules is two rules matching disjoint sets
1250 * of fields. E.g., if one rule matches only on port number, while another only
1251 * on dl_type, any packet from that specific port and with that specific
1252 * dl_type could match both, if the rules also have the same priority. */
1254 classifier_rule_overlaps(const struct classifier
*cls
,
1255 const struct cls_rule
*target
, ovs_version_t version
)
1257 struct cls_subtable
*subtable
;
1259 /* Iterate subtables in the descending max priority order. */
1260 PVECTOR_FOR_EACH_PRIORITY (subtable
, target
->priority
, 2,
1261 sizeof(struct cls_subtable
), &cls
->subtables
) {
1263 struct minimask mask
;
1264 uint64_t storage
[FLOW_U64S
];
1266 const struct cls_rule
*rule
;
1268 minimask_combine(&m
.mask
, target
->match
.mask
, &subtable
->mask
,
1271 RCULIST_FOR_EACH (rule
, node
, &subtable
->rules_list
) {
1272 if (rule
->priority
== target
->priority
1273 && miniflow_equal_in_minimask(target
->match
.flow
,
1274 rule
->match
.flow
, &m
.mask
)
1275 && cls_rule_visible_in_version(rule
, version
)) {
1283 /* Returns true if 'rule' exactly matches 'criteria' or if 'rule' is more
1284 * specific than 'criteria'. That is, 'rule' matches 'criteria' and this
1285 * function returns true if, for every field:
1287 * - 'criteria' and 'rule' specify the same (non-wildcarded) value for the
1290 * - 'criteria' wildcards the field,
1292 * Conversely, 'rule' does not match 'criteria' and this function returns false
1293 * if, for at least one field:
1295 * - 'criteria' and 'rule' specify different values for the field, or
1297 * - 'criteria' specifies a value for the field but 'rule' wildcards it.
1299 * Equivalently, the truth table for whether a field matches is:
1304 * r +---------+---------+
1305 * i wild | yes | yes |
1307 * e +---------+---------+
1308 * r exact | no |if values|
1310 * a +---------+---------+
1312 * This is the matching rule used by OpenFlow 1.0 non-strict OFPT_FLOW_MOD
1313 * commands and by OpenFlow 1.0 aggregate and flow stats.
1315 * Ignores rule->priority. */
1317 cls_rule_is_loose_match(const struct cls_rule
*rule
,
1318 const struct minimatch
*criteria
)
1320 return (!minimask_has_extra(rule
->match
.mask
, criteria
->mask
)
1321 && miniflow_equal_in_minimask(rule
->match
.flow
, criteria
->flow
,
1328 rule_matches(const struct cls_rule
*rule
, const struct cls_rule
*target
,
1329 ovs_version_t version
)
1331 /* Rule may only match a target if it is visible in target's version. */
1332 return cls_rule_visible_in_version(rule
, version
)
1333 && (!target
|| miniflow_equal_in_minimask(rule
->match
.flow
,
1335 target
->match
.mask
));
1338 static const struct cls_rule
*
1339 search_subtable(const struct cls_subtable
*subtable
,
1340 struct cls_cursor
*cursor
)
1343 || !minimask_has_extra(&subtable
->mask
, cursor
->target
->match
.mask
)) {
1344 const struct cls_rule
*rule
;
1346 RCULIST_FOR_EACH (rule
, node
, &subtable
->rules_list
) {
1347 if (rule_matches(rule
, cursor
->target
, cursor
->version
)) {
1355 /* Initializes 'cursor' for iterating through rules in 'cls', and returns the
1358 * - If 'target' is null, or if the 'target' is a catchall target, the
1359 * cursor will visit every rule in 'cls' that is visible in 'version'.
1361 * - If 'target' is nonnull, the cursor will visit each 'rule' in 'cls'
1362 * such that cls_rule_is_loose_match(rule, target) returns true and that
1363 * the rule is visible in 'version'.
1365 * Ignores target->priority. */
1367 cls_cursor_start(const struct classifier
*cls
, const struct cls_rule
*target
,
1368 ovs_version_t version
)
1370 struct cls_cursor cursor
;
1371 struct cls_subtable
*subtable
;
1373 memset(&cursor
, 0x0, sizeof cursor
);
1375 cursor
.target
= target
&& !cls_rule_is_catchall(target
) ? target
: NULL
;
1376 cursor
.version
= version
;
1379 /* Find first rule. */
1380 PVECTOR_CURSOR_FOR_EACH (subtable
, &cursor
.subtables
,
1381 &cursor
.cls
->subtables
) {
1382 const struct cls_rule
*rule
= search_subtable(subtable
, &cursor
);
1385 cursor
.subtable
= subtable
;
1394 static const struct cls_rule
*
1395 cls_cursor_next(struct cls_cursor
*cursor
)
1397 const struct cls_rule
*rule
;
1398 const struct cls_subtable
*subtable
;
1400 rule
= cursor
->rule
;
1401 subtable
= cursor
->subtable
;
1402 RCULIST_FOR_EACH_CONTINUE (rule
, node
, &subtable
->rules_list
) {
1403 if (rule_matches(rule
, cursor
->target
, cursor
->version
)) {
1408 PVECTOR_CURSOR_FOR_EACH_CONTINUE (subtable
, &cursor
->subtables
) {
1409 rule
= search_subtable(subtable
, cursor
);
1411 cursor
->subtable
= subtable
;
1419 /* Sets 'cursor->rule' to the next matching cls_rule in 'cursor''s iteration,
1420 * or to null if all matching rules have been visited. */
1422 cls_cursor_advance(struct cls_cursor
*cursor
)
1424 cursor
->rule
= cls_cursor_next(cursor
);
1427 static struct cls_subtable
*
1428 find_subtable(const struct classifier
*cls
, const struct minimask
*mask
)
1430 struct cls_subtable
*subtable
;
1432 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, minimask_hash(mask
, 0),
1433 &cls
->subtables_map
) {
1434 if (minimask_equal(mask
, &subtable
->mask
)) {
1441 /* Initializes 'map' with a subset of 'miniflow''s maps that includes only the
1442 * portions with u64-offset 'i' such that 'start' <= i < 'end'. Does not copy
1443 * any data from 'miniflow' to 'map'. */
1444 static struct flowmap
1445 miniflow_get_map_in_range(const struct miniflow
*miniflow
, uint8_t start
,
1451 map
= miniflow
->map
;
1453 /* Clear the bits before 'start'. */
1454 while (start
>= MAP_T_BITS
) {
1455 start
-= MAP_T_BITS
;
1457 map
.bits
[start
/ MAP_T_BITS
] = 0;
1460 flowmap_clear(&map
, ofs
, start
);
1463 /* Clear the bits starting at 'end'. */
1464 if (end
< FLOW_U64S
) {
1465 /* flowmap_clear() can handle at most MAP_T_BITS at a time. */
1466 ovs_assert(FLOW_U64S
- end
<= MAP_T_BITS
);
1467 flowmap_clear(&map
, end
, FLOW_U64S
- end
);
1473 subtable_destroy_cb(struct cls_subtable
*subtable
)
1477 ovs_assert(ovsrcu_get_protected(struct trie_node
*, &subtable
->ports_trie
)
1479 ovs_assert(cmap_is_empty(&subtable
->rules
));
1480 ovs_assert(rculist_is_empty(&subtable
->rules_list
));
1482 for (i
= 0; i
< subtable
->n_indices
; i
++) {
1483 ccmap_destroy(&subtable
->indices
[i
]);
1485 cmap_destroy(&subtable
->rules
);
1487 ovsrcu_postpone(free
, subtable
);
1490 /* The new subtable will be visible to the readers only after this. */
1491 static struct cls_subtable
*
1492 insert_subtable(struct classifier
*cls
, const struct minimask
*mask
)
1494 uint32_t hash
= minimask_hash(mask
, 0);
1495 struct cls_subtable
*subtable
;
1497 struct flowmap stage_map
;
1499 size_t count
= miniflow_n_values(&mask
->masks
);
1501 subtable
= xzalloc(sizeof *subtable
+ MINIFLOW_VALUES_SIZE(count
));
1502 cmap_init(&subtable
->rules
);
1503 miniflow_clone(CONST_CAST(struct miniflow
*, &subtable
->mask
.masks
),
1504 &mask
->masks
, count
);
1506 /* Init indices for segmented lookup, if any. */
1508 for (i
= 0; i
< cls
->n_flow_segments
; i
++) {
1509 stage_map
= miniflow_get_map_in_range(&mask
->masks
, prev
,
1510 cls
->flow_segments
[i
]);
1511 /* Add an index if it adds mask bits. */
1512 if (!flowmap_is_empty(stage_map
)) {
1513 ccmap_init(&subtable
->indices
[index
]);
1514 *CONST_CAST(struct flowmap
*, &subtable
->index_maps
[index
])
1518 prev
= cls
->flow_segments
[i
];
1520 /* Map for the final stage. */
1521 *CONST_CAST(struct flowmap
*, &subtable
->index_maps
[index
])
1522 = miniflow_get_map_in_range(&mask
->masks
, prev
, FLOW_U64S
);
1523 /* Check if the final stage adds any bits. */
1525 if (flowmap_is_empty(subtable
->index_maps
[index
])) {
1526 /* Remove the last index, as it has the same fields as the rules
1529 ccmap_destroy(&subtable
->indices
[index
]);
1532 *CONST_CAST(uint8_t *, &subtable
->n_indices
) = index
;
1534 for (i
= 0; i
< cls
->n_tries
; i
++) {
1535 const struct mf_field
*field
1536 = ovsrcu_get(struct mf_field
*, &cls
->tries
[i
].field
);
1537 subtable
->trie_plen
[i
]
1538 = field
? minimask_get_prefix_len(mask
, field
) : 0;
1542 ovsrcu_set_hidden(&subtable
->ports_trie
, NULL
);
1543 *CONST_CAST(int *, &subtable
->ports_mask_len
)
1544 = 32 - ctz32(ntohl(miniflow_get_ports(&mask
->masks
)));
1546 /* List of rules. */
1547 rculist_init(&subtable
->rules_list
);
1549 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, hash
);
1554 /* RCU readers may still access the subtable before it is actually freed. */
1556 destroy_subtable(struct classifier
*cls
, struct cls_subtable
*subtable
)
1558 pvector_remove(&cls
->subtables
, subtable
);
1559 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
1560 minimask_hash(&subtable
->mask
, 0));
1562 ovsrcu_postpone(subtable_destroy_cb
, subtable
);
1565 static unsigned int be_get_bit_at(const ovs_be32 value
[], unsigned int ofs
);
1567 /* Return 'true' if can skip rest of the subtable based on the prefix trie
1568 * lookup results. */
1570 check_tries(struct trie_ctx trie_ctx
[CLS_MAX_TRIES
], unsigned int n_tries
,
1571 const unsigned int field_plen
[CLS_MAX_TRIES
],
1572 const struct flowmap range_map
, const struct flow
*flow
,
1573 struct flow_wildcards
*wc
)
1577 /* Check if we could avoid fully unwildcarding the next level of
1578 * fields using the prefix tries. The trie checks are done only as
1579 * needed to avoid folding in additional bits to the wildcards mask. */
1580 for (j
= 0; j
< n_tries
; j
++) {
1581 /* Is the trie field relevant for this subtable? */
1582 if (field_plen
[j
]) {
1583 struct trie_ctx
*ctx
= &trie_ctx
[j
];
1584 const struct mf_field
*ctx_field
1585 = ovsrcu_get(struct mf_field
*, &ctx
->trie
->field
);
1587 /* Is the trie field within the current range of fields? */
1589 || !flowmap_is_set(&range_map
, ctx_field
->flow_be32ofs
/ 2)) {
1593 /* On-demand trie lookup. */
1594 if (!ctx
->lookup_done
) {
1595 memset(&ctx
->match_plens
, 0, sizeof ctx
->match_plens
);
1596 ctx
->maskbits
= trie_lookup(ctx
->trie
, flow
, &ctx
->match_plens
);
1597 ctx
->lookup_done
= true;
1599 /* Possible to skip the rest of the subtable if subtable's
1600 * prefix on the field is not included in the lookup result. */
1601 if (!be_get_bit_at(&ctx
->match_plens
.be32
, field_plen
[j
] - 1)) {
1602 /* We want the trie lookup to never result in unwildcarding
1603 * any bits that would not be unwildcarded otherwise.
1604 * Since the trie is shared by the whole classifier, it is
1605 * possible that the 'maskbits' contain bits that are
1606 * irrelevant for the partition relevant for the current
1607 * packet. Hence the checks below. */
1609 /* Check that the trie result will not unwildcard more bits
1610 * than this subtable would otherwise. */
1611 if (ctx
->maskbits
<= field_plen
[j
]) {
1612 /* Unwildcard the bits and skip the rest. */
1613 mask_set_prefix_bits(wc
, ctx_field
->flow_be32ofs
,
1615 /* Note: Prerequisite already unwildcarded, as the only
1616 * prerequisite of the supported trie lookup fields is
1617 * the ethertype, which is always unwildcarded. */
1620 /* Can skip if the field is already unwildcarded. */
1621 if (mask_prefix_bits_set(wc
, ctx_field
->flow_be32ofs
,
1631 /* Returns true if 'target' satisifies 'flow'/'mask', that is, if each bit
1632 * for which 'flow', for which 'mask' has a bit set, specifies a particular
1633 * value has the correct value in 'target'.
1635 * This function is equivalent to miniflow_equal_flow_in_minimask(flow,
1636 * target, mask) but this is faster because of the invariant that
1637 * flow->map and mask->masks.map are the same, and that this version
1638 * takes the 'wc'. */
1640 miniflow_and_mask_matches_flow(const struct miniflow
*flow
,
1641 const struct minimask
*mask
,
1642 const struct flow
*target
)
1644 const uint64_t *flowp
= miniflow_get_values(flow
);
1645 const uint64_t *maskp
= miniflow_get_values(&mask
->masks
);
1646 const uint64_t *target_u64
= (const uint64_t *)target
;
1649 FLOWMAP_FOR_EACH_MAP (map
, mask
->masks
.map
) {
1652 MAP_FOR_EACH_INDEX (idx
, map
) {
1653 if ((*flowp
++ ^ target_u64
[idx
]) & *maskp
++) {
1657 target_u64
+= MAP_T_BITS
;
1662 static inline const struct cls_match
*
1663 find_match(const struct cls_subtable
*subtable
, ovs_version_t version
,
1664 const struct flow
*flow
, uint32_t hash
)
1666 const struct cls_match
*head
, *rule
;
1668 CMAP_FOR_EACH_WITH_HASH (head
, cmap_node
, hash
, &subtable
->rules
) {
1669 if (OVS_LIKELY(miniflow_and_mask_matches_flow(&head
->flow
,
1672 /* Return highest priority rule that is visible. */
1673 CLS_MATCH_FOR_EACH (rule
, head
) {
1674 if (OVS_LIKELY(cls_match_visible_in_version(rule
, version
))) {
1684 static const struct cls_match
*
1685 find_match_wc(const struct cls_subtable
*subtable
, ovs_version_t version
,
1686 const struct flow
*flow
, struct trie_ctx trie_ctx
[CLS_MAX_TRIES
],
1687 unsigned int n_tries
, struct flow_wildcards
*wc
)
1689 if (OVS_UNLIKELY(!wc
)) {
1690 return find_match(subtable
, version
, flow
,
1691 flow_hash_in_minimask(flow
, &subtable
->mask
, 0));
1694 uint32_t basis
= 0, hash
;
1695 const struct cls_match
*rule
= NULL
;
1696 struct flowmap stages_map
= FLOWMAP_EMPTY_INITIALIZER
;
1697 unsigned int mask_offset
= 0;
1700 /* Try to finish early by checking fields in segments. */
1701 for (i
= 0; i
< subtable
->n_indices
; i
++) {
1702 if (check_tries(trie_ctx
, n_tries
, subtable
->trie_plen
,
1703 subtable
->index_maps
[i
], flow
, wc
)) {
1704 /* 'wc' bits for the trie field set, now unwildcard the preceding
1705 * bits used so far. */
1709 /* Accumulate the map used so far. */
1710 stages_map
= flowmap_or(stages_map
, subtable
->index_maps
[i
]);
1712 hash
= flow_hash_in_minimask_range(flow
, &subtable
->mask
,
1713 subtable
->index_maps
[i
],
1714 &mask_offset
, &basis
);
1716 if (!ccmap_find(&subtable
->indices
[i
], hash
)) {
1720 /* Trie check for the final range. */
1721 if (check_tries(trie_ctx
, n_tries
, subtable
->trie_plen
,
1722 subtable
->index_maps
[i
], flow
, wc
)) {
1725 hash
= flow_hash_in_minimask_range(flow
, &subtable
->mask
,
1726 subtable
->index_maps
[i
],
1727 &mask_offset
, &basis
);
1728 rule
= find_match(subtable
, version
, flow
, hash
);
1729 if (!rule
&& subtable
->ports_mask_len
) {
1730 /* The final stage had ports, but there was no match. Instead of
1731 * unwildcarding all the ports bits, use the ports trie to figure out a
1732 * smaller set of bits to unwildcard. */
1734 ovs_be32 value
, plens
, mask
;
1736 mask
= miniflow_get_ports(&subtable
->mask
.masks
);
1737 value
= ((OVS_FORCE ovs_be32
*)flow
)[TP_PORTS_OFS32
] & mask
;
1738 mbits
= trie_lookup_value(&subtable
->ports_trie
, &value
, &plens
, 32);
1740 ((OVS_FORCE ovs_be32
*)&wc
->masks
)[TP_PORTS_OFS32
] |=
1741 mask
& be32_prefix_mask(mbits
);
1746 /* Must unwildcard all the fields, as they were looked at. */
1747 flow_wildcards_fold_minimask(wc
, &subtable
->mask
);
1751 /* Unwildcard the bits in stages so far, as they were used in determining
1752 * there is no match. */
1753 flow_wildcards_fold_minimask_in_map(wc
, &subtable
->mask
, stages_map
);
1757 static struct cls_match
*
1758 find_equal(const struct cls_subtable
*subtable
, const struct miniflow
*flow
,
1761 struct cls_match
*head
;
1763 CMAP_FOR_EACH_WITH_HASH (head
, cmap_node
, hash
, &subtable
->rules
) {
1764 if (miniflow_equal(&head
->flow
, flow
)) {
1771 /* A longest-prefix match tree. */
1773 /* Return at least 'plen' bits of the 'prefix', starting at bit offset 'ofs'.
1774 * Prefixes are in the network byte order, and the offset 0 corresponds to
1775 * the most significant bit of the first byte. The offset can be read as
1776 * "how many bits to skip from the start of the prefix starting at 'pr'". */
1778 raw_get_prefix(const ovs_be32 pr
[], unsigned int ofs
, unsigned int plen
)
1782 pr
+= ofs
/ 32; /* Where to start. */
1783 ofs
%= 32; /* How many bits to skip at 'pr'. */
1785 prefix
= ntohl(*pr
) << ofs
; /* Get the first 32 - ofs bits. */
1786 if (plen
> 32 - ofs
) { /* Need more than we have already? */
1787 prefix
|= ntohl(*++pr
) >> (32 - ofs
);
1789 /* Return with possible unwanted bits at the end. */
1793 /* Return min(TRIE_PREFIX_BITS, plen) bits of the 'prefix', starting at bit
1794 * offset 'ofs'. Prefixes are in the network byte order, and the offset 0
1795 * corresponds to the most significant bit of the first byte. The offset can
1796 * be read as "how many bits to skip from the start of the prefix starting at
1799 trie_get_prefix(const ovs_be32 pr
[], unsigned int ofs
, unsigned int plen
)
1804 if (plen
> TRIE_PREFIX_BITS
) {
1805 plen
= TRIE_PREFIX_BITS
; /* Get at most TRIE_PREFIX_BITS. */
1807 /* Return with unwanted bits cleared. */
1808 return raw_get_prefix(pr
, ofs
, plen
) & ~0u << (32 - plen
);
1811 /* Return the number of equal bits in 'n_bits' of 'prefix's MSBs and a 'value'
1812 * starting at "MSB 0"-based offset 'ofs'. */
1814 prefix_equal_bits(uint32_t prefix
, unsigned int n_bits
, const ovs_be32 value
[],
1817 uint64_t diff
= prefix
^ raw_get_prefix(value
, ofs
, n_bits
);
1818 /* Set the bit after the relevant bits to limit the result. */
1819 return raw_clz64(diff
<< 32 | UINT64_C(1) << (63 - n_bits
));
1822 /* Return the number of equal bits in 'node' prefix and a 'prefix' of length
1823 * 'plen', starting at "MSB 0"-based offset 'ofs'. */
1825 trie_prefix_equal_bits(const struct trie_node
*node
, const ovs_be32 prefix
[],
1826 unsigned int ofs
, unsigned int plen
)
1828 return prefix_equal_bits(node
->prefix
, MIN(node
->n_bits
, plen
- ofs
),
1832 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' can
1833 * be greater than 31. */
1835 be_get_bit_at(const ovs_be32 value
[], unsigned int ofs
)
1837 return (((const uint8_t *)value
)[ofs
/ 8] >> (7 - ofs
% 8)) & 1u;
1840 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' must
1841 * be between 0 and 31, inclusive. */
1843 get_bit_at(const uint32_t prefix
, unsigned int ofs
)
1845 return (prefix
>> (31 - ofs
)) & 1u;
1848 /* Create new branch. */
1849 static struct trie_node
*
1850 trie_branch_create(const ovs_be32
*prefix
, unsigned int ofs
, unsigned int plen
,
1851 unsigned int n_rules
)
1853 struct trie_node
*node
= xmalloc(sizeof *node
);
1855 node
->prefix
= trie_get_prefix(prefix
, ofs
, plen
);
1857 if (plen
<= TRIE_PREFIX_BITS
) {
1858 node
->n_bits
= plen
;
1859 ovsrcu_set_hidden(&node
->edges
[0], NULL
);
1860 ovsrcu_set_hidden(&node
->edges
[1], NULL
);
1861 node
->n_rules
= n_rules
;
1862 } else { /* Need intermediate nodes. */
1863 struct trie_node
*subnode
= trie_branch_create(prefix
,
1864 ofs
+ TRIE_PREFIX_BITS
,
1865 plen
- TRIE_PREFIX_BITS
,
1867 int bit
= get_bit_at(subnode
->prefix
, 0);
1868 node
->n_bits
= TRIE_PREFIX_BITS
;
1869 ovsrcu_set_hidden(&node
->edges
[bit
], subnode
);
1870 ovsrcu_set_hidden(&node
->edges
[!bit
], NULL
);
1877 trie_node_destroy(const struct trie_node
*node
)
1879 ovsrcu_postpone(free
, CONST_CAST(struct trie_node
*, node
));
1882 /* Copy a trie node for modification and postpone delete the old one. */
1883 static struct trie_node
*
1884 trie_node_rcu_realloc(const struct trie_node
*node
)
1886 struct trie_node
*new_node
= xmalloc(sizeof *node
);
1889 trie_node_destroy(node
);
1895 trie_destroy(rcu_trie_ptr
*trie
)
1897 struct trie_node
*node
= ovsrcu_get_protected(struct trie_node
*, trie
);
1900 ovsrcu_set_hidden(trie
, NULL
);
1901 trie_destroy(&node
->edges
[0]);
1902 trie_destroy(&node
->edges
[1]);
1903 trie_node_destroy(node
);
1908 trie_is_leaf(const struct trie_node
*trie
)
1911 return !ovsrcu_get(struct trie_node
*, &trie
->edges
[0])
1912 && !ovsrcu_get(struct trie_node
*, &trie
->edges
[1]);
1916 mask_set_prefix_bits(struct flow_wildcards
*wc
, uint8_t be32ofs
,
1917 unsigned int n_bits
)
1919 ovs_be32
*mask
= &((ovs_be32
*)&wc
->masks
)[be32ofs
];
1922 for (i
= 0; i
< n_bits
/ 32; i
++) {
1923 mask
[i
] = OVS_BE32_MAX
;
1926 mask
[i
] |= htonl(~0u << (32 - n_bits
% 32));
1931 mask_prefix_bits_set(const struct flow_wildcards
*wc
, uint8_t be32ofs
,
1932 unsigned int n_bits
)
1934 ovs_be32
*mask
= &((ovs_be32
*)&wc
->masks
)[be32ofs
];
1936 ovs_be32 zeroes
= 0;
1938 for (i
= 0; i
< n_bits
/ 32; i
++) {
1942 zeroes
|= ~mask
[i
] & htonl(~0u << (32 - n_bits
% 32));
1945 return !zeroes
; /* All 'n_bits' bits set. */
1948 static rcu_trie_ptr
*
1949 trie_next_edge(struct trie_node
*node
, const ovs_be32 value
[],
1952 return node
->edges
+ be_get_bit_at(value
, ofs
);
1955 static const struct trie_node
*
1956 trie_next_node(const struct trie_node
*node
, const ovs_be32 value
[],
1959 return ovsrcu_get(struct trie_node
*,
1960 &node
->edges
[be_get_bit_at(value
, ofs
)]);
1963 /* Set the bit at ("MSB 0"-based) offset 'ofs'. 'ofs' can be greater than 31.
1966 be_set_bit_at(ovs_be32 value
[], unsigned int ofs
)
1968 ((uint8_t *)value
)[ofs
/ 8] |= 1u << (7 - ofs
% 8);
1971 /* Returns the number of bits in the prefix mask necessary to determine a
1972 * mismatch, in case there are longer prefixes in the tree below the one that
1974 * '*plens' will have a bit set for each prefix length that may have matching
1975 * rules. The caller is responsible for clearing the '*plens' prior to
1979 trie_lookup_value(const rcu_trie_ptr
*trie
, const ovs_be32 value
[],
1980 ovs_be32 plens
[], unsigned int n_bits
)
1982 const struct trie_node
*prev
= NULL
;
1983 const struct trie_node
*node
= ovsrcu_get(struct trie_node
*, trie
);
1984 unsigned int match_len
= 0; /* Number of matching bits. */
1986 for (; node
; prev
= node
, node
= trie_next_node(node
, value
, match_len
)) {
1987 unsigned int eqbits
;
1988 /* Check if this edge can be followed. */
1989 eqbits
= prefix_equal_bits(node
->prefix
, node
->n_bits
, value
,
1991 match_len
+= eqbits
;
1992 if (eqbits
< node
->n_bits
) { /* Mismatch, nothing more to be found. */
1993 /* Bit at offset 'match_len' differed. */
1994 return match_len
+ 1; /* Includes the first mismatching bit. */
1996 /* Full match, check if rules exist at this prefix length. */
1997 if (node
->n_rules
> 0) {
1998 be_set_bit_at(plens
, match_len
- 1);
2000 if (match_len
>= n_bits
) {
2001 return n_bits
; /* Full prefix. */
2004 /* node == NULL. Full match so far, but we tried to follow an
2005 * non-existing branch. Need to exclude the other branch if it exists
2006 * (it does not if we were called on an empty trie or 'prev' is a leaf
2008 return !prev
|| trie_is_leaf(prev
) ? match_len
: match_len
+ 1;
2012 trie_lookup(const struct cls_trie
*trie
, const struct flow
*flow
,
2013 union trie_prefix
*plens
)
2015 const struct mf_field
*mf
= ovsrcu_get(struct mf_field
*, &trie
->field
);
2017 /* Check that current flow matches the prerequisites for the trie
2018 * field. Some match fields are used for multiple purposes, so we
2019 * must check that the trie is relevant for this flow. */
2020 if (mf
&& mf_are_prereqs_ok(mf
, flow
, NULL
)) {
2021 return trie_lookup_value(&trie
->root
,
2022 &((ovs_be32
*)flow
)[mf
->flow_be32ofs
],
2023 &plens
->be32
, mf
->n_bits
);
2025 memset(plens
, 0xff, sizeof *plens
); /* All prefixes, no skipping. */
2026 return 0; /* Value not used in this case. */
2029 /* Returns the length of a prefix match mask for the field 'mf' in 'minimask'.
2030 * Returns the u32 offset to the miniflow data in '*miniflow_index', if
2031 * 'miniflow_index' is not NULL. */
2033 minimask_get_prefix_len(const struct minimask
*minimask
,
2034 const struct mf_field
*mf
)
2036 unsigned int n_bits
= 0, mask_tz
= 0; /* Non-zero when end of mask seen. */
2037 uint8_t be32_ofs
= mf
->flow_be32ofs
;
2038 uint8_t be32_end
= be32_ofs
+ mf
->n_bytes
/ 4;
2040 for (; be32_ofs
< be32_end
; ++be32_ofs
) {
2041 uint32_t mask
= ntohl(minimask_get_be32(minimask
, be32_ofs
));
2043 /* Validate mask, count the mask length. */
2046 return 0; /* No bits allowed after mask ended. */
2049 if (~mask
& (~mask
+ 1)) {
2050 return 0; /* Mask not contiguous. */
2052 mask_tz
= ctz32(mask
);
2053 n_bits
+= 32 - mask_tz
;
2061 * This is called only when mask prefix is known to be CIDR and non-zero.
2062 * Relies on the fact that the flow and mask have the same map, and since
2063 * the mask is CIDR, the storage for the flow field exists even if it
2064 * happened to be zeros.
2066 static const ovs_be32
*
2067 minimatch_get_prefix(const struct minimatch
*match
, rcu_field_ptr
*field
)
2069 struct mf_field
*mf
= ovsrcu_get_protected(struct mf_field
*, field
);
2070 size_t u64_ofs
= mf
->flow_be32ofs
/ 2;
2072 return (OVS_FORCE
const ovs_be32
*)miniflow_get__(match
->flow
, u64_ofs
)
2073 + (mf
->flow_be32ofs
& 1);
2076 /* Insert rule in to the prefix tree.
2077 * 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2080 trie_insert(struct cls_trie
*trie
, const struct cls_rule
*rule
, int mlen
)
2082 trie_insert_prefix(&trie
->root
,
2083 minimatch_get_prefix(&rule
->match
, &trie
->field
), mlen
);
2087 trie_insert_prefix(rcu_trie_ptr
*edge
, const ovs_be32
*prefix
, int mlen
)
2089 struct trie_node
*node
;
2092 /* Walk the tree. */
2093 for (; (node
= ovsrcu_get_protected(struct trie_node
*, edge
));
2094 edge
= trie_next_edge(node
, prefix
, ofs
)) {
2095 unsigned int eqbits
= trie_prefix_equal_bits(node
, prefix
, ofs
, mlen
);
2097 if (eqbits
< node
->n_bits
) {
2098 /* Mismatch, new node needs to be inserted above. */
2099 int old_branch
= get_bit_at(node
->prefix
, eqbits
);
2100 struct trie_node
*new_parent
;
2102 new_parent
= trie_branch_create(prefix
, ofs
- eqbits
, eqbits
,
2103 ofs
== mlen
? 1 : 0);
2104 /* Copy the node to modify it. */
2105 node
= trie_node_rcu_realloc(node
);
2106 /* Adjust the new node for its new position in the tree. */
2107 node
->prefix
<<= eqbits
;
2108 node
->n_bits
-= eqbits
;
2109 ovsrcu_set_hidden(&new_parent
->edges
[old_branch
], node
);
2111 /* Check if need a new branch for the new rule. */
2113 ovsrcu_set_hidden(&new_parent
->edges
[!old_branch
],
2114 trie_branch_create(prefix
, ofs
, mlen
- ofs
,
2117 ovsrcu_set(edge
, new_parent
); /* Publish changes. */
2120 /* Full match so far. */
2123 /* Full match at the current node, rule needs to be added here. */
2128 /* Must insert a new tree branch for the new rule. */
2129 ovsrcu_set(edge
, trie_branch_create(prefix
, ofs
, mlen
- ofs
, 1));
2132 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2135 trie_remove(struct cls_trie
*trie
, const struct cls_rule
*rule
, int mlen
)
2137 trie_remove_prefix(&trie
->root
,
2138 minimatch_get_prefix(&rule
->match
, &trie
->field
), mlen
);
2141 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2144 trie_remove_prefix(rcu_trie_ptr
*root
, const ovs_be32
*prefix
, int mlen
)
2146 struct trie_node
*node
;
2147 rcu_trie_ptr
*edges
[sizeof(union trie_prefix
) * CHAR_BIT
];
2148 int depth
= 0, ofs
= 0;
2150 /* Walk the tree. */
2151 for (edges
[0] = root
;
2152 (node
= ovsrcu_get_protected(struct trie_node
*, edges
[depth
]));
2153 edges
[++depth
] = trie_next_edge(node
, prefix
, ofs
)) {
2154 unsigned int eqbits
= trie_prefix_equal_bits(node
, prefix
, ofs
, mlen
);
2156 if (eqbits
< node
->n_bits
) {
2157 /* Mismatch, nothing to be removed. This should never happen, as
2158 * only rules in the classifier are ever removed. */
2159 break; /* Log a warning. */
2161 /* Full match so far. */
2165 /* Full prefix match at the current node, remove rule here. */
2166 if (!node
->n_rules
) {
2167 break; /* Log a warning. */
2171 /* Check if can prune the tree. */
2172 while (!node
->n_rules
) {
2173 struct trie_node
*next
,
2174 *edge0
= ovsrcu_get_protected(struct trie_node
*,
2176 *edge1
= ovsrcu_get_protected(struct trie_node
*,
2179 if (edge0
&& edge1
) {
2180 break; /* A branching point, cannot prune. */
2183 /* Else have at most one child node, remove this node. */
2184 next
= edge0
? edge0
: edge1
;
2187 if (node
->n_bits
+ next
->n_bits
> TRIE_PREFIX_BITS
) {
2188 break; /* Cannot combine. */
2190 next
= trie_node_rcu_realloc(next
); /* Modify. */
2192 /* Combine node with next. */
2193 next
->prefix
= node
->prefix
| next
->prefix
>> node
->n_bits
;
2194 next
->n_bits
+= node
->n_bits
;
2196 /* Update the parent's edge. */
2197 ovsrcu_set(edges
[depth
], next
); /* Publish changes. */
2198 trie_node_destroy(node
);
2200 if (next
|| !depth
) {
2201 /* Branch not pruned or at root, nothing more to do. */
2204 node
= ovsrcu_get_protected(struct trie_node
*,
2210 /* Cannot go deeper. This should never happen, since only rules
2211 * that actually exist in the classifier are ever removed. */
2215 #define CLS_MATCH_POISON (struct cls_match *)(UINTPTR_MAX / 0xf * 0xb)
2218 cls_match_free_cb(struct cls_match
*rule
)
2220 ovsrcu_set_hidden(&rule
->next
, CLS_MATCH_POISON
);