]>
git.proxmox.com Git - mirror_ovs.git/blob - lib/classifier.c
2 * Copyright (c) 2009, 2010, 2011, 2012, 2013 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"
20 #include <netinet/in.h>
21 #include "byte-order.h"
22 #include "dynamic-string.h"
27 #include "ovs-thread.h"
31 VLOG_DEFINE_THIS_MODULE(classifier
);
36 /* Ports trie depends on both ports sharing the same ovs_be32. */
37 #define TP_PORTS_OFS32 (offsetof(struct flow, tp_src) / 4)
38 BUILD_ASSERT_DECL(TP_PORTS_OFS32
== offsetof(struct flow
, tp_dst
) / 4);
40 /* Prefix trie for a 'field' */
42 const struct mf_field
*field
; /* Trie field, or NULL. */
43 struct trie_node
*root
; /* NULL if none. */
46 struct cls_subtable_entry
{
47 struct cls_subtable
*subtable
;
49 unsigned int max_priority
;
52 struct cls_subtables
{
53 size_t count
; /* One past last valid array element. */
54 size_t alloc_size
; /* Number of allocated elements. */
55 struct cls_subtable_entry
*array
;
59 CLS_MAX_INDICES
= 3 /* Maximum number of lookup indices per subtable. */
62 struct cls_classifier
{
63 int n_rules
; /* Total number of rules. */
64 uint8_t n_flow_segments
;
65 uint8_t flow_segments
[CLS_MAX_INDICES
]; /* Flow segment boundaries to use
66 * for staged lookup. */
67 struct hmap subtables_map
; /* Contains "struct cls_subtable"s. */
68 struct cls_subtables subtables
;
69 struct hmap partitions
; /* Contains "struct cls_partition"s. */
70 struct cls_trie tries
[CLS_MAX_TRIES
]; /* Prefix tries. */
74 /* A set of rules that all have the same fields wildcarded. */
76 struct hmap_node hmap_node
; /* Within struct cls_classifier 'subtables_map'
78 struct hmap rules
; /* Contains "struct cls_rule"s. */
79 int n_rules
; /* Number of rules, including duplicates. */
80 unsigned int max_priority
; /* Max priority of any rule in the subtable. */
81 unsigned int max_count
; /* Count of max_priority rules. */
82 tag_type tag
; /* Tag generated from mask for partitioning. */
83 uint8_t n_indices
; /* How many indices to use. */
84 uint8_t index_ofs
[CLS_MAX_INDICES
]; /* u32 flow segment boundaries. */
85 struct hindex indices
[CLS_MAX_INDICES
]; /* Staged lookup indices. */
86 unsigned int trie_plen
[CLS_MAX_TRIES
]; /* Trie prefix length in 'mask'. */
88 struct trie_node
*ports_trie
; /* NULL if none. */
89 struct minimask mask
; /* Wildcards for fields. */
90 /* 'mask' must be the last field. */
93 /* Associates a metadata value (that is, a value of the OpenFlow 1.1+ metadata
94 * field) with tags for the "cls_subtable"s that contain rules that match that
96 struct cls_partition
{
97 struct hmap_node hmap_node
; /* In struct cls_classifier's 'partitions'
99 ovs_be64 metadata
; /* metadata value for this partition. */
100 tag_type tags
; /* OR of each flow's cls_subtable tag. */
101 struct tag_tracker tracker
; /* Tracks the bits in 'tags'. */
104 /* Internal representation of a rule in a "struct cls_subtable". */
106 struct cls_rule
*cls_rule
;
107 struct hindex_node index_nodes
[CLS_MAX_INDICES
]; /* Within subtable's
109 struct hmap_node hmap_node
; /* Within struct cls_subtable 'rules'. */
110 unsigned int priority
; /* Larger numbers are higher priorities. */
111 struct cls_partition
*partition
;
112 struct list list
; /* List of identical, lower-priority rules. */
113 struct miniflow flow
; /* Matching rule. Mask is in the subtable. */
114 /* 'flow' must be the last field. */
117 static struct cls_match
*
118 cls_match_alloc(struct cls_rule
*rule
)
120 int count
= count_1bits(rule
->match
.flow
.map
);
122 struct cls_match
*cls_match
123 = xmalloc(sizeof *cls_match
- sizeof cls_match
->flow
.inline_values
124 + MINIFLOW_VALUES_SIZE(count
));
126 cls_match
->cls_rule
= rule
;
127 miniflow_clone_inline(&cls_match
->flow
, &rule
->match
.flow
, count
);
128 cls_match
->priority
= rule
->priority
;
129 rule
->cls_match
= cls_match
;
134 static struct cls_subtable
*find_subtable(const struct cls_classifier
*,
135 const struct minimask
*);
136 static struct cls_subtable
*insert_subtable(struct cls_classifier
*,
137 const struct minimask
*);
139 static void destroy_subtable(struct cls_classifier
*, struct cls_subtable
*);
141 static void update_subtables_after_insertion(struct cls_classifier
*,
142 struct cls_subtable
*,
143 unsigned int new_priority
);
144 static void update_subtables_after_removal(struct cls_classifier
*,
145 struct cls_subtable
*,
146 unsigned int del_priority
);
148 static struct cls_match
*find_match_wc(const struct cls_subtable
*,
149 const struct flow
*, struct trie_ctx
*,
150 unsigned int n_tries
,
151 struct flow_wildcards
*);
152 static struct cls_match
*find_equal(struct cls_subtable
*,
153 const struct miniflow
*, uint32_t hash
);
154 static struct cls_match
*insert_rule(struct cls_classifier
*,
155 struct cls_subtable
*, struct cls_rule
*);
157 /* Iterates RULE over HEAD and all of the cls_rules on HEAD->list. */
158 #define FOR_EACH_RULE_IN_LIST(RULE, HEAD) \
159 for ((RULE) = (HEAD); (RULE) != NULL; (RULE) = next_rule_in_list(RULE))
160 #define FOR_EACH_RULE_IN_LIST_SAFE(RULE, NEXT, HEAD) \
161 for ((RULE) = (HEAD); \
162 (RULE) != NULL && ((NEXT) = next_rule_in_list(RULE), true); \
165 static struct cls_match
*next_rule_in_list__(struct cls_match
*);
166 static struct cls_match
*next_rule_in_list(struct cls_match
*);
168 static unsigned int minimask_get_prefix_len(const struct minimask
*,
169 const struct mf_field
*);
170 static void trie_init(struct cls_classifier
*, int trie_idx
,
171 const struct mf_field
*);
172 static unsigned int trie_lookup(const struct cls_trie
*, const struct flow
*,
173 unsigned int *checkbits
);
174 static unsigned int trie_lookup_value(const struct trie_node
*,
175 const ovs_be32 value
[],
176 unsigned int *checkbits
);
177 static void trie_destroy(struct trie_node
*);
178 static void trie_insert(struct cls_trie
*, const struct cls_rule
*, int mlen
);
179 static void trie_insert_prefix(struct trie_node
**, const ovs_be32
*prefix
,
181 static void trie_remove(struct cls_trie
*, const struct cls_rule
*, int mlen
);
182 static void trie_remove_prefix(struct trie_node
**, const ovs_be32
*prefix
,
184 static void mask_set_prefix_bits(struct flow_wildcards
*, uint8_t be32ofs
,
186 static bool mask_prefix_bits_set(const struct flow_wildcards
*,
187 uint8_t be32ofs
, unsigned int nbits
);
190 cls_subtables_init(struct cls_subtables
*subtables
)
192 memset(subtables
, 0, sizeof *subtables
);
196 cls_subtables_destroy(struct cls_subtables
*subtables
)
198 free(subtables
->array
);
199 memset(subtables
, 0, sizeof *subtables
);
202 /* Subtables insertion. */
204 cls_subtables_push_back(struct cls_subtables
*subtables
,
205 struct cls_subtable_entry a
)
207 if (subtables
->count
== subtables
->alloc_size
) {
208 subtables
->array
= x2nrealloc(subtables
->array
, &subtables
->alloc_size
,
212 subtables
->array
[subtables
->count
++] = a
;
215 /* Move subtable entry at 'from' to 'to', shifting the elements in between
216 * (including the one at 'to') accordingly. */
218 cls_subtables_move(struct cls_subtable_entry
*to
,
219 struct cls_subtable_entry
*from
)
222 struct cls_subtable_entry temp
= *from
;
225 /* Shift entries (from,to] backwards to make space at 'to'. */
226 memmove(from
, from
+ 1, (to
- from
) * sizeof *to
);
228 /* Shift entries [to,from) forward to make space at 'to'. */
229 memmove(to
+ 1, to
, (from
- to
) * sizeof *to
);
236 /* Subtables removal. */
238 cls_subtables_remove(struct cls_subtables
*subtables
,
239 struct cls_subtable_entry
*elem
)
241 ssize_t size
= (&subtables
->array
[subtables
->count
]
242 - (elem
+ 1)) * sizeof *elem
;
244 memmove(elem
, elem
+ 1, size
);
249 #define CLS_SUBTABLES_FOR_EACH(SUBTABLE, ITER, SUBTABLES) \
250 for ((ITER) = (SUBTABLES)->array; \
251 (ITER) < &(SUBTABLES)->array[(SUBTABLES)->count] \
252 && OVS_LIKELY((SUBTABLE) = (ITER)->subtable); \
254 #define CLS_SUBTABLES_FOR_EACH_CONTINUE(SUBTABLE, ITER, SUBTABLES) \
256 (ITER) < &(SUBTABLES)->array[(SUBTABLES)->count] \
257 && OVS_LIKELY((SUBTABLE) = (ITER)->subtable); \
259 #define CLS_SUBTABLES_FOR_EACH_REVERSE(SUBTABLE, ITER, SUBTABLES) \
260 for ((ITER) = &(SUBTABLES)->array[(SUBTABLES)->count]; \
261 (ITER) > (SUBTABLES)->array \
262 && OVS_LIKELY((SUBTABLE) = (--(ITER))->subtable);)
265 cls_subtables_verify(struct cls_subtables
*subtables
)
267 struct cls_subtable
*table
;
268 struct cls_subtable_entry
*iter
;
269 unsigned int priority
= 0;
271 CLS_SUBTABLES_FOR_EACH_REVERSE (table
, iter
, subtables
) {
272 if (iter
->max_priority
!= table
->max_priority
) {
273 VLOG_WARN("Subtable %p has mismatching priority in cache (%u != %u)",
274 table
, iter
->max_priority
, table
->max_priority
);
276 if (iter
->max_priority
< priority
) {
277 VLOG_WARN("Subtable cache is out of order (%u < %u)",
278 iter
->max_priority
, priority
);
280 priority
= iter
->max_priority
;
285 cls_subtables_reset(struct cls_classifier
*cls
)
287 struct cls_subtables old
= cls
->subtables
;
288 struct cls_subtable
*subtable
;
290 VLOG_WARN("Resetting subtable cache.");
292 cls_subtables_verify(&cls
->subtables
);
294 cls_subtables_init(&cls
->subtables
);
296 HMAP_FOR_EACH (subtable
, hmap_node
, &cls
->subtables_map
) {
297 struct cls_match
*head
;
298 struct cls_subtable_entry elem
;
299 struct cls_subtable
*table
;
300 struct cls_subtable_entry
*iter
, *from
= NULL
;
301 unsigned int new_max
= 0;
302 unsigned int max_count
= 0;
305 /* Verify max_priority. */
306 HMAP_FOR_EACH (head
, hmap_node
, &subtable
->rules
) {
307 if (head
->priority
> new_max
) {
308 new_max
= head
->priority
;
310 } else if (head
->priority
== new_max
) {
314 if (new_max
!= subtable
->max_priority
||
315 max_count
!= subtable
->max_count
) {
316 VLOG_WARN("subtable %p (%u rules) has mismatching max_priority "
317 "(%u) or max_count (%u). Highest priority found was %u, "
319 subtable
, subtable
->n_rules
, subtable
->max_priority
,
320 subtable
->max_count
, new_max
, max_count
);
321 subtable
->max_priority
= new_max
;
322 subtable
->max_count
= max_count
;
325 /* Locate the subtable from the old cache. */
327 CLS_SUBTABLES_FOR_EACH (table
, iter
, &old
) {
328 if (table
== subtable
) {
329 if (iter
->max_priority
!= new_max
) {
330 VLOG_WARN("Subtable %p has wrong max priority (%u != %u) "
332 subtable
, iter
->max_priority
, new_max
);
335 VLOG_WARN("Subtable %p duplicated in the old cache.",
342 VLOG_WARN("Subtable %p not found from the old cache.", subtable
);
345 elem
.subtable
= subtable
;
346 elem
.tag
= subtable
->tag
;
347 elem
.max_priority
= subtable
->max_priority
;
348 cls_subtables_push_back(&cls
->subtables
, elem
);
350 /* Possibly move 'subtable' earlier in the priority array. If
351 * we break out of the loop, then the subtable (at 'from')
352 * should be moved to the position right after the current
353 * element. If the loop terminates normally, then 'iter' will
354 * be at the first array element and we'll move the subtable
355 * to the front of the array. */
356 CLS_SUBTABLES_FOR_EACH_REVERSE (table
, iter
, &cls
->subtables
) {
357 if (table
== subtable
) {
358 from
= iter
; /* Locate the subtable as we go. */
359 } else if (table
->max_priority
>= new_max
) {
360 ovs_assert(from
!= NULL
);
361 iter
++; /* After this. */
366 /* Move subtable at 'from' to 'iter'. */
367 cls_subtables_move(iter
, from
);
370 /* Verify that the old and the new have the same size. */
371 if (old
.count
!= cls
->subtables
.count
) {
372 VLOG_WARN("subtables cache sizes differ: old (%"PRIuSIZE
373 ") != new (%"PRIuSIZE
").",
374 old
.count
, cls
->subtables
.count
);
377 cls_subtables_destroy(&old
);
379 cls_subtables_verify(&cls
->subtables
);
383 /* flow/miniflow/minimask/minimatch utilities.
384 * These are only used by the classifier, so place them here to allow
385 * for better optimization. */
387 static inline uint64_t
388 miniflow_get_map_in_range(const struct miniflow
*miniflow
,
389 uint8_t start
, uint8_t end
, unsigned int *offset
)
391 uint64_t map
= miniflow
->map
;
395 uint64_t msk
= (UINT64_C(1) << start
) - 1; /* 'start' LSBs set */
396 *offset
= count_1bits(map
& msk
);
399 if (end
< FLOW_U32S
) {
400 uint64_t msk
= (UINT64_C(1) << end
) - 1; /* 'end' LSBs set */
406 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
407 * 'mask', given 'basis'.
409 * The hash values returned by this function are the same as those returned by
410 * miniflow_hash_in_minimask(), only the form of the arguments differ. */
411 static inline uint32_t
412 flow_hash_in_minimask(const struct flow
*flow
, const struct minimask
*mask
,
415 const uint32_t *mask_values
= miniflow_get_u32_values(&mask
->masks
);
416 const uint32_t *flow_u32
= (const uint32_t *)flow
;
417 const uint32_t *p
= mask_values
;
422 for (map
= mask
->masks
.map
; map
; map
= zero_rightmost_1bit(map
)) {
423 hash
= mhash_add(hash
, flow_u32
[raw_ctz(map
)] & *p
++);
426 return mhash_finish(hash
, (p
- mask_values
) * 4);
429 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
430 * 'mask', given 'basis'.
432 * The hash values returned by this function are the same as those returned by
433 * flow_hash_in_minimask(), only the form of the arguments differ. */
434 static inline uint32_t
435 miniflow_hash_in_minimask(const struct miniflow
*flow
,
436 const struct minimask
*mask
, uint32_t basis
)
438 const uint32_t *mask_values
= miniflow_get_u32_values(&mask
->masks
);
439 const uint32_t *p
= mask_values
;
440 uint32_t hash
= basis
;
443 MINIFLOW_FOR_EACH_IN_MAP(flow_u32
, flow
, mask
->masks
.map
) {
444 hash
= mhash_add(hash
, flow_u32
& *p
++);
447 return mhash_finish(hash
, (p
- mask_values
) * 4);
450 /* Returns a hash value for the bits of range [start, end) in 'flow',
451 * where there are 1-bits in 'mask', given 'hash'.
453 * The hash values returned by this function are the same as those returned by
454 * minimatch_hash_range(), only the form of the arguments differ. */
455 static inline uint32_t
456 flow_hash_in_minimask_range(const struct flow
*flow
,
457 const struct minimask
*mask
,
458 uint8_t start
, uint8_t end
, uint32_t *basis
)
460 const uint32_t *mask_values
= miniflow_get_u32_values(&mask
->masks
);
461 const uint32_t *flow_u32
= (const uint32_t *)flow
;
463 uint64_t map
= miniflow_get_map_in_range(&mask
->masks
, start
, end
,
465 const uint32_t *p
= mask_values
+ offset
;
466 uint32_t hash
= *basis
;
468 for (; map
; map
= zero_rightmost_1bit(map
)) {
469 hash
= mhash_add(hash
, flow_u32
[raw_ctz(map
)] & *p
++);
472 *basis
= hash
; /* Allow continuation from the unfinished value. */
473 return mhash_finish(hash
, (p
- mask_values
) * 4);
476 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask. */
478 flow_wildcards_fold_minimask(struct flow_wildcards
*wc
,
479 const struct minimask
*mask
)
481 flow_union_with_miniflow(&wc
->masks
, &mask
->masks
);
484 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask
485 * in range [start, end). */
487 flow_wildcards_fold_minimask_range(struct flow_wildcards
*wc
,
488 const struct minimask
*mask
,
489 uint8_t start
, uint8_t end
)
491 uint32_t *dst_u32
= (uint32_t *)&wc
->masks
;
493 uint64_t map
= miniflow_get_map_in_range(&mask
->masks
, start
, end
,
495 const uint32_t *p
= miniflow_get_u32_values(&mask
->masks
) + offset
;
497 for (; map
; map
= zero_rightmost_1bit(map
)) {
498 dst_u32
[raw_ctz(map
)] |= *p
++;
502 /* Returns a hash value for 'flow', given 'basis'. */
503 static inline uint32_t
504 miniflow_hash(const struct miniflow
*flow
, uint32_t basis
)
506 const uint32_t *values
= miniflow_get_u32_values(flow
);
507 const uint32_t *p
= values
;
508 uint32_t hash
= basis
;
509 uint64_t hash_map
= 0;
512 for (map
= flow
->map
; map
; map
= zero_rightmost_1bit(map
)) {
514 hash
= mhash_add(hash
, *p
);
515 hash_map
|= rightmost_1bit(map
);
519 hash
= mhash_add(hash
, hash_map
);
520 hash
= mhash_add(hash
, hash_map
>> 32);
522 return mhash_finish(hash
, p
- values
);
525 /* Returns a hash value for 'mask', given 'basis'. */
526 static inline uint32_t
527 minimask_hash(const struct minimask
*mask
, uint32_t basis
)
529 return miniflow_hash(&mask
->masks
, basis
);
532 /* Returns a hash value for 'match', given 'basis'. */
533 static inline uint32_t
534 minimatch_hash(const struct minimatch
*match
, uint32_t basis
)
536 return miniflow_hash(&match
->flow
, minimask_hash(&match
->mask
, basis
));
539 /* Returns a hash value for the bits of range [start, end) in 'minimatch',
542 * The hash values returned by this function are the same as those returned by
543 * flow_hash_in_minimask_range(), only the form of the arguments differ. */
544 static inline uint32_t
545 minimatch_hash_range(const struct minimatch
*match
, uint8_t start
, uint8_t end
,
549 const uint32_t *p
, *q
;
550 uint32_t hash
= *basis
;
553 n
= count_1bits(miniflow_get_map_in_range(&match
->mask
.masks
, start
, end
,
555 q
= miniflow_get_u32_values(&match
->mask
.masks
) + offset
;
556 p
= miniflow_get_u32_values(&match
->flow
) + offset
;
558 for (i
= 0; i
< n
; i
++) {
559 hash
= mhash_add(hash
, p
[i
] & q
[i
]);
561 *basis
= hash
; /* Allow continuation from the unfinished value. */
562 return mhash_finish(hash
, (offset
+ n
) * 4);
568 /* Initializes 'rule' to match packets specified by 'match' at the given
569 * 'priority'. 'match' must satisfy the invariant described in the comment at
570 * the definition of struct match.
572 * The caller must eventually destroy 'rule' with cls_rule_destroy().
574 * (OpenFlow uses priorities between 0 and UINT16_MAX, inclusive, but
575 * internally Open vSwitch supports a wider range.) */
577 cls_rule_init(struct cls_rule
*rule
,
578 const struct match
*match
, unsigned int priority
)
580 minimatch_init(&rule
->match
, match
);
581 rule
->priority
= priority
;
582 rule
->cls_match
= NULL
;
585 /* Same as cls_rule_init() for initialization from a "struct minimatch". */
587 cls_rule_init_from_minimatch(struct cls_rule
*rule
,
588 const struct minimatch
*match
,
589 unsigned int priority
)
591 minimatch_clone(&rule
->match
, match
);
592 rule
->priority
= priority
;
593 rule
->cls_match
= NULL
;
596 /* Initializes 'dst' as a copy of 'src'.
598 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
600 cls_rule_clone(struct cls_rule
*dst
, const struct cls_rule
*src
)
602 minimatch_clone(&dst
->match
, &src
->match
);
603 dst
->priority
= src
->priority
;
604 dst
->cls_match
= NULL
;
607 /* Initializes 'dst' with the data in 'src', destroying 'src'.
609 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
611 cls_rule_move(struct cls_rule
*dst
, struct cls_rule
*src
)
613 minimatch_move(&dst
->match
, &src
->match
);
614 dst
->priority
= src
->priority
;
615 dst
->cls_match
= NULL
;
618 /* Frees memory referenced by 'rule'. Doesn't free 'rule' itself (it's
619 * normally embedded into a larger structure).
621 * ('rule' must not currently be in a classifier.) */
623 cls_rule_destroy(struct cls_rule
*rule
)
625 ovs_assert(!rule
->cls_match
);
626 minimatch_destroy(&rule
->match
);
629 /* Returns true if 'a' and 'b' match the same packets at the same priority,
630 * false if they differ in some way. */
632 cls_rule_equal(const struct cls_rule
*a
, const struct cls_rule
*b
)
634 return a
->priority
== b
->priority
&& minimatch_equal(&a
->match
, &b
->match
);
637 /* Returns a hash value for 'rule', folding in 'basis'. */
639 cls_rule_hash(const struct cls_rule
*rule
, uint32_t basis
)
641 return minimatch_hash(&rule
->match
, hash_int(rule
->priority
, basis
));
644 /* Appends a string describing 'rule' to 's'. */
646 cls_rule_format(const struct cls_rule
*rule
, struct ds
*s
)
648 minimatch_format(&rule
->match
, s
, rule
->priority
);
651 /* Returns true if 'rule' matches every packet, false otherwise. */
653 cls_rule_is_catchall(const struct cls_rule
*rule
)
655 return minimask_is_catchall(&rule
->match
.mask
);
658 /* Initializes 'cls' as a classifier that initially contains no classification
661 classifier_init(struct classifier
*cls_
, const uint8_t *flow_segments
)
663 struct cls_classifier
*cls
= xmalloc(sizeof *cls
);
665 fat_rwlock_init(&cls_
->rwlock
);
670 hmap_init(&cls
->subtables_map
);
671 cls_subtables_init(&cls
->subtables
);
672 hmap_init(&cls
->partitions
);
673 cls
->n_flow_segments
= 0;
675 while (cls
->n_flow_segments
< CLS_MAX_INDICES
676 && *flow_segments
< FLOW_U32S
) {
677 cls
->flow_segments
[cls
->n_flow_segments
++] = *flow_segments
++;
683 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
684 * caller's responsibility. */
686 classifier_destroy(struct classifier
*cls_
)
689 struct cls_classifier
*cls
= cls_
->cls
;
690 struct cls_subtable
*partition
, *next_partition
;
691 struct cls_subtable
*subtable
, *next_subtable
;
694 fat_rwlock_destroy(&cls_
->rwlock
);
699 for (i
= 0; i
< cls
->n_tries
; i
++) {
700 trie_destroy(cls
->tries
[i
].root
);
703 HMAP_FOR_EACH_SAFE (subtable
, next_subtable
, hmap_node
,
704 &cls
->subtables_map
) {
705 destroy_subtable(cls
, subtable
);
707 hmap_destroy(&cls
->subtables_map
);
709 HMAP_FOR_EACH_SAFE (partition
, next_partition
, hmap_node
,
711 hmap_remove(&cls
->partitions
, &partition
->hmap_node
);
714 hmap_destroy(&cls
->partitions
);
716 cls_subtables_destroy(&cls
->subtables
);
721 /* We use uint64_t as a set for the fields below. */
722 BUILD_ASSERT_DECL(MFF_N_IDS
<= 64);
724 /* Set the fields for which prefix lookup should be performed. */
726 classifier_set_prefix_fields(struct classifier
*cls_
,
727 const enum mf_field_id
*trie_fields
,
728 unsigned int n_fields
)
730 struct cls_classifier
*cls
= cls_
->cls
;
734 for (i
= 0, trie
= 0; i
< n_fields
&& trie
< CLS_MAX_TRIES
; i
++) {
735 const struct mf_field
*field
= mf_from_id(trie_fields
[i
]);
736 if (field
->flow_be32ofs
< 0 || field
->n_bits
% 32) {
737 /* Incompatible field. This is the only place where we
738 * enforce these requirements, but the rest of the trie code
739 * depends on the flow_be32ofs to be non-negative and the
740 * field length to be a multiple of 32 bits. */
744 if (fields
& (UINT64_C(1) << trie_fields
[i
])) {
745 /* Duplicate field, there is no need to build more than
746 * one index for any one field. */
749 fields
|= UINT64_C(1) << trie_fields
[i
];
751 if (trie
>= cls
->n_tries
|| field
!= cls
->tries
[trie
].field
) {
752 trie_init(cls
, trie
, field
);
757 /* Destroy the rest. */
758 for (i
= trie
; i
< cls
->n_tries
; i
++) {
759 trie_init(cls
, i
, NULL
);
765 trie_init(struct cls_classifier
*cls
, int trie_idx
,
766 const struct mf_field
*field
)
768 struct cls_trie
*trie
= &cls
->tries
[trie_idx
];
769 struct cls_subtable
*subtable
;
770 struct cls_subtable_entry
*iter
;
772 if (trie_idx
< cls
->n_tries
) {
773 trie_destroy(trie
->root
);
778 /* Add existing rules to the trie. */
779 CLS_SUBTABLES_FOR_EACH (subtable
, iter
, &cls
->subtables
) {
782 plen
= field
? minimask_get_prefix_len(&subtable
->mask
, field
) : 0;
783 /* Initialize subtable's prefix length on this field. */
784 subtable
->trie_plen
[trie_idx
] = plen
;
787 struct cls_match
*head
;
789 HMAP_FOR_EACH (head
, hmap_node
, &subtable
->rules
) {
790 struct cls_match
*match
;
792 FOR_EACH_RULE_IN_LIST (match
, head
) {
793 trie_insert(trie
, match
->cls_rule
, plen
);
800 /* Returns true if 'cls' contains no classification rules, false otherwise. */
802 classifier_is_empty(const struct classifier
*cls
)
804 return cls
->cls
->n_rules
== 0;
807 /* Returns the number of rules in 'cls'. */
809 classifier_count(const struct classifier
*cls
)
811 return cls
->cls
->n_rules
;
815 hash_metadata(ovs_be64 metadata_
)
817 uint64_t metadata
= (OVS_FORCE
uint64_t) metadata_
;
818 return hash_uint64(metadata
);
821 static struct cls_partition
*
822 find_partition(const struct cls_classifier
*cls
, ovs_be64 metadata
,
825 struct cls_partition
*partition
;
827 HMAP_FOR_EACH_IN_BUCKET (partition
, hmap_node
, hash
, &cls
->partitions
) {
828 if (partition
->metadata
== metadata
) {
836 static struct cls_partition
*
837 create_partition(struct cls_classifier
*cls
, struct cls_subtable
*subtable
,
840 uint32_t hash
= hash_metadata(metadata
);
841 struct cls_partition
*partition
= find_partition(cls
, metadata
, hash
);
843 partition
= xmalloc(sizeof *partition
);
844 partition
->metadata
= metadata
;
846 tag_tracker_init(&partition
->tracker
);
847 hmap_insert(&cls
->partitions
, &partition
->hmap_node
, hash
);
849 tag_tracker_add(&partition
->tracker
, &partition
->tags
, subtable
->tag
);
853 static inline ovs_be32
minimatch_get_ports(const struct minimatch
*match
)
855 /* Could optimize to use the same map if needed for fast path. */
856 return MINIFLOW_GET_BE32(&match
->flow
, tp_src
)
857 & MINIFLOW_GET_BE32(&match
->mask
.masks
, tp_src
);
860 /* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller
861 * must not modify or free it.
863 * If 'cls' already contains an identical rule (including wildcards, values of
864 * fixed fields, and priority), replaces the old rule by 'rule' and returns the
865 * rule that was replaced. The caller takes ownership of the returned rule and
866 * is thus responsible for destroying it with cls_rule_destroy(), freeing the
867 * memory block in which it resides, etc., as necessary.
869 * Returns NULL if 'cls' does not contain a rule with an identical key, after
870 * inserting the new rule. In this case, no rules are displaced by the new
871 * rule, even rules that cannot have any effect because the new rule matches a
872 * superset of their flows and has higher priority. */
874 classifier_replace(struct classifier
*cls_
, struct cls_rule
*rule
)
876 struct cls_classifier
*cls
= cls_
->cls
;
877 struct cls_match
*old_rule
;
878 struct cls_subtable
*subtable
;
880 subtable
= find_subtable(cls
, &rule
->match
.mask
);
882 subtable
= insert_subtable(cls
, &rule
->match
.mask
);
885 old_rule
= insert_rule(cls
, subtable
, rule
);
889 rule
->cls_match
->partition
= NULL
;
890 if (minimask_get_metadata_mask(&rule
->match
.mask
) == OVS_BE64_MAX
) {
891 ovs_be64 metadata
= miniflow_get_metadata(&rule
->match
.flow
);
892 rule
->cls_match
->partition
= create_partition(cls
, subtable
,
899 for (i
= 0; i
< cls
->n_tries
; i
++) {
900 if (subtable
->trie_plen
[i
]) {
901 trie_insert(&cls
->tries
[i
], rule
, subtable
->trie_plen
[i
]);
906 if (subtable
->ports_mask_len
) {
907 /* We mask the value to be inserted to always have the wildcarded
908 * bits in known (zero) state, so we can include them in comparison
909 * and they will always match (== their original value does not
911 ovs_be32 masked_ports
= minimatch_get_ports(&rule
->match
);
913 trie_insert_prefix(&subtable
->ports_trie
, &masked_ports
,
914 subtable
->ports_mask_len
);
919 struct cls_rule
*old_cls_rule
= old_rule
->cls_rule
;
921 rule
->cls_match
->partition
= old_rule
->partition
;
922 old_cls_rule
->cls_match
= NULL
;
928 /* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller
929 * must not modify or free it.
931 * 'cls' must not contain an identical rule (including wildcards, values of
932 * fixed fields, and priority). Use classifier_find_rule_exactly() to find
935 classifier_insert(struct classifier
*cls
, struct cls_rule
*rule
)
937 struct cls_rule
*displaced_rule
= classifier_replace(cls
, rule
);
938 ovs_assert(!displaced_rule
);
941 /* Removes 'rule' from 'cls'. It is the caller's responsibility to destroy
942 * 'rule' with cls_rule_destroy(), freeing the memory block in which 'rule'
943 * resides, etc., as necessary. */
945 classifier_remove(struct classifier
*cls_
, struct cls_rule
*rule
)
947 struct cls_classifier
*cls
= cls_
->cls
;
948 struct cls_partition
*partition
;
949 struct cls_match
*cls_match
= rule
->cls_match
;
950 struct cls_match
*head
;
951 struct cls_subtable
*subtable
;
954 ovs_assert(cls_match
);
956 subtable
= find_subtable(cls
, &rule
->match
.mask
);
957 ovs_assert(subtable
);
959 if (subtable
->ports_mask_len
) {
960 ovs_be32 masked_ports
= minimatch_get_ports(&rule
->match
);
962 trie_remove_prefix(&subtable
->ports_trie
,
963 &masked_ports
, subtable
->ports_mask_len
);
965 for (i
= 0; i
< cls
->n_tries
; i
++) {
966 if (subtable
->trie_plen
[i
]) {
967 trie_remove(&cls
->tries
[i
], rule
, subtable
->trie_plen
[i
]);
971 /* Remove rule node from indices. */
972 for (i
= 0; i
< subtable
->n_indices
; i
++) {
973 hindex_remove(&subtable
->indices
[i
], &cls_match
->index_nodes
[i
]);
976 head
= find_equal(subtable
, &rule
->match
.flow
, cls_match
->hmap_node
.hash
);
977 if (head
!= cls_match
) {
978 list_remove(&cls_match
->list
);
979 } else if (list_is_empty(&cls_match
->list
)) {
980 hmap_remove(&subtable
->rules
, &cls_match
->hmap_node
);
982 struct cls_match
*next
= CONTAINER_OF(cls_match
->list
.next
,
983 struct cls_match
, list
);
985 list_remove(&cls_match
->list
);
986 hmap_replace(&subtable
->rules
, &cls_match
->hmap_node
,
990 partition
= cls_match
->partition
;
992 tag_tracker_subtract(&partition
->tracker
, &partition
->tags
,
994 if (!partition
->tags
) {
995 hmap_remove(&cls
->partitions
, &partition
->hmap_node
);
1000 if (--subtable
->n_rules
== 0) {
1001 destroy_subtable(cls
, subtable
);
1003 update_subtables_after_removal(cls
, subtable
, cls_match
->priority
);
1008 rule
->cls_match
= NULL
;
1012 /* Prefix tree context. Valid when 'lookup_done' is true. Can skip all
1013 * subtables which have more than 'match_plen' bits in their corresponding
1014 * field at offset 'be32ofs'. If skipped, 'maskbits' prefix bits should be
1015 * unwildcarded to quarantee datapath flow matches only packets it should. */
1017 const struct cls_trie
*trie
;
1018 bool lookup_done
; /* Status of the lookup. */
1019 uint8_t be32ofs
; /* U32 offset of the field in question. */
1020 unsigned int match_plen
; /* Longest prefix than could possibly match. */
1021 unsigned int maskbits
; /* Prefix length needed to avoid false matches. */
1025 trie_ctx_init(struct trie_ctx
*ctx
, const struct cls_trie
*trie
)
1028 ctx
->be32ofs
= trie
->field
->flow_be32ofs
;
1029 ctx
->lookup_done
= false;
1033 lookahead_subtable(const struct cls_subtable_entry
*subtables
)
1035 ovs_prefetch_range(subtables
->subtable
, sizeof *subtables
->subtable
);
1038 /* Finds and returns the highest-priority rule in 'cls' that matches 'flow'.
1039 * Returns a null pointer if no rules in 'cls' match 'flow'. If multiple rules
1040 * of equal priority match 'flow', returns one arbitrarily.
1042 * If a rule is found and 'wc' is non-null, bitwise-OR's 'wc' with the
1043 * set of bits that were significant in the lookup. At some point
1044 * earlier, 'wc' should have been initialized (e.g., by
1045 * flow_wildcards_init_catchall()). */
1047 classifier_lookup(const struct classifier
*cls_
, const struct flow
*flow
,
1048 struct flow_wildcards
*wc
)
1050 struct cls_classifier
*cls
= cls_
->cls
;
1051 const struct cls_partition
*partition
;
1053 struct cls_match
*best
;
1054 struct trie_ctx trie_ctx
[CLS_MAX_TRIES
];
1056 struct cls_subtable_entry
*subtables
= cls
->subtables
.array
;
1057 int n_subtables
= cls
->subtables
.count
;
1058 int64_t best_priority
= -1;
1060 /* Prefetch the subtables array. */
1061 ovs_prefetch_range(subtables
, n_subtables
* sizeof *subtables
);
1063 /* Determine 'tags' such that, if 'subtable->tag' doesn't intersect them,
1064 * then 'flow' cannot possibly match in 'subtable':
1066 * - If flow->metadata maps to a given 'partition', then we can use
1067 * 'tags' for 'partition->tags'.
1069 * - If flow->metadata has no partition, then no rule in 'cls' has an
1070 * exact-match for flow->metadata. That means that we don't need to
1071 * search any subtable that includes flow->metadata in its mask.
1073 * In either case, we always need to search any cls_subtables that do not
1074 * include flow->metadata in its mask. One way to do that would be to
1075 * check the "cls_subtable"s explicitly for that, but that would require an
1076 * extra branch per subtable. Instead, we mark such a cls_subtable's
1077 * 'tags' as TAG_ALL and make sure that 'tags' is never empty. This means
1078 * that 'tags' always intersects such a cls_subtable's 'tags', so we don't
1079 * need a special case.
1081 partition
= (hmap_is_empty(&cls
->partitions
)
1083 : find_partition(cls
, flow
->metadata
,
1084 hash_metadata(flow
->metadata
)));
1085 tags
= partition
? partition
->tags
: TAG_ARBITRARY
;
1087 /* Initialize trie contexts for match_find_wc(). */
1088 for (i
= 0; i
< cls
->n_tries
; i
++) {
1089 trie_ctx_init(&trie_ctx
[i
], &cls
->tries
[i
]);
1092 /* Prefetch the first subtables. */
1093 if (n_subtables
> 1) {
1094 lookahead_subtable(subtables
);
1095 lookahead_subtable(subtables
+ 1);
1099 for (i
= 0; OVS_LIKELY(i
< n_subtables
); i
++) {
1100 struct cls_match
*rule
;
1102 if ((int64_t)subtables
[i
].max_priority
<= best_priority
) {
1103 /* Subtables are in descending priority order,
1104 * can not find anything better. */
1108 /* Prefetch a forthcoming subtable. */
1109 if (i
+ 2 < n_subtables
) {
1110 lookahead_subtable(&subtables
[i
+ 2]);
1113 if (!tag_intersects(tags
, subtables
[i
].tag
)) {
1117 rule
= find_match_wc(subtables
[i
].subtable
, flow
, trie_ctx
,
1119 if (rule
&& (int64_t)rule
->priority
> best_priority
) {
1120 best_priority
= (int64_t)rule
->priority
;
1125 return best
? best
->cls_rule
: NULL
;
1128 /* Returns true if 'target' satisifies 'match', that is, if each bit for which
1129 * 'match' specifies a particular value has the correct value in 'target'.
1131 * 'flow' and 'mask' have the same mask! */
1133 miniflow_and_mask_matches_miniflow(const struct miniflow
*flow
,
1134 const struct minimask
*mask
,
1135 const struct miniflow
*target
)
1137 const uint32_t *flowp
= miniflow_get_u32_values(flow
);
1138 const uint32_t *maskp
= miniflow_get_u32_values(&mask
->masks
);
1139 uint32_t target_u32
;
1141 MINIFLOW_FOR_EACH_IN_MAP(target_u32
, target
, mask
->masks
.map
) {
1142 if ((*flowp
++ ^ target_u32
) & *maskp
++) {
1150 static inline struct cls_match
*
1151 find_match_miniflow(const struct cls_subtable
*subtable
,
1152 const struct miniflow
*flow
,
1155 struct cls_match
*rule
;
1157 HMAP_FOR_EACH_WITH_HASH (rule
, hmap_node
, hash
, &subtable
->rules
) {
1158 if (miniflow_and_mask_matches_miniflow(&rule
->flow
, &subtable
->mask
,
1167 /* Finds and returns the highest-priority rule in 'cls' that matches
1168 * 'miniflow'. Returns a null pointer if no rules in 'cls' match 'flow'.
1169 * If multiple rules of equal priority match 'flow', returns one arbitrarily.
1171 * This function is optimized for the userspace datapath, which only ever has
1172 * one priority value for it's flows!
1174 struct cls_rule
*classifier_lookup_miniflow_first(const struct classifier
*cls_
,
1175 const struct miniflow
*flow
)
1177 struct cls_classifier
*cls
= cls_
->cls
;
1178 struct cls_subtable
*subtable
;
1179 struct cls_subtable_entry
*iter
;
1181 CLS_SUBTABLES_FOR_EACH (subtable
, iter
, &cls
->subtables
) {
1182 struct cls_match
*rule
;
1184 rule
= find_match_miniflow(subtable
, flow
,
1185 miniflow_hash_in_minimask(flow
,
1189 return rule
->cls_rule
;
1196 /* Finds and returns a rule in 'cls' with exactly the same priority and
1197 * matching criteria as 'target'. Returns a null pointer if 'cls' doesn't
1198 * contain an exact match. */
1200 classifier_find_rule_exactly(const struct classifier
*cls_
,
1201 const struct cls_rule
*target
)
1203 struct cls_classifier
*cls
= cls_
->cls
;
1204 struct cls_match
*head
, *rule
;
1205 struct cls_subtable
*subtable
;
1207 subtable
= find_subtable(cls
, &target
->match
.mask
);
1212 /* Skip if there is no hope. */
1213 if (target
->priority
> subtable
->max_priority
) {
1217 head
= find_equal(subtable
, &target
->match
.flow
,
1218 miniflow_hash_in_minimask(&target
->match
.flow
,
1219 &target
->match
.mask
, 0));
1220 FOR_EACH_RULE_IN_LIST (rule
, head
) {
1221 if (target
->priority
>= rule
->priority
) {
1222 return target
->priority
== rule
->priority
? rule
->cls_rule
: NULL
;
1228 /* Finds and returns a rule in 'cls' with priority 'priority' and exactly the
1229 * same matching criteria as 'target'. Returns a null pointer if 'cls' doesn't
1230 * contain an exact match. */
1232 classifier_find_match_exactly(const struct classifier
*cls
,
1233 const struct match
*target
,
1234 unsigned int priority
)
1236 struct cls_rule
*retval
;
1239 cls_rule_init(&cr
, target
, priority
);
1240 retval
= classifier_find_rule_exactly(cls
, &cr
);
1241 cls_rule_destroy(&cr
);
1246 /* Checks if 'target' would overlap any other rule in 'cls'. Two rules are
1247 * considered to overlap if both rules have the same priority and a packet
1248 * could match both. */
1250 classifier_rule_overlaps(const struct classifier
*cls_
,
1251 const struct cls_rule
*target
)
1253 struct cls_classifier
*cls
= cls_
->cls
;
1254 struct cls_subtable
*subtable
;
1255 struct cls_subtable_entry
*iter
;
1257 /* Iterate subtables in the descending max priority order. */
1258 CLS_SUBTABLES_FOR_EACH (subtable
, iter
, &cls
->subtables
) {
1259 uint32_t storage
[FLOW_U32S
];
1260 struct minimask mask
;
1261 struct cls_match
*head
;
1263 if (target
->priority
> iter
->max_priority
) {
1264 break; /* Can skip this and the rest of the subtables. */
1267 minimask_combine(&mask
, &target
->match
.mask
, &subtable
->mask
, storage
);
1268 HMAP_FOR_EACH (head
, hmap_node
, &subtable
->rules
) {
1269 struct cls_match
*rule
;
1271 FOR_EACH_RULE_IN_LIST (rule
, head
) {
1272 if (rule
->priority
< target
->priority
) {
1273 break; /* Rules in descending priority order. */
1275 if (rule
->priority
== target
->priority
1276 && miniflow_equal_in_minimask(&target
->match
.flow
,
1277 &rule
->flow
, &mask
)) {
1287 /* Returns true if 'rule' exactly matches 'criteria' or if 'rule' is more
1288 * specific than 'criteria'. That is, 'rule' matches 'criteria' and this
1289 * function returns true if, for every field:
1291 * - 'criteria' and 'rule' specify the same (non-wildcarded) value for the
1294 * - 'criteria' wildcards the field,
1296 * Conversely, 'rule' does not match 'criteria' and this function returns false
1297 * if, for at least one field:
1299 * - 'criteria' and 'rule' specify different values for the field, or
1301 * - 'criteria' specifies a value for the field but 'rule' wildcards it.
1303 * Equivalently, the truth table for whether a field matches is:
1308 * r +---------+---------+
1309 * i wild | yes | yes |
1311 * e +---------+---------+
1312 * r exact | no |if values|
1314 * a +---------+---------+
1316 * This is the matching rule used by OpenFlow 1.0 non-strict OFPT_FLOW_MOD
1317 * commands and by OpenFlow 1.0 aggregate and flow stats.
1319 * Ignores rule->priority. */
1321 cls_rule_is_loose_match(const struct cls_rule
*rule
,
1322 const struct minimatch
*criteria
)
1324 return (!minimask_has_extra(&rule
->match
.mask
, &criteria
->mask
)
1325 && miniflow_equal_in_minimask(&rule
->match
.flow
, &criteria
->flow
,
1332 rule_matches(const struct cls_match
*rule
, const struct cls_rule
*target
)
1335 || miniflow_equal_in_minimask(&rule
->flow
,
1336 &target
->match
.flow
,
1337 &target
->match
.mask
));
1340 static struct cls_match
*
1341 search_subtable(const struct cls_subtable
*subtable
,
1342 const struct cls_rule
*target
)
1344 if (!target
|| !minimask_has_extra(&subtable
->mask
, &target
->match
.mask
)) {
1345 struct cls_match
*rule
;
1347 HMAP_FOR_EACH (rule
, hmap_node
, &subtable
->rules
) {
1348 if (rule_matches(rule
, target
)) {
1356 /* Initializes 'cursor' for iterating through rules in 'cls':
1358 * - If 'target' is null, the cursor will visit every rule in 'cls'.
1360 * - If 'target' is nonnull, the cursor will visit each 'rule' in 'cls'
1361 * such that cls_rule_is_loose_match(rule, target) returns true.
1363 * Ignores target->priority. */
1365 cls_cursor_init(struct cls_cursor
*cursor
, const struct classifier
*cls
,
1366 const struct cls_rule
*target
)
1368 cursor
->cls
= cls
->cls
;
1369 cursor
->target
= target
&& !cls_rule_is_catchall(target
) ? target
: NULL
;
1372 /* Returns the first matching cls_rule in 'cursor''s iteration, or a null
1373 * pointer if there are no matches. */
1375 cls_cursor_first(struct cls_cursor
*cursor
)
1377 struct cls_subtable
*subtable
;
1379 HMAP_FOR_EACH (subtable
, hmap_node
, &cursor
->cls
->subtables_map
) {
1380 struct cls_match
*rule
= search_subtable(subtable
, cursor
->target
);
1382 cursor
->subtable
= subtable
;
1383 return rule
->cls_rule
;
1390 /* Returns the next matching cls_rule in 'cursor''s iteration, or a null
1391 * pointer if there are no more matches. */
1393 cls_cursor_next(struct cls_cursor
*cursor
, const struct cls_rule
*rule_
)
1395 struct cls_match
*rule
= CONST_CAST(struct cls_match
*, rule_
->cls_match
);
1396 const struct cls_subtable
*subtable
;
1397 struct cls_match
*next
;
1399 next
= next_rule_in_list__(rule
);
1400 if (next
->priority
< rule
->priority
) {
1401 return next
->cls_rule
;
1404 /* 'next' is the head of the list, that is, the rule that is included in
1405 * the subtable's hmap. (This is important when the classifier contains
1406 * rules that differ only in priority.) */
1408 HMAP_FOR_EACH_CONTINUE (rule
, hmap_node
, &cursor
->subtable
->rules
) {
1409 if (rule_matches(rule
, cursor
->target
)) {
1410 return rule
->cls_rule
;
1414 subtable
= cursor
->subtable
;
1415 HMAP_FOR_EACH_CONTINUE (subtable
, hmap_node
, &cursor
->cls
->subtables_map
) {
1416 rule
= search_subtable(subtable
, cursor
->target
);
1418 cursor
->subtable
= subtable
;
1419 return rule
->cls_rule
;
1426 static struct cls_subtable
*
1427 find_subtable(const struct cls_classifier
*cls
, const struct minimask
*mask
)
1429 struct cls_subtable
*subtable
;
1431 HMAP_FOR_EACH_IN_BUCKET (subtable
, hmap_node
, minimask_hash(mask
, 0),
1432 &cls
->subtables_map
) {
1433 if (minimask_equal(mask
, &subtable
->mask
)) {
1440 static struct cls_subtable
*
1441 insert_subtable(struct cls_classifier
*cls
, const struct minimask
*mask
)
1443 uint32_t hash
= minimask_hash(mask
, 0);
1444 struct cls_subtable
*subtable
;
1446 struct flow_wildcards old
, new;
1448 struct cls_subtable_entry elem
;
1449 int count
= count_1bits(mask
->masks
.map
);
1451 subtable
= xzalloc(sizeof *subtable
- sizeof mask
->masks
.inline_values
1452 + MINIFLOW_VALUES_SIZE(count
));
1453 hmap_init(&subtable
->rules
);
1454 miniflow_clone_inline(&subtable
->mask
.masks
, &mask
->masks
, count
);
1456 /* Init indices for segmented lookup, if any. */
1457 flow_wildcards_init_catchall(&new);
1460 for (i
= 0; i
< cls
->n_flow_segments
; i
++) {
1461 flow_wildcards_fold_minimask_range(&new, mask
, prev
,
1462 cls
->flow_segments
[i
]);
1463 /* Add an index if it adds mask bits. */
1464 if (!flow_wildcards_equal(&new, &old
)) {
1465 hindex_init(&subtable
->indices
[index
]);
1466 subtable
->index_ofs
[index
] = cls
->flow_segments
[i
];
1470 prev
= cls
->flow_segments
[i
];
1472 /* Check if the rest of the subtable's mask adds any bits,
1473 * and remove the last index if it doesn't. */
1475 flow_wildcards_fold_minimask_range(&new, mask
, prev
, FLOW_U32S
);
1476 if (flow_wildcards_equal(&new, &old
)) {
1478 subtable
->index_ofs
[index
] = 0;
1479 hindex_destroy(&subtable
->indices
[index
]);
1482 subtable
->n_indices
= index
;
1484 subtable
->tag
= (minimask_get_metadata_mask(mask
) == OVS_BE64_MAX
1485 ? tag_create_deterministic(hash
)
1488 for (i
= 0; i
< cls
->n_tries
; i
++) {
1489 subtable
->trie_plen
[i
] = minimask_get_prefix_len(mask
,
1490 cls
->tries
[i
].field
);
1494 subtable
->ports_trie
= NULL
;
1495 subtable
->ports_mask_len
1496 = 32 - ctz32(ntohl(MINIFLOW_GET_BE32(&mask
->masks
, tp_src
)));
1498 hmap_insert(&cls
->subtables_map
, &subtable
->hmap_node
, hash
);
1499 elem
.subtable
= subtable
;
1500 elem
.tag
= subtable
->tag
;
1501 elem
.max_priority
= subtable
->max_priority
;
1502 cls_subtables_push_back(&cls
->subtables
, elem
);
1508 destroy_subtable(struct cls_classifier
*cls
, struct cls_subtable
*subtable
)
1511 struct cls_subtable
*table
= NULL
;
1512 struct cls_subtable_entry
*iter
;
1514 CLS_SUBTABLES_FOR_EACH (table
, iter
, &cls
->subtables
) {
1515 if (table
== subtable
) {
1516 cls_subtables_remove(&cls
->subtables
, iter
);
1521 trie_destroy(subtable
->ports_trie
);
1523 for (i
= 0; i
< subtable
->n_indices
; i
++) {
1524 hindex_destroy(&subtable
->indices
[i
]);
1526 minimask_destroy(&subtable
->mask
);
1527 hmap_remove(&cls
->subtables_map
, &subtable
->hmap_node
);
1528 hmap_destroy(&subtable
->rules
);
1532 /* This function performs the following updates for 'subtable' in 'cls'
1533 * following the addition of a new rule with priority 'new_priority' to
1536 * - Update 'subtable->max_priority' and 'subtable->max_count' if necessary.
1538 * - Update 'subtable''s position in 'cls->subtables' if necessary.
1540 * This function should only be called after adding a new rule, not after
1541 * replacing a rule by an identical one or modifying a rule in-place. */
1543 update_subtables_after_insertion(struct cls_classifier
*cls
,
1544 struct cls_subtable
*subtable
,
1545 unsigned int new_priority
)
1547 if (new_priority
== subtable
->max_priority
) {
1548 ++subtable
->max_count
;
1549 } else if (new_priority
> subtable
->max_priority
) {
1550 struct cls_subtable
*table
;
1551 struct cls_subtable_entry
*iter
, *from
= NULL
;
1553 subtable
->max_priority
= new_priority
;
1554 subtable
->max_count
= 1;
1556 /* Possibly move 'subtable' earlier in the priority array. If
1557 * we break out of the loop, then the subtable (at 'from')
1558 * should be moved to the position right after the current
1559 * element. If the loop terminates normally, then 'iter' will
1560 * be at the first array element and we'll move the subtable
1561 * to the front of the array. */
1562 CLS_SUBTABLES_FOR_EACH_REVERSE (table
, iter
, &cls
->subtables
) {
1563 if (table
== subtable
) {
1564 from
= iter
; /* Locate the subtable as we go. */
1565 iter
->max_priority
= new_priority
;
1566 } else if (table
->max_priority
>= new_priority
) {
1568 /* Corrupted cache? */
1569 cls_subtables_reset(cls
);
1570 VLOG_ABORT("update_subtables_after_insertion(): Subtable priority list corrupted.");
1573 iter
++; /* After this. */
1578 /* Move subtable at 'from' to 'iter'. */
1579 cls_subtables_move(iter
, from
);
1583 /* This function performs the following updates for 'subtable' in 'cls'
1584 * following the deletion of a rule with priority 'del_priority' from
1587 * - Update 'subtable->max_priority' and 'subtable->max_count' if necessary.
1589 * - Update 'subtable''s position in 'cls->subtables' if necessary.
1591 * This function should only be called after removing a rule, not after
1592 * replacing a rule by an identical one or modifying a rule in-place. */
1594 update_subtables_after_removal(struct cls_classifier
*cls
,
1595 struct cls_subtable
*subtable
,
1596 unsigned int del_priority
)
1598 if (del_priority
== subtable
->max_priority
&& --subtable
->max_count
== 0) {
1599 struct cls_match
*head
;
1600 struct cls_subtable
*table
;
1601 struct cls_subtable_entry
*iter
, *from
= NULL
;
1603 subtable
->max_priority
= 0;
1604 HMAP_FOR_EACH (head
, hmap_node
, &subtable
->rules
) {
1605 if (head
->priority
> subtable
->max_priority
) {
1606 subtable
->max_priority
= head
->priority
;
1607 subtable
->max_count
= 1;
1608 } else if (head
->priority
== subtable
->max_priority
) {
1609 ++subtable
->max_count
;
1613 /* Possibly move 'subtable' later in the priority array.
1614 * After the loop the 'iter' will point right after the position
1615 * at which the subtable should be moved (either at a subtable
1616 * with an equal or lower priority, or just past the array),
1617 * so it is decremented once. */
1618 CLS_SUBTABLES_FOR_EACH (table
, iter
, &cls
->subtables
) {
1619 if (table
== subtable
) {
1620 from
= iter
; /* Locate the subtable as we go. */
1621 iter
->max_priority
= subtable
->max_priority
;
1622 } else if (table
->max_priority
<= subtable
->max_priority
) {
1624 /* Corrupted cache? */
1625 cls_subtables_reset(cls
);
1626 VLOG_ABORT("update_subtables_after_removal(): Subtable priority list corrupted.");
1632 /* Now at one past the destination. */
1635 /* Move subtable at 'from' to 'iter'. */
1636 cls_subtables_move(iter
, from
);
1645 /* Return 'true' if can skip rest of the subtable based on the prefix trie
1646 * lookup results. */
1648 check_tries(struct trie_ctx trie_ctx
[CLS_MAX_TRIES
], unsigned int n_tries
,
1649 const unsigned int field_plen
[CLS_MAX_TRIES
],
1650 const struct range ofs
, const struct flow
*flow
,
1651 struct flow_wildcards
*wc
)
1655 /* Check if we could avoid fully unwildcarding the next level of
1656 * fields using the prefix tries. The trie checks are done only as
1657 * needed to avoid folding in additional bits to the wildcards mask. */
1658 for (j
= 0; j
< n_tries
; j
++) {
1659 /* Is the trie field relevant for this subtable? */
1660 if (field_plen
[j
]) {
1661 struct trie_ctx
*ctx
= &trie_ctx
[j
];
1662 uint8_t be32ofs
= ctx
->be32ofs
;
1664 /* Is the trie field within the current range of fields? */
1665 if (be32ofs
>= ofs
.start
&& be32ofs
< ofs
.end
) {
1666 /* On-demand trie lookup. */
1667 if (!ctx
->lookup_done
) {
1668 ctx
->match_plen
= trie_lookup(ctx
->trie
, flow
,
1670 ctx
->lookup_done
= true;
1672 /* Possible to skip the rest of the subtable if subtable's
1673 * prefix on the field is longer than what is known to match
1674 * based on the trie lookup. */
1675 if (field_plen
[j
] > ctx
->match_plen
) {
1676 /* RFC: We want the trie lookup to never result in
1677 * unwildcarding any bits that would not be unwildcarded
1678 * otherwise. Since the trie is shared by the whole
1679 * classifier, it is possible that the 'maskbits' contain
1680 * bits that are irrelevant for the partition of the
1681 * classifier relevant for the current flow. */
1683 /* Can skip if the field is already unwildcarded. */
1684 if (mask_prefix_bits_set(wc
, be32ofs
, ctx
->maskbits
)) {
1687 /* Check that the trie result will not unwildcard more bits
1688 * than this stage will. */
1689 if (ctx
->maskbits
<= field_plen
[j
]) {
1690 /* Unwildcard the bits and skip the rest. */
1691 mask_set_prefix_bits(wc
, be32ofs
, ctx
->maskbits
);
1692 /* Note: Prerequisite already unwildcarded, as the only
1693 * prerequisite of the supported trie lookup fields is
1694 * the ethertype, which is currently always
1706 /* Returns true if 'target' satisifies 'flow'/'mask', that is, if each bit
1707 * for which 'flow', for which 'mask' has a bit set, specifies a particular
1708 * value has the correct value in 'target'.
1710 * This function is equivalent to miniflow_equal_flow_in_minimask(flow,
1711 * target, mask) but it is faster because of the invariant that
1712 * flow->map and mask->masks.map are the same. */
1714 miniflow_and_mask_matches_flow(const struct miniflow
*flow
,
1715 const struct minimask
*mask
,
1716 const struct flow
*target
)
1718 const uint32_t *flowp
= miniflow_get_u32_values(flow
);
1719 const uint32_t *maskp
= miniflow_get_u32_values(&mask
->masks
);
1720 uint32_t target_u32
;
1722 FLOW_FOR_EACH_IN_MAP(target_u32
, target
, mask
->masks
.map
) {
1723 if ((*flowp
++ ^ target_u32
) & *maskp
++) {
1731 static inline struct cls_match
*
1732 find_match(const struct cls_subtable
*subtable
, const struct flow
*flow
,
1735 struct cls_match
*rule
;
1737 HMAP_FOR_EACH_WITH_HASH (rule
, hmap_node
, hash
, &subtable
->rules
) {
1738 if (miniflow_and_mask_matches_flow(&rule
->flow
, &subtable
->mask
,
1747 static struct cls_match
*
1748 find_match_wc(const struct cls_subtable
*subtable
, const struct flow
*flow
,
1749 struct trie_ctx trie_ctx
[CLS_MAX_TRIES
], unsigned int n_tries
,
1750 struct flow_wildcards
*wc
)
1752 uint32_t basis
= 0, hash
;
1753 struct cls_match
*rule
= NULL
;
1757 if (OVS_UNLIKELY(!wc
)) {
1758 return find_match(subtable
, flow
,
1759 flow_hash_in_minimask(flow
, &subtable
->mask
, 0));
1763 /* Try to finish early by checking fields in segments. */
1764 for (i
= 0; i
< subtable
->n_indices
; i
++) {
1765 struct hindex_node
*inode
;
1766 ofs
.end
= subtable
->index_ofs
[i
];
1768 if (check_tries(trie_ctx
, n_tries
, subtable
->trie_plen
, ofs
, flow
,
1772 hash
= flow_hash_in_minimask_range(flow
, &subtable
->mask
, ofs
.start
,
1774 ofs
.start
= ofs
.end
;
1775 inode
= hindex_node_with_hash(&subtable
->indices
[i
], hash
);
1777 /* No match, can stop immediately, but must fold in the mask
1778 * covered so far. */
1782 /* If we have narrowed down to a single rule already, check whether
1783 * that rule matches. If it does match, then we're done. If it does
1784 * not match, then we know that we will never get a match, but we do
1785 * not yet know how many wildcards we need to fold into 'wc' so we
1786 * continue iterating through indices to find that out. (We won't
1787 * waste time calling miniflow_and_mask_matches_flow() again because
1788 * we've set 'rule' nonnull.)
1790 * This check shows a measurable benefit with non-trivial flow tables.
1792 * (Rare) hash collisions may cause us to miss the opportunity for this
1794 if (!inode
->s
&& !rule
) {
1795 ASSIGN_CONTAINER(rule
, inode
- i
, index_nodes
);
1796 if (miniflow_and_mask_matches_flow(&rule
->flow
, &subtable
->mask
,
1802 ofs
.end
= FLOW_U32S
;
1803 /* Trie check for the final range. */
1804 if (check_tries(trie_ctx
, n_tries
, subtable
->trie_plen
, ofs
, flow
, wc
)) {
1808 /* Multiple potential matches exist, look for one. */
1809 hash
= flow_hash_in_minimask_range(flow
, &subtable
->mask
, ofs
.start
,
1811 rule
= find_match(subtable
, flow
, hash
);
1813 /* We already narrowed the matching candidates down to just 'rule',
1814 * but it didn't match. */
1817 if (!rule
&& subtable
->ports_mask_len
) {
1818 /* Ports are always part of the final range, if any.
1819 * No match was found for the ports. Use the ports trie to figure out
1820 * which ports bits to unwildcard. */
1822 ovs_be32 value
, mask
;
1824 mask
= MINIFLOW_GET_BE32(&subtable
->mask
.masks
, tp_src
);
1825 value
= ((OVS_FORCE ovs_be32
*)flow
)[TP_PORTS_OFS32
] & mask
;
1826 trie_lookup_value(subtable
->ports_trie
, &value
, &mbits
);
1828 ((OVS_FORCE ovs_be32
*)&wc
->masks
)[TP_PORTS_OFS32
] |=
1829 mask
& htonl(~0 << (32 - mbits
));
1831 ofs
.start
= TP_PORTS_OFS32
;
1835 /* Must unwildcard all the fields, as they were looked at. */
1836 flow_wildcards_fold_minimask(wc
, &subtable
->mask
);
1840 /* Must unwildcard the fields looked up so far, if any. */
1842 flow_wildcards_fold_minimask_range(wc
, &subtable
->mask
, 0, ofs
.start
);
1847 static struct cls_match
*
1848 find_equal(struct cls_subtable
*subtable
, const struct miniflow
*flow
,
1851 struct cls_match
*head
;
1853 HMAP_FOR_EACH_WITH_HASH (head
, hmap_node
, hash
, &subtable
->rules
) {
1854 if (miniflow_equal(&head
->flow
, flow
)) {
1861 static struct cls_match
*
1862 insert_rule(struct cls_classifier
*cls
, struct cls_subtable
*subtable
,
1863 struct cls_rule
*new)
1865 struct cls_match
*cls_match
= cls_match_alloc(new);
1866 struct cls_match
*head
;
1867 struct cls_match
*old
= NULL
;
1869 uint32_t basis
= 0, hash
;
1870 uint8_t prev_be32ofs
= 0;
1872 /* Add new node to segment indices. */
1873 for (i
= 0; i
< subtable
->n_indices
; i
++) {
1874 hash
= minimatch_hash_range(&new->match
, prev_be32ofs
,
1875 subtable
->index_ofs
[i
], &basis
);
1876 hindex_insert(&subtable
->indices
[i
], &cls_match
->index_nodes
[i
], hash
);
1877 prev_be32ofs
= subtable
->index_ofs
[i
];
1879 hash
= minimatch_hash_range(&new->match
, prev_be32ofs
, FLOW_U32S
, &basis
);
1880 head
= find_equal(subtable
, &new->match
.flow
, hash
);
1882 hmap_insert(&subtable
->rules
, &cls_match
->hmap_node
, hash
);
1883 list_init(&cls_match
->list
);
1886 /* Scan the list for the insertion point that will keep the list in
1887 * order of decreasing priority. */
1888 struct cls_match
*rule
;
1890 cls_match
->hmap_node
.hash
= hash
; /* Otherwise done by hmap_insert. */
1892 FOR_EACH_RULE_IN_LIST (rule
, head
) {
1893 if (cls_match
->priority
>= rule
->priority
) {
1895 /* 'new' is the new highest-priority flow in the list. */
1896 hmap_replace(&subtable
->rules
,
1897 &rule
->hmap_node
, &cls_match
->hmap_node
);
1900 if (cls_match
->priority
== rule
->priority
) {
1901 list_replace(&cls_match
->list
, &rule
->list
);
1905 list_insert(&rule
->list
, &cls_match
->list
);
1911 /* Insert 'new' at the end of the list. */
1912 list_push_back(&head
->list
, &cls_match
->list
);
1917 update_subtables_after_insertion(cls
, subtable
, cls_match
->priority
);
1919 /* Remove old node from indices. */
1920 for (i
= 0; i
< subtable
->n_indices
; i
++) {
1921 hindex_remove(&subtable
->indices
[i
], &old
->index_nodes
[i
]);
1927 static struct cls_match
*
1928 next_rule_in_list__(struct cls_match
*rule
)
1930 struct cls_match
*next
= OBJECT_CONTAINING(rule
->list
.next
, next
, list
);
1934 static struct cls_match
*
1935 next_rule_in_list(struct cls_match
*rule
)
1937 struct cls_match
*next
= next_rule_in_list__(rule
);
1938 return next
->priority
< rule
->priority
? next
: NULL
;
1941 /* A longest-prefix match tree. */
1943 uint32_t prefix
; /* Prefix bits for this node, MSB first. */
1944 uint8_t nbits
; /* Never zero, except for the root node. */
1945 unsigned int n_rules
; /* Number of rules that have this prefix. */
1946 struct trie_node
*edges
[2]; /* Both NULL if leaf. */
1949 /* Max bits per node. Must fit in struct trie_node's 'prefix'.
1950 * Also tested with 16, 8, and 5 to stress the implementation. */
1951 #define TRIE_PREFIX_BITS 32
1953 /* Return at least 'plen' bits of the 'prefix', starting at bit offset 'ofs'.
1954 * Prefixes are in the network byte order, and the offset 0 corresponds to
1955 * the most significant bit of the first byte. The offset can be read as
1956 * "how many bits to skip from the start of the prefix starting at 'pr'". */
1958 raw_get_prefix(const ovs_be32 pr
[], unsigned int ofs
, unsigned int plen
)
1962 pr
+= ofs
/ 32; /* Where to start. */
1963 ofs
%= 32; /* How many bits to skip at 'pr'. */
1965 prefix
= ntohl(*pr
) << ofs
; /* Get the first 32 - ofs bits. */
1966 if (plen
> 32 - ofs
) { /* Need more than we have already? */
1967 prefix
|= ntohl(*++pr
) >> (32 - ofs
);
1969 /* Return with possible unwanted bits at the end. */
1973 /* Return min(TRIE_PREFIX_BITS, plen) bits of the 'prefix', starting at bit
1974 * offset 'ofs'. Prefixes are in the network byte order, and the offset 0
1975 * corresponds to the most significant bit of the first byte. The offset can
1976 * be read as "how many bits to skip from the start of the prefix starting at
1979 trie_get_prefix(const ovs_be32 pr
[], unsigned int ofs
, unsigned int plen
)
1984 if (plen
> TRIE_PREFIX_BITS
) {
1985 plen
= TRIE_PREFIX_BITS
; /* Get at most TRIE_PREFIX_BITS. */
1987 /* Return with unwanted bits cleared. */
1988 return raw_get_prefix(pr
, ofs
, plen
) & ~0u << (32 - plen
);
1991 /* Return the number of equal bits in 'nbits' of 'prefix's MSBs and a 'value'
1992 * starting at "MSB 0"-based offset 'ofs'. */
1994 prefix_equal_bits(uint32_t prefix
, unsigned int nbits
, const ovs_be32 value
[],
1997 uint64_t diff
= prefix
^ raw_get_prefix(value
, ofs
, nbits
);
1998 /* Set the bit after the relevant bits to limit the result. */
1999 return raw_clz64(diff
<< 32 | UINT64_C(1) << (63 - nbits
));
2002 /* Return the number of equal bits in 'node' prefix and a 'prefix' of length
2003 * 'plen', starting at "MSB 0"-based offset 'ofs'. */
2005 trie_prefix_equal_bits(const struct trie_node
*node
, const ovs_be32 prefix
[],
2006 unsigned int ofs
, unsigned int plen
)
2008 return prefix_equal_bits(node
->prefix
, MIN(node
->nbits
, plen
- ofs
),
2012 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' can
2013 * be greater than 31. */
2015 be_get_bit_at(const ovs_be32 value
[], unsigned int ofs
)
2017 return (((const uint8_t *)value
)[ofs
/ 8] >> (7 - ofs
% 8)) & 1u;
2020 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' must
2021 * be between 0 and 31, inclusive. */
2023 get_bit_at(const uint32_t prefix
, unsigned int ofs
)
2025 return (prefix
>> (31 - ofs
)) & 1u;
2028 /* Create new branch. */
2029 static struct trie_node
*
2030 trie_branch_create(const ovs_be32
*prefix
, unsigned int ofs
, unsigned int plen
,
2031 unsigned int n_rules
)
2033 struct trie_node
*node
= xmalloc(sizeof *node
);
2035 node
->prefix
= trie_get_prefix(prefix
, ofs
, plen
);
2037 if (plen
<= TRIE_PREFIX_BITS
) {
2039 node
->edges
[0] = NULL
;
2040 node
->edges
[1] = NULL
;
2041 node
->n_rules
= n_rules
;
2042 } else { /* Need intermediate nodes. */
2043 struct trie_node
*subnode
= trie_branch_create(prefix
,
2044 ofs
+ TRIE_PREFIX_BITS
,
2045 plen
- TRIE_PREFIX_BITS
,
2047 int bit
= get_bit_at(subnode
->prefix
, 0);
2048 node
->nbits
= TRIE_PREFIX_BITS
;
2049 node
->edges
[bit
] = subnode
;
2050 node
->edges
[!bit
] = NULL
;
2057 trie_node_destroy(struct trie_node
*node
)
2063 trie_destroy(struct trie_node
*node
)
2066 trie_destroy(node
->edges
[0]);
2067 trie_destroy(node
->edges
[1]);
2073 trie_is_leaf(const struct trie_node
*trie
)
2075 return !trie
->edges
[0] && !trie
->edges
[1]; /* No children. */
2079 mask_set_prefix_bits(struct flow_wildcards
*wc
, uint8_t be32ofs
,
2082 ovs_be32
*mask
= &((ovs_be32
*)&wc
->masks
)[be32ofs
];
2085 for (i
= 0; i
< nbits
/ 32; i
++) {
2086 mask
[i
] = OVS_BE32_MAX
;
2089 mask
[i
] |= htonl(~0u << (32 - nbits
% 32));
2094 mask_prefix_bits_set(const struct flow_wildcards
*wc
, uint8_t be32ofs
,
2097 ovs_be32
*mask
= &((ovs_be32
*)&wc
->masks
)[be32ofs
];
2099 ovs_be32 zeroes
= 0;
2101 for (i
= 0; i
< nbits
/ 32; i
++) {
2105 zeroes
|= ~mask
[i
] & htonl(~0u << (32 - nbits
% 32));
2108 return !zeroes
; /* All 'nbits' bits set. */
2111 static struct trie_node
**
2112 trie_next_edge(struct trie_node
*node
, const ovs_be32 value
[],
2115 return node
->edges
+ be_get_bit_at(value
, ofs
);
2118 static const struct trie_node
*
2119 trie_next_node(const struct trie_node
*node
, const ovs_be32 value
[],
2122 return node
->edges
[be_get_bit_at(value
, ofs
)];
2125 /* Return the prefix mask length necessary to find the longest-prefix match for
2126 * the '*value' in the prefix tree 'node'.
2127 * '*checkbits' is set to the number of bits in the prefix mask necessary to
2128 * determine a mismatch, in case there are longer prefixes in the tree below
2129 * the one that matched.
2132 trie_lookup_value(const struct trie_node
*node
, const ovs_be32 value
[],
2133 unsigned int *checkbits
)
2135 unsigned int plen
= 0, match_len
= 0;
2136 const struct trie_node
*prev
= NULL
;
2138 for (; node
; prev
= node
, node
= trie_next_node(node
, value
, plen
)) {
2139 unsigned int eqbits
;
2140 /* Check if this edge can be followed. */
2141 eqbits
= prefix_equal_bits(node
->prefix
, node
->nbits
, value
, plen
);
2143 if (eqbits
< node
->nbits
) { /* Mismatch, nothing more to be found. */
2144 /* Bit at offset 'plen' differed. */
2145 *checkbits
= plen
+ 1; /* Includes the first mismatching bit. */
2148 /* Full match, check if rules exist at this prefix length. */
2149 if (node
->n_rules
> 0) {
2153 /* Dead end, exclude the other branch if it exists. */
2154 *checkbits
= !prev
|| trie_is_leaf(prev
) ? plen
: plen
+ 1;
2159 trie_lookup(const struct cls_trie
*trie
, const struct flow
*flow
,
2160 unsigned int *checkbits
)
2162 const struct mf_field
*mf
= trie
->field
;
2164 /* Check that current flow matches the prerequisites for the trie
2165 * field. Some match fields are used for multiple purposes, so we
2166 * must check that the trie is relevant for this flow. */
2167 if (mf_are_prereqs_ok(mf
, flow
)) {
2168 return trie_lookup_value(trie
->root
,
2169 &((ovs_be32
*)flow
)[mf
->flow_be32ofs
],
2172 *checkbits
= 0; /* Value not used in this case. */
2176 /* Returns the length of a prefix match mask for the field 'mf' in 'minimask'.
2177 * Returns the u32 offset to the miniflow data in '*miniflow_index', if
2178 * 'miniflow_index' is not NULL. */
2180 minimask_get_prefix_len(const struct minimask
*minimask
,
2181 const struct mf_field
*mf
)
2183 unsigned int nbits
= 0, mask_tz
= 0; /* Non-zero when end of mask seen. */
2184 uint8_t u32_ofs
= mf
->flow_be32ofs
;
2185 uint8_t u32_end
= u32_ofs
+ mf
->n_bytes
/ 4;
2187 for (; u32_ofs
< u32_end
; ++u32_ofs
) {
2189 mask
= ntohl((OVS_FORCE ovs_be32
)minimask_get(minimask
, u32_ofs
));
2191 /* Validate mask, count the mask length. */
2194 return 0; /* No bits allowed after mask ended. */
2197 if (~mask
& (~mask
+ 1)) {
2198 return 0; /* Mask not contiguous. */
2200 mask_tz
= ctz32(mask
);
2201 nbits
+= 32 - mask_tz
;
2209 * This is called only when mask prefix is known to be CIDR and non-zero.
2210 * Relies on the fact that the flow and mask have the same map, and since
2211 * the mask is CIDR, the storage for the flow field exists even if it
2212 * happened to be zeros.
2214 static const ovs_be32
*
2215 minimatch_get_prefix(const struct minimatch
*match
, const struct mf_field
*mf
)
2217 return miniflow_get_be32_values(&match
->flow
) +
2218 count_1bits(match
->flow
.map
& ((UINT64_C(1) << mf
->flow_be32ofs
) - 1));
2221 /* Insert rule in to the prefix tree.
2222 * 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2225 trie_insert(struct cls_trie
*trie
, const struct cls_rule
*rule
, int mlen
)
2227 trie_insert_prefix(&trie
->root
,
2228 minimatch_get_prefix(&rule
->match
, trie
->field
), mlen
);
2232 trie_insert_prefix(struct trie_node
**edge
, const ovs_be32
*prefix
, int mlen
)
2234 struct trie_node
*node
;
2237 /* Walk the tree. */
2238 for (; (node
= *edge
) != NULL
;
2239 edge
= trie_next_edge(node
, prefix
, ofs
)) {
2240 unsigned int eqbits
= trie_prefix_equal_bits(node
, prefix
, ofs
, mlen
);
2242 if (eqbits
< node
->nbits
) {
2243 /* Mismatch, new node needs to be inserted above. */
2244 int old_branch
= get_bit_at(node
->prefix
, eqbits
);
2246 /* New parent node. */
2247 *edge
= trie_branch_create(prefix
, ofs
- eqbits
, eqbits
,
2248 ofs
== mlen
? 1 : 0);
2250 /* Adjust old node for its new position in the tree. */
2251 node
->prefix
<<= eqbits
;
2252 node
->nbits
-= eqbits
;
2253 (*edge
)->edges
[old_branch
] = node
;
2255 /* Check if need a new branch for the new rule. */
2257 (*edge
)->edges
[!old_branch
]
2258 = trie_branch_create(prefix
, ofs
, mlen
- ofs
, 1);
2262 /* Full match so far. */
2265 /* Full match at the current node, rule needs to be added here. */
2270 /* Must insert a new tree branch for the new rule. */
2271 *edge
= trie_branch_create(prefix
, ofs
, mlen
- ofs
, 1);
2274 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2277 trie_remove(struct cls_trie
*trie
, const struct cls_rule
*rule
, int mlen
)
2279 trie_remove_prefix(&trie
->root
,
2280 minimatch_get_prefix(&rule
->match
, trie
->field
), mlen
);
2283 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2286 trie_remove_prefix(struct trie_node
**root
, const ovs_be32
*prefix
, int mlen
)
2288 struct trie_node
*node
;
2289 struct trie_node
**edges
[sizeof(union mf_value
) * 8];
2290 int depth
= 0, ofs
= 0;
2292 /* Walk the tree. */
2293 for (edges
[0] = root
;
2294 (node
= *edges
[depth
]) != NULL
;
2295 edges
[++depth
] = trie_next_edge(node
, prefix
, ofs
)) {
2296 unsigned int eqbits
= trie_prefix_equal_bits(node
, prefix
, ofs
, mlen
);
2298 if (eqbits
< node
->nbits
) {
2299 /* Mismatch, nothing to be removed. This should never happen, as
2300 * only rules in the classifier are ever removed. */
2301 break; /* Log a warning. */
2303 /* Full match so far. */
2307 /* Full prefix match at the current node, remove rule here. */
2308 if (!node
->n_rules
) {
2309 break; /* Log a warning. */
2313 /* Check if can prune the tree. */
2314 while (!node
->n_rules
&& !(node
->edges
[0] && node
->edges
[1])) {
2315 /* No rules and at most one child node, remove this node. */
2316 struct trie_node
*next
;
2317 next
= node
->edges
[0] ? node
->edges
[0] : node
->edges
[1];
2320 if (node
->nbits
+ next
->nbits
> TRIE_PREFIX_BITS
) {
2321 break; /* Cannot combine. */
2323 /* Combine node with next. */
2324 next
->prefix
= node
->prefix
| next
->prefix
>> node
->nbits
;
2325 next
->nbits
+= node
->nbits
;
2327 trie_node_destroy(node
);
2328 /* Update the parent's edge. */
2329 *edges
[depth
] = next
;
2330 if (next
|| !depth
) {
2331 /* Branch not pruned or at root, nothing more to do. */
2334 node
= *edges
[--depth
];
2339 /* Cannot go deeper. This should never happen, since only rules
2340 * that actually exist in the classifier are ever removed. */
2341 VLOG_WARN("Trying to remove non-existing rule from a prefix trie.");