]> git.proxmox.com Git - ovs.git/blob - lib/classifier.c
Remove assumption that there are 64 or fewer fields.
[ovs.git] / lib / classifier.c
1 /*
2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include <config.h>
18 #include "classifier.h"
19 #include <errno.h>
20 #include <netinet/in.h>
21 #include "byte-order.h"
22 #include "dynamic-string.h"
23 #include "flow.h"
24 #include "hash.h"
25 #include "cmap.h"
26 #include "list.h"
27 #include "odp-util.h"
28 #include "ofp-util.h"
29 #include "packets.h"
30 #include "tag.h"
31 #include "util.h"
32 #include "vlog.h"
33
34 VLOG_DEFINE_THIS_MODULE(classifier);
35
36 struct trie_ctx;
37
38 /* Ports trie depends on both ports sharing the same ovs_be32. */
39 #define TP_PORTS_OFS32 (offsetof(struct flow, tp_src) / 4)
40 BUILD_ASSERT_DECL(TP_PORTS_OFS32 == offsetof(struct flow, tp_dst) / 4);
41
42 /* A set of rules that all have the same fields wildcarded. */
43 struct cls_subtable {
44 /* The fields are only used by writers and iterators. */
45 struct cmap_node cmap_node; /* Within struct classifier 'subtables_map'. */
46
47 /* The fields are only used by writers. */
48 int n_rules OVS_GUARDED; /* Number of rules, including
49 * duplicates. */
50 unsigned int max_priority OVS_GUARDED; /* Max priority of any rule in
51 * the subtable. */
52 unsigned int max_count OVS_GUARDED; /* Count of max_priority rules. */
53
54 /* These fields are accessed by readers who care about wildcarding. */
55 tag_type tag; /* Tag generated from mask for partitioning (const). */
56 uint8_t n_indices; /* How many indices to use (const). */
57 uint8_t index_ofs[CLS_MAX_INDICES]; /* u32 segment boundaries (const). */
58 unsigned int trie_plen[CLS_MAX_TRIES]; /* Trie prefix length in 'mask'
59 * (runtime configurable). */
60 int ports_mask_len; /* (const) */
61 struct cmap indices[CLS_MAX_INDICES]; /* Staged lookup indices. */
62 rcu_trie_ptr ports_trie; /* NULL if none. */
63
64 /* These fields are accessed by all readers. */
65 struct cmap rules; /* Contains "struct cls_rule"s. */
66 struct minimask mask; /* Wildcards for fields (const). */
67 /* 'mask' must be the last field. */
68 };
69
70 /* Associates a metadata value (that is, a value of the OpenFlow 1.1+ metadata
71 * field) with tags for the "cls_subtable"s that contain rules that match that
72 * metadata value. */
73 struct cls_partition {
74 struct cmap_node cmap_node; /* In struct classifier's 'partitions' map. */
75 ovs_be64 metadata; /* metadata value for this partition. */
76 tag_type tags; /* OR of each flow's cls_subtable tag. */
77 struct tag_tracker tracker OVS_GUARDED; /* Tracks the bits in 'tags'. */
78 };
79
80 /* Internal representation of a rule in a "struct cls_subtable". */
81 struct cls_match {
82 /* Accessed only by writers and iterators. */
83 struct list list OVS_GUARDED; /* List of identical, lower-priority rules. */
84
85 /* Accessed only by writers. */
86 struct cls_partition *partition OVS_GUARDED;
87
88 /* Accessed by readers interested in wildcarding. */
89 unsigned int priority; /* Larger numbers are higher priorities. */
90 struct cmap_node index_nodes[CLS_MAX_INDICES]; /* Within subtable's
91 * 'indices'. */
92 /* Accessed by all readers. */
93 struct cmap_node cmap_node; /* Within struct cls_subtable 'rules'. */
94 struct cls_rule *cls_rule;
95 struct miniflow flow; /* Matching rule. Mask is in the subtable. */
96 /* 'flow' must be the last field. */
97 };
98
99 static struct cls_match *
100 cls_match_alloc(struct cls_rule *rule)
101 {
102 int count = count_1bits(rule->match.flow.map);
103
104 struct cls_match *cls_match
105 = xmalloc(sizeof *cls_match - sizeof cls_match->flow.inline_values
106 + MINIFLOW_VALUES_SIZE(count));
107
108 cls_match->cls_rule = rule;
109 miniflow_clone_inline(&cls_match->flow, &rule->match.flow, count);
110 cls_match->priority = rule->priority;
111 rule->cls_match = cls_match;
112
113 return cls_match;
114 }
115
116 static struct cls_subtable *find_subtable(const struct classifier *cls,
117 const struct minimask *)
118 OVS_REQUIRES(cls->mutex);
119 static struct cls_subtable *insert_subtable(struct classifier *cls,
120 const struct minimask *)
121 OVS_REQUIRES(cls->mutex);
122 static void destroy_subtable(struct classifier *cls, struct cls_subtable *)
123 OVS_REQUIRES(cls->mutex);
124 static struct cls_match *insert_rule(struct classifier *cls,
125 struct cls_subtable *, struct cls_rule *)
126 OVS_REQUIRES(cls->mutex);
127
128 static struct cls_match *find_match_wc(const struct cls_subtable *,
129 const struct flow *, struct trie_ctx *,
130 unsigned int n_tries,
131 struct flow_wildcards *);
132 static struct cls_match *find_equal(struct cls_subtable *,
133 const struct miniflow *, uint32_t hash);
134
135 /* Iterates RULE over HEAD and all of the cls_rules on HEAD->list.
136 * Classifier's mutex must be held while iterating, as the list is
137 * protoceted by it. */
138 #define FOR_EACH_RULE_IN_LIST(RULE, HEAD) \
139 for ((RULE) = (HEAD); (RULE) != NULL; (RULE) = next_rule_in_list(RULE))
140 #define FOR_EACH_RULE_IN_LIST_SAFE(RULE, NEXT, HEAD) \
141 for ((RULE) = (HEAD); \
142 (RULE) != NULL && ((NEXT) = next_rule_in_list(RULE), true); \
143 (RULE) = (NEXT))
144
145 static struct cls_match *next_rule_in_list__(struct cls_match *);
146 static struct cls_match *next_rule_in_list(struct cls_match *);
147
148 static unsigned int minimask_get_prefix_len(const struct minimask *,
149 const struct mf_field *);
150 static void trie_init(struct classifier *cls, int trie_idx,
151 const struct mf_field *)
152 OVS_REQUIRES(cls->mutex);
153 static unsigned int trie_lookup(const struct cls_trie *, const struct flow *,
154 union mf_value *plens);
155 static unsigned int trie_lookup_value(const rcu_trie_ptr *,
156 const ovs_be32 value[], ovs_be32 plens[],
157 unsigned int value_bits);
158 static void trie_destroy(rcu_trie_ptr *);
159 static void trie_insert(struct cls_trie *, const struct cls_rule *, int mlen);
160 static void trie_insert_prefix(rcu_trie_ptr *, const ovs_be32 *prefix,
161 int mlen);
162 static void trie_remove(struct cls_trie *, const struct cls_rule *, int mlen);
163 static void trie_remove_prefix(rcu_trie_ptr *, const ovs_be32 *prefix,
164 int mlen);
165 static void mask_set_prefix_bits(struct flow_wildcards *, uint8_t be32ofs,
166 unsigned int n_bits);
167 static bool mask_prefix_bits_set(const struct flow_wildcards *,
168 uint8_t be32ofs, unsigned int n_bits);
169 \f
170 /* flow/miniflow/minimask/minimatch utilities.
171 * These are only used by the classifier, so place them here to allow
172 * for better optimization. */
173
174 static inline uint64_t
175 miniflow_get_map_in_range(const struct miniflow *miniflow,
176 uint8_t start, uint8_t end, unsigned int *offset)
177 {
178 uint64_t map = miniflow->map;
179 *offset = 0;
180
181 if (start > 0) {
182 uint64_t msk = (UINT64_C(1) << start) - 1; /* 'start' LSBs set */
183 *offset = count_1bits(map & msk);
184 map &= ~msk;
185 }
186 if (end < FLOW_U32S) {
187 uint64_t msk = (UINT64_C(1) << end) - 1; /* 'end' LSBs set */
188 map &= msk;
189 }
190 return map;
191 }
192
193 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
194 * 'mask', given 'basis'.
195 *
196 * The hash values returned by this function are the same as those returned by
197 * miniflow_hash_in_minimask(), only the form of the arguments differ. */
198 static inline uint32_t
199 flow_hash_in_minimask(const struct flow *flow, const struct minimask *mask,
200 uint32_t basis)
201 {
202 const uint32_t *mask_values = miniflow_get_u32_values(&mask->masks);
203 const uint32_t *flow_u32 = (const uint32_t *)flow;
204 const uint32_t *p = mask_values;
205 uint32_t hash;
206 uint64_t map;
207
208 hash = basis;
209 for (map = mask->masks.map; map; map = zero_rightmost_1bit(map)) {
210 hash = hash_add(hash, flow_u32[raw_ctz(map)] & *p++);
211 }
212
213 return hash_finish(hash, (p - mask_values) * 4);
214 }
215
216 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
217 * 'mask', given 'basis'.
218 *
219 * The hash values returned by this function are the same as those returned by
220 * flow_hash_in_minimask(), only the form of the arguments differ. */
221 static inline uint32_t
222 miniflow_hash_in_minimask(const struct miniflow *flow,
223 const struct minimask *mask, uint32_t basis)
224 {
225 const uint32_t *mask_values = miniflow_get_u32_values(&mask->masks);
226 const uint32_t *p = mask_values;
227 uint32_t hash = basis;
228 uint32_t flow_u32;
229
230 MINIFLOW_FOR_EACH_IN_MAP(flow_u32, flow, mask->masks.map) {
231 hash = hash_add(hash, flow_u32 & *p++);
232 }
233
234 return hash_finish(hash, (p - mask_values) * 4);
235 }
236
237 /* Returns a hash value for the bits of range [start, end) in 'flow',
238 * where there are 1-bits in 'mask', given 'hash'.
239 *
240 * The hash values returned by this function are the same as those returned by
241 * minimatch_hash_range(), only the form of the arguments differ. */
242 static inline uint32_t
243 flow_hash_in_minimask_range(const struct flow *flow,
244 const struct minimask *mask,
245 uint8_t start, uint8_t end, uint32_t *basis)
246 {
247 const uint32_t *mask_values = miniflow_get_u32_values(&mask->masks);
248 const uint32_t *flow_u32 = (const uint32_t *)flow;
249 unsigned int offset;
250 uint64_t map = miniflow_get_map_in_range(&mask->masks, start, end,
251 &offset);
252 const uint32_t *p = mask_values + offset;
253 uint32_t hash = *basis;
254
255 for (; map; map = zero_rightmost_1bit(map)) {
256 hash = hash_add(hash, flow_u32[raw_ctz(map)] & *p++);
257 }
258
259 *basis = hash; /* Allow continuation from the unfinished value. */
260 return hash_finish(hash, (p - mask_values) * 4);
261 }
262
263 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask. */
264 static inline void
265 flow_wildcards_fold_minimask(struct flow_wildcards *wc,
266 const struct minimask *mask)
267 {
268 flow_union_with_miniflow(&wc->masks, &mask->masks);
269 }
270
271 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask
272 * in range [start, end). */
273 static inline void
274 flow_wildcards_fold_minimask_range(struct flow_wildcards *wc,
275 const struct minimask *mask,
276 uint8_t start, uint8_t end)
277 {
278 uint32_t *dst_u32 = (uint32_t *)&wc->masks;
279 unsigned int offset;
280 uint64_t map = miniflow_get_map_in_range(&mask->masks, start, end,
281 &offset);
282 const uint32_t *p = miniflow_get_u32_values(&mask->masks) + offset;
283
284 for (; map; map = zero_rightmost_1bit(map)) {
285 dst_u32[raw_ctz(map)] |= *p++;
286 }
287 }
288
289 /* Returns a hash value for 'flow', given 'basis'. */
290 static inline uint32_t
291 miniflow_hash(const struct miniflow *flow, uint32_t basis)
292 {
293 const uint32_t *values = miniflow_get_u32_values(flow);
294 const uint32_t *p = values;
295 uint32_t hash = basis;
296 uint64_t hash_map = 0;
297 uint64_t map;
298
299 for (map = flow->map; map; map = zero_rightmost_1bit(map)) {
300 if (*p) {
301 hash = hash_add(hash, *p);
302 hash_map |= rightmost_1bit(map);
303 }
304 p++;
305 }
306 hash = hash_add(hash, hash_map);
307 hash = hash_add(hash, hash_map >> 32);
308
309 return hash_finish(hash, p - values);
310 }
311
312 /* Returns a hash value for 'mask', given 'basis'. */
313 static inline uint32_t
314 minimask_hash(const struct minimask *mask, uint32_t basis)
315 {
316 return miniflow_hash(&mask->masks, basis);
317 }
318
319 /* Returns a hash value for 'match', given 'basis'. */
320 static inline uint32_t
321 minimatch_hash(const struct minimatch *match, uint32_t basis)
322 {
323 return miniflow_hash(&match->flow, minimask_hash(&match->mask, basis));
324 }
325
326 /* Returns a hash value for the bits of range [start, end) in 'minimatch',
327 * given 'basis'.
328 *
329 * The hash values returned by this function are the same as those returned by
330 * flow_hash_in_minimask_range(), only the form of the arguments differ. */
331 static inline uint32_t
332 minimatch_hash_range(const struct minimatch *match, uint8_t start, uint8_t end,
333 uint32_t *basis)
334 {
335 unsigned int offset;
336 const uint32_t *p, *q;
337 uint32_t hash = *basis;
338 int n, i;
339
340 n = count_1bits(miniflow_get_map_in_range(&match->mask.masks, start, end,
341 &offset));
342 q = miniflow_get_u32_values(&match->mask.masks) + offset;
343 p = miniflow_get_u32_values(&match->flow) + offset;
344
345 for (i = 0; i < n; i++) {
346 hash = hash_add(hash, p[i] & q[i]);
347 }
348 *basis = hash; /* Allow continuation from the unfinished value. */
349 return hash_finish(hash, (offset + n) * 4);
350 }
351
352 \f
353 /* cls_rule. */
354
355 /* Initializes 'rule' to match packets specified by 'match' at the given
356 * 'priority'. 'match' must satisfy the invariant described in the comment at
357 * the definition of struct match.
358 *
359 * The caller must eventually destroy 'rule' with cls_rule_destroy().
360 *
361 * (OpenFlow uses priorities between 0 and UINT16_MAX, inclusive, but
362 * internally Open vSwitch supports a wider range.) */
363 void
364 cls_rule_init(struct cls_rule *rule,
365 const struct match *match, unsigned int priority)
366 {
367 minimatch_init(&rule->match, match);
368 rule->priority = priority;
369 rule->cls_match = NULL;
370 }
371
372 /* Same as cls_rule_init() for initialization from a "struct minimatch". */
373 void
374 cls_rule_init_from_minimatch(struct cls_rule *rule,
375 const struct minimatch *match,
376 unsigned int priority)
377 {
378 minimatch_clone(&rule->match, match);
379 rule->priority = priority;
380 rule->cls_match = NULL;
381 }
382
383 /* Initializes 'dst' as a copy of 'src'.
384 *
385 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
386 void
387 cls_rule_clone(struct cls_rule *dst, const struct cls_rule *src)
388 {
389 minimatch_clone(&dst->match, &src->match);
390 dst->priority = src->priority;
391 dst->cls_match = NULL;
392 }
393
394 /* Initializes 'dst' with the data in 'src', destroying 'src'.
395 *
396 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
397 void
398 cls_rule_move(struct cls_rule *dst, struct cls_rule *src)
399 {
400 minimatch_move(&dst->match, &src->match);
401 dst->priority = src->priority;
402 dst->cls_match = NULL;
403 }
404
405 /* Frees memory referenced by 'rule'. Doesn't free 'rule' itself (it's
406 * normally embedded into a larger structure).
407 *
408 * ('rule' must not currently be in a classifier.) */
409 void
410 cls_rule_destroy(struct cls_rule *rule)
411 {
412 ovs_assert(!rule->cls_match);
413 minimatch_destroy(&rule->match);
414 }
415
416 /* Returns true if 'a' and 'b' match the same packets at the same priority,
417 * false if they differ in some way. */
418 bool
419 cls_rule_equal(const struct cls_rule *a, const struct cls_rule *b)
420 {
421 return a->priority == b->priority && minimatch_equal(&a->match, &b->match);
422 }
423
424 /* Returns a hash value for 'rule', folding in 'basis'. */
425 uint32_t
426 cls_rule_hash(const struct cls_rule *rule, uint32_t basis)
427 {
428 return minimatch_hash(&rule->match, hash_int(rule->priority, basis));
429 }
430
431 /* Appends a string describing 'rule' to 's'. */
432 void
433 cls_rule_format(const struct cls_rule *rule, struct ds *s)
434 {
435 minimatch_format(&rule->match, s, rule->priority);
436 }
437
438 /* Returns true if 'rule' matches every packet, false otherwise. */
439 bool
440 cls_rule_is_catchall(const struct cls_rule *rule)
441 {
442 return minimask_is_catchall(&rule->match.mask);
443 }
444 \f
445 /* Initializes 'cls' as a classifier that initially contains no classification
446 * rules. */
447 void
448 classifier_init(struct classifier *cls, const uint8_t *flow_segments)
449 OVS_EXCLUDED(cls->mutex)
450 {
451 ovs_mutex_init(&cls->mutex);
452 ovs_mutex_lock(&cls->mutex);
453 cls->n_rules = 0;
454 cmap_init(&cls->subtables_map);
455 pvector_init(&cls->subtables);
456 cmap_init(&cls->partitions);
457 cls->n_flow_segments = 0;
458 if (flow_segments) {
459 while (cls->n_flow_segments < CLS_MAX_INDICES
460 && *flow_segments < FLOW_U32S) {
461 cls->flow_segments[cls->n_flow_segments++] = *flow_segments++;
462 }
463 }
464 cls->n_tries = 0;
465 for (int i = 0; i < CLS_MAX_TRIES; i++) {
466 trie_init(cls, i, NULL);
467 }
468 ovs_mutex_unlock(&cls->mutex);
469 }
470
471 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
472 * caller's responsibility.
473 * May only be called after all the readers have been terminated. */
474 void
475 classifier_destroy(struct classifier *cls)
476 OVS_EXCLUDED(cls->mutex)
477 {
478 if (cls) {
479 struct cls_partition *partition;
480 struct cls_subtable *subtable;
481 int i;
482
483 ovs_mutex_lock(&cls->mutex);
484 for (i = 0; i < cls->n_tries; i++) {
485 trie_destroy(&cls->tries[i].root);
486 }
487
488 CMAP_FOR_EACH_SAFE (subtable, cmap_node, &cls->subtables_map) {
489 destroy_subtable(cls, subtable);
490 }
491 cmap_destroy(&cls->subtables_map);
492
493 CMAP_FOR_EACH_SAFE (partition, cmap_node, &cls->partitions) {
494 ovsrcu_postpone(free, partition);
495 }
496 cmap_destroy(&cls->partitions);
497
498 pvector_destroy(&cls->subtables);
499 ovs_mutex_unlock(&cls->mutex);
500 ovs_mutex_destroy(&cls->mutex);
501 }
502 }
503
504 /* Set the fields for which prefix lookup should be performed. */
505 bool
506 classifier_set_prefix_fields(struct classifier *cls,
507 const enum mf_field_id *trie_fields,
508 unsigned int n_fields)
509 OVS_EXCLUDED(cls->mutex)
510 {
511 const struct mf_field * new_fields[CLS_MAX_TRIES];
512 struct mf_bitmap fields = MF_BITMAP_INITIALIZER;
513 int i, n_tries = 0;
514 bool changed = false;
515
516 ovs_mutex_lock(&cls->mutex);
517 for (i = 0; i < n_fields && n_tries < CLS_MAX_TRIES; i++) {
518 const struct mf_field *field = mf_from_id(trie_fields[i]);
519 if (field->flow_be32ofs < 0 || field->n_bits % 32) {
520 /* Incompatible field. This is the only place where we
521 * enforce these requirements, but the rest of the trie code
522 * depends on the flow_be32ofs to be non-negative and the
523 * field length to be a multiple of 32 bits. */
524 continue;
525 }
526
527 if (bitmap_is_set(fields.bm, trie_fields[i])) {
528 /* Duplicate field, there is no need to build more than
529 * one index for any one field. */
530 continue;
531 }
532 bitmap_set1(fields.bm, trie_fields[i]);
533
534 new_fields[n_tries] = NULL;
535 if (n_tries >= cls->n_tries || field != cls->tries[n_tries].field) {
536 new_fields[n_tries] = field;
537 changed = true;
538 }
539 n_tries++;
540 }
541
542 if (changed || n_tries < cls->n_tries) {
543 struct cls_subtable *subtable;
544
545 /* Trie configuration needs to change. Disable trie lookups
546 * for the tries that are changing and wait all the current readers
547 * with the old configuration to be done. */
548 changed = false;
549 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
550 for (i = 0; i < cls->n_tries; i++) {
551 if ((i < n_tries && new_fields[i]) || i >= n_tries) {
552 if (subtable->trie_plen[i]) {
553 subtable->trie_plen[i] = 0;
554 changed = true;
555 }
556 }
557 }
558 }
559 /* Synchronize if any readers were using tries. The readers may
560 * temporarily function without the trie lookup based optimizations. */
561 if (changed) {
562 /* ovsrcu_synchronize() functions as a memory barrier, so it does
563 * not matter that subtable->trie_plen is not atomic. */
564 ovsrcu_synchronize();
565 }
566
567 /* Now set up the tries. */
568 for (i = 0; i < n_tries; i++) {
569 if (new_fields[i]) {
570 trie_init(cls, i, new_fields[i]);
571 }
572 }
573 /* Destroy the rest, if any. */
574 for (; i < cls->n_tries; i++) {
575 trie_init(cls, i, NULL);
576 }
577
578 cls->n_tries = n_tries;
579 ovs_mutex_unlock(&cls->mutex);
580 return true;
581 }
582
583 ovs_mutex_unlock(&cls->mutex);
584 return false; /* No change. */
585 }
586
587 static void
588 trie_init(struct classifier *cls, int trie_idx, const struct mf_field *field)
589 OVS_REQUIRES(cls->mutex)
590 {
591 struct cls_trie *trie = &cls->tries[trie_idx];
592 struct cls_subtable *subtable;
593
594 if (trie_idx < cls->n_tries) {
595 trie_destroy(&trie->root);
596 } else {
597 ovsrcu_set_hidden(&trie->root, NULL);
598 }
599 trie->field = field;
600
601 /* Add existing rules to the new trie. */
602 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
603 unsigned int plen;
604
605 plen = field ? minimask_get_prefix_len(&subtable->mask, field) : 0;
606 if (plen) {
607 struct cls_match *head;
608
609 CMAP_FOR_EACH (head, cmap_node, &subtable->rules) {
610 struct cls_match *match;
611
612 FOR_EACH_RULE_IN_LIST (match, head) {
613 trie_insert(trie, match->cls_rule, plen);
614 }
615 }
616 }
617 /* Initialize subtable's prefix length on this field. This will
618 * allow readers to use the trie. */
619 atomic_thread_fence(memory_order_release);
620 subtable->trie_plen[trie_idx] = plen;
621 }
622 }
623
624 /* Returns true if 'cls' contains no classification rules, false otherwise.
625 * Checking the cmap requires no locking. */
626 bool
627 classifier_is_empty(const struct classifier *cls)
628 {
629 return cmap_is_empty(&cls->subtables_map);
630 }
631
632 /* Returns the number of rules in 'cls'. */
633 int
634 classifier_count(const struct classifier *cls)
635 OVS_NO_THREAD_SAFETY_ANALYSIS
636 {
637 /* n_rules is an int, so in the presence of concurrent writers this will
638 * return either the old or a new value. */
639 return cls->n_rules;
640 }
641
642 static uint32_t
643 hash_metadata(ovs_be64 metadata_)
644 {
645 uint64_t metadata = (OVS_FORCE uint64_t) metadata_;
646 return hash_uint64(metadata);
647 }
648
649 static struct cls_partition *
650 find_partition(const struct classifier *cls, ovs_be64 metadata, uint32_t hash)
651 {
652 struct cls_partition *partition;
653
654 CMAP_FOR_EACH_WITH_HASH (partition, cmap_node, hash, &cls->partitions) {
655 if (partition->metadata == metadata) {
656 return partition;
657 }
658 }
659
660 return NULL;
661 }
662
663 static struct cls_partition *
664 create_partition(struct classifier *cls, struct cls_subtable *subtable,
665 ovs_be64 metadata)
666 OVS_REQUIRES(cls->mutex)
667 {
668 uint32_t hash = hash_metadata(metadata);
669 struct cls_partition *partition = find_partition(cls, metadata, hash);
670 if (!partition) {
671 partition = xmalloc(sizeof *partition);
672 partition->metadata = metadata;
673 partition->tags = 0;
674 tag_tracker_init(&partition->tracker);
675 cmap_insert(&cls->partitions, &partition->cmap_node, hash);
676 }
677 tag_tracker_add(&partition->tracker, &partition->tags, subtable->tag);
678 return partition;
679 }
680
681 static inline ovs_be32 minimatch_get_ports(const struct minimatch *match)
682 {
683 /* Could optimize to use the same map if needed for fast path. */
684 return MINIFLOW_GET_BE32(&match->flow, tp_src)
685 & MINIFLOW_GET_BE32(&match->mask.masks, tp_src);
686 }
687
688 /* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller
689 * must not modify or free it.
690 *
691 * If 'cls' already contains an identical rule (including wildcards, values of
692 * fixed fields, and priority), replaces the old rule by 'rule' and returns the
693 * rule that was replaced. The caller takes ownership of the returned rule and
694 * is thus responsible for destroying it with cls_rule_destroy(), freeing the
695 * memory block in which it resides, etc., as necessary.
696 *
697 * Returns NULL if 'cls' does not contain a rule with an identical key, after
698 * inserting the new rule. In this case, no rules are displaced by the new
699 * rule, even rules that cannot have any effect because the new rule matches a
700 * superset of their flows and has higher priority. */
701 struct cls_rule *
702 classifier_replace(struct classifier *cls, struct cls_rule *rule)
703 OVS_EXCLUDED(cls->mutex)
704 {
705 struct cls_match *old_rule;
706 struct cls_subtable *subtable;
707 struct cls_rule *old_cls_rule = NULL;
708
709 ovs_mutex_lock(&cls->mutex);
710 subtable = find_subtable(cls, &rule->match.mask);
711 if (!subtable) {
712 subtable = insert_subtable(cls, &rule->match.mask);
713 }
714
715 old_rule = insert_rule(cls, subtable, rule);
716 if (!old_rule) {
717 old_cls_rule = NULL;
718
719 rule->cls_match->partition = NULL;
720 if (minimask_get_metadata_mask(&rule->match.mask) == OVS_BE64_MAX) {
721 ovs_be64 metadata = miniflow_get_metadata(&rule->match.flow);
722 rule->cls_match->partition = create_partition(cls, subtable,
723 metadata);
724 }
725
726 cls->n_rules++;
727
728 for (int i = 0; i < cls->n_tries; i++) {
729 if (subtable->trie_plen[i]) {
730 trie_insert(&cls->tries[i], rule, subtable->trie_plen[i]);
731 }
732 }
733
734 /* Ports trie. */
735 if (subtable->ports_mask_len) {
736 /* We mask the value to be inserted to always have the wildcarded
737 * bits in known (zero) state, so we can include them in comparison
738 * and they will always match (== their original value does not
739 * matter). */
740 ovs_be32 masked_ports = minimatch_get_ports(&rule->match);
741
742 trie_insert_prefix(&subtable->ports_trie, &masked_ports,
743 subtable->ports_mask_len);
744 }
745 } else {
746 old_cls_rule = old_rule->cls_rule;
747 rule->cls_match->partition = old_rule->partition;
748 old_cls_rule->cls_match = NULL;
749
750 /* 'old_rule' contains a cmap_node, which may not be freed
751 * immediately. */
752 ovsrcu_postpone(free, old_rule);
753 }
754 ovs_mutex_unlock(&cls->mutex);
755 return old_cls_rule;
756 }
757
758 /* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller
759 * must not modify or free it.
760 *
761 * 'cls' must not contain an identical rule (including wildcards, values of
762 * fixed fields, and priority). Use classifier_find_rule_exactly() to find
763 * such a rule. */
764 void
765 classifier_insert(struct classifier *cls, struct cls_rule *rule)
766 {
767 struct cls_rule *displaced_rule = classifier_replace(cls, rule);
768 ovs_assert(!displaced_rule);
769 }
770
771 /* Removes 'rule' from 'cls'. It is the caller's responsibility to destroy
772 * 'rule' with cls_rule_destroy(), freeing the memory block in which 'rule'
773 * resides, etc., as necessary. */
774 void
775 classifier_remove(struct classifier *cls, struct cls_rule *rule)
776 OVS_EXCLUDED(cls->mutex)
777 {
778 struct cls_partition *partition;
779 struct cls_match *cls_match = rule->cls_match;
780 struct cls_match *head;
781 struct cls_subtable *subtable;
782 int i;
783 uint32_t basis = 0, hash, ihash[CLS_MAX_INDICES];
784 uint8_t prev_be32ofs = 0;
785
786 ovs_assert(cls_match);
787
788 ovs_mutex_lock(&cls->mutex);
789 subtable = find_subtable(cls, &rule->match.mask);
790 ovs_assert(subtable);
791
792 if (subtable->ports_mask_len) {
793 ovs_be32 masked_ports = minimatch_get_ports(&rule->match);
794
795 trie_remove_prefix(&subtable->ports_trie,
796 &masked_ports, subtable->ports_mask_len);
797 }
798 for (i = 0; i < cls->n_tries; i++) {
799 if (subtable->trie_plen[i]) {
800 trie_remove(&cls->tries[i], rule, subtable->trie_plen[i]);
801 }
802 }
803
804 /* Remove rule node from indices. */
805 for (i = 0; i < subtable->n_indices; i++) {
806 ihash[i] = minimatch_hash_range(&rule->match, prev_be32ofs,
807 subtable->index_ofs[i], &basis);
808 cmap_remove(&subtable->indices[i], &cls_match->index_nodes[i],
809 ihash[i]);
810 prev_be32ofs = subtable->index_ofs[i];
811 }
812 hash = minimatch_hash_range(&rule->match, prev_be32ofs, FLOW_U32S, &basis);
813
814 head = find_equal(subtable, &rule->match.flow, hash);
815 if (head != cls_match) {
816 list_remove(&cls_match->list);
817 } else if (list_is_empty(&cls_match->list)) {
818 cmap_remove(&subtable->rules, &cls_match->cmap_node, hash);
819 } else {
820 struct cls_match *next = CONTAINER_OF(cls_match->list.next,
821 struct cls_match, list);
822
823 list_remove(&cls_match->list);
824 cmap_replace(&subtable->rules, &cls_match->cmap_node,
825 &next->cmap_node, hash);
826 }
827
828 partition = cls_match->partition;
829 if (partition) {
830 tag_tracker_subtract(&partition->tracker, &partition->tags,
831 subtable->tag);
832 if (!partition->tags) {
833 cmap_remove(&cls->partitions, &partition->cmap_node,
834 hash_metadata(partition->metadata));
835 ovsrcu_postpone(free, partition);
836 }
837 }
838
839 if (--subtable->n_rules == 0) {
840 destroy_subtable(cls, subtable);
841 } else if (subtable->max_priority == cls_match->priority
842 && --subtable->max_count == 0) {
843 /* Find the new 'max_priority' and 'max_count'. */
844 struct cls_match *head;
845 unsigned int max_priority = 0;
846
847 CMAP_FOR_EACH (head, cmap_node, &subtable->rules) {
848 if (head->priority > max_priority) {
849 max_priority = head->priority;
850 subtable->max_count = 1;
851 } else if (head->priority == max_priority) {
852 ++subtable->max_count;
853 }
854 }
855 subtable->max_priority = max_priority;
856 pvector_change_priority(&cls->subtables, subtable, max_priority);
857 }
858
859 cls->n_rules--;
860
861 rule->cls_match = NULL;
862 ovsrcu_postpone(free, cls_match);
863 ovs_mutex_unlock(&cls->mutex);
864 }
865
866 /* Prefix tree context. Valid when 'lookup_done' is true. Can skip all
867 * subtables which have a prefix match on the trie field, but whose prefix
868 * length is not indicated in 'match_plens'. For example, a subtable that
869 * has a 8-bit trie field prefix match can be skipped if
870 * !be_get_bit_at(&match_plens, 8 - 1). If skipped, 'maskbits' prefix bits
871 * must be unwildcarded to make datapath flow only match packets it should. */
872 struct trie_ctx {
873 const struct cls_trie *trie;
874 bool lookup_done; /* Status of the lookup. */
875 uint8_t be32ofs; /* U32 offset of the field in question. */
876 unsigned int maskbits; /* Prefix length needed to avoid false matches. */
877 union mf_value match_plens; /* Bitmask of prefix lengths with possible
878 * matches. */
879 };
880
881 static void
882 trie_ctx_init(struct trie_ctx *ctx, const struct cls_trie *trie)
883 {
884 ctx->trie = trie;
885 ctx->be32ofs = trie->field->flow_be32ofs;
886 ctx->lookup_done = false;
887 }
888
889 /* Finds and returns the highest-priority rule in 'cls' that matches 'flow'.
890 * Returns a null pointer if no rules in 'cls' match 'flow'. If multiple rules
891 * of equal priority match 'flow', returns one arbitrarily.
892 *
893 * If a rule is found and 'wc' is non-null, bitwise-OR's 'wc' with the
894 * set of bits that were significant in the lookup. At some point
895 * earlier, 'wc' should have been initialized (e.g., by
896 * flow_wildcards_init_catchall()). */
897 struct cls_rule *
898 classifier_lookup(const struct classifier *cls, const struct flow *flow,
899 struct flow_wildcards *wc)
900 {
901 const struct cls_partition *partition;
902 tag_type tags;
903 int64_t best_priority = -1;
904 const struct cls_match *best;
905 struct trie_ctx trie_ctx[CLS_MAX_TRIES];
906 struct cls_subtable *subtable;
907
908 /* Synchronize for cls->n_tries and subtable->trie_plen. They can change
909 * when table configuration changes, which happens typically only on
910 * startup. */
911 atomic_thread_fence(memory_order_acquire);
912
913 /* Determine 'tags' such that, if 'subtable->tag' doesn't intersect them,
914 * then 'flow' cannot possibly match in 'subtable':
915 *
916 * - If flow->metadata maps to a given 'partition', then we can use
917 * 'tags' for 'partition->tags'.
918 *
919 * - If flow->metadata has no partition, then no rule in 'cls' has an
920 * exact-match for flow->metadata. That means that we don't need to
921 * search any subtable that includes flow->metadata in its mask.
922 *
923 * In either case, we always need to search any cls_subtables that do not
924 * include flow->metadata in its mask. One way to do that would be to
925 * check the "cls_subtable"s explicitly for that, but that would require an
926 * extra branch per subtable. Instead, we mark such a cls_subtable's
927 * 'tags' as TAG_ALL and make sure that 'tags' is never empty. This means
928 * that 'tags' always intersects such a cls_subtable's 'tags', so we don't
929 * need a special case.
930 */
931 partition = (cmap_is_empty(&cls->partitions)
932 ? NULL
933 : find_partition(cls, flow->metadata,
934 hash_metadata(flow->metadata)));
935 tags = partition ? partition->tags : TAG_ARBITRARY;
936
937 /* Initialize trie contexts for match_find_wc(). */
938 for (int i = 0; i < cls->n_tries; i++) {
939 trie_ctx_init(&trie_ctx[i], &cls->tries[i]);
940 }
941
942 best = NULL;
943 PVECTOR_FOR_EACH_PRIORITY(subtable, best_priority, 2,
944 sizeof(struct cls_subtable), &cls->subtables) {
945 struct cls_match *rule;
946
947 if (!tag_intersects(tags, subtable->tag)) {
948 continue;
949 }
950
951 rule = find_match_wc(subtable, flow, trie_ctx, cls->n_tries, wc);
952 if (rule && (int64_t)rule->priority > best_priority) {
953 best_priority = (int64_t)rule->priority;
954 best = rule;
955 }
956 }
957
958 return best ? best->cls_rule : NULL;
959 }
960
961 /* Returns true if 'target' satisifies 'match', that is, if each bit for which
962 * 'match' specifies a particular value has the correct value in 'target'.
963 *
964 * 'flow' and 'mask' have the same mask! */
965 static bool
966 miniflow_and_mask_matches_miniflow(const struct miniflow *flow,
967 const struct minimask *mask,
968 const struct miniflow *target)
969 {
970 const uint32_t *flowp = miniflow_get_u32_values(flow);
971 const uint32_t *maskp = miniflow_get_u32_values(&mask->masks);
972 uint32_t target_u32;
973
974 MINIFLOW_FOR_EACH_IN_MAP(target_u32, target, mask->masks.map) {
975 if ((*flowp++ ^ target_u32) & *maskp++) {
976 return false;
977 }
978 }
979
980 return true;
981 }
982
983 static inline struct cls_match *
984 find_match_miniflow(const struct cls_subtable *subtable,
985 const struct miniflow *flow,
986 uint32_t hash)
987 {
988 struct cls_match *rule;
989
990 CMAP_FOR_EACH_WITH_HASH (rule, cmap_node, hash, &subtable->rules) {
991 if (miniflow_and_mask_matches_miniflow(&rule->flow, &subtable->mask,
992 flow)) {
993 return rule;
994 }
995 }
996
997 return NULL;
998 }
999
1000 /* For each miniflow in 'flows' performs a classifier lookup writing the result
1001 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
1002 * NULL it is skipped.
1003 *
1004 * This function is optimized for use in the userspace datapath and therefore
1005 * does not implement a lot of features available in the standard
1006 * classifier_lookup() function. Specifically, it does not implement
1007 * priorities, instead returning any rule which matches the flow. */
1008 void
1009 classifier_lookup_miniflow_batch(const struct classifier *cls,
1010 const struct miniflow **flows,
1011 struct cls_rule **rules, size_t len)
1012 {
1013 struct cls_subtable *subtable;
1014 size_t i, begin = 0;
1015
1016 memset(rules, 0, len * sizeof *rules);
1017 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
1018 for (i = begin; i < len; i++) {
1019 struct cls_match *match;
1020 uint32_t hash;
1021
1022 if (OVS_UNLIKELY(rules[i] || !flows[i])) {
1023 continue;
1024 }
1025
1026 hash = miniflow_hash_in_minimask(flows[i], &subtable->mask, 0);
1027 match = find_match_miniflow(subtable, flows[i], hash);
1028 if (OVS_UNLIKELY(match)) {
1029 rules[i] = match->cls_rule;
1030 }
1031 }
1032
1033 while (begin < len && (rules[begin] || !flows[begin])) {
1034 begin++;
1035 }
1036 if (begin >= len) {
1037 break;
1038 }
1039 }
1040 }
1041
1042 /* Finds and returns a rule in 'cls' with exactly the same priority and
1043 * matching criteria as 'target'. Returns a null pointer if 'cls' doesn't
1044 * contain an exact match. */
1045 struct cls_rule *
1046 classifier_find_rule_exactly(const struct classifier *cls,
1047 const struct cls_rule *target)
1048 OVS_EXCLUDED(cls->mutex)
1049 {
1050 struct cls_match *head, *rule;
1051 struct cls_subtable *subtable;
1052
1053 ovs_mutex_lock(&cls->mutex);
1054 subtable = find_subtable(cls, &target->match.mask);
1055 if (!subtable) {
1056 goto out;
1057 }
1058
1059 /* Skip if there is no hope. */
1060 if (target->priority > subtable->max_priority) {
1061 goto out;
1062 }
1063
1064 head = find_equal(subtable, &target->match.flow,
1065 miniflow_hash_in_minimask(&target->match.flow,
1066 &target->match.mask, 0));
1067 FOR_EACH_RULE_IN_LIST (rule, head) {
1068 if (target->priority >= rule->priority) {
1069 ovs_mutex_unlock(&cls->mutex);
1070 return target->priority == rule->priority ? rule->cls_rule : NULL;
1071 }
1072 }
1073 out:
1074 ovs_mutex_unlock(&cls->mutex);
1075 return NULL;
1076 }
1077
1078 /* Finds and returns a rule in 'cls' with priority 'priority' and exactly the
1079 * same matching criteria as 'target'. Returns a null pointer if 'cls' doesn't
1080 * contain an exact match. */
1081 struct cls_rule *
1082 classifier_find_match_exactly(const struct classifier *cls,
1083 const struct match *target,
1084 unsigned int priority)
1085 {
1086 struct cls_rule *retval;
1087 struct cls_rule cr;
1088
1089 cls_rule_init(&cr, target, priority);
1090 retval = classifier_find_rule_exactly(cls, &cr);
1091 cls_rule_destroy(&cr);
1092
1093 return retval;
1094 }
1095
1096 /* Checks if 'target' would overlap any other rule in 'cls'. Two rules are
1097 * considered to overlap if both rules have the same priority and a packet
1098 * could match both. */
1099 bool
1100 classifier_rule_overlaps(const struct classifier *cls,
1101 const struct cls_rule *target)
1102 OVS_EXCLUDED(cls->mutex)
1103 {
1104 struct cls_subtable *subtable;
1105 int64_t stop_at_priority = (int64_t)target->priority - 1;
1106
1107 ovs_mutex_lock(&cls->mutex);
1108 /* Iterate subtables in the descending max priority order. */
1109 PVECTOR_FOR_EACH_PRIORITY (subtable, stop_at_priority, 2,
1110 sizeof(struct cls_subtable), &cls->subtables) {
1111 uint32_t storage[FLOW_U32S];
1112 struct minimask mask;
1113 struct cls_match *head;
1114
1115 minimask_combine(&mask, &target->match.mask, &subtable->mask, storage);
1116 CMAP_FOR_EACH (head, cmap_node, &subtable->rules) {
1117 struct cls_match *rule;
1118
1119 FOR_EACH_RULE_IN_LIST (rule, head) {
1120 if (rule->priority < target->priority) {
1121 break; /* Rules in descending priority order. */
1122 }
1123 if (rule->priority == target->priority
1124 && miniflow_equal_in_minimask(&target->match.flow,
1125 &rule->flow, &mask)) {
1126 ovs_mutex_unlock(&cls->mutex);
1127 return true;
1128 }
1129 }
1130 }
1131 }
1132
1133 ovs_mutex_unlock(&cls->mutex);
1134 return false;
1135 }
1136
1137 /* Returns true if 'rule' exactly matches 'criteria' or if 'rule' is more
1138 * specific than 'criteria'. That is, 'rule' matches 'criteria' and this
1139 * function returns true if, for every field:
1140 *
1141 * - 'criteria' and 'rule' specify the same (non-wildcarded) value for the
1142 * field, or
1143 *
1144 * - 'criteria' wildcards the field,
1145 *
1146 * Conversely, 'rule' does not match 'criteria' and this function returns false
1147 * if, for at least one field:
1148 *
1149 * - 'criteria' and 'rule' specify different values for the field, or
1150 *
1151 * - 'criteria' specifies a value for the field but 'rule' wildcards it.
1152 *
1153 * Equivalently, the truth table for whether a field matches is:
1154 *
1155 * rule
1156 *
1157 * c wildcard exact
1158 * r +---------+---------+
1159 * i wild | yes | yes |
1160 * t card | | |
1161 * e +---------+---------+
1162 * r exact | no |if values|
1163 * i | |are equal|
1164 * a +---------+---------+
1165 *
1166 * This is the matching rule used by OpenFlow 1.0 non-strict OFPT_FLOW_MOD
1167 * commands and by OpenFlow 1.0 aggregate and flow stats.
1168 *
1169 * Ignores rule->priority. */
1170 bool
1171 cls_rule_is_loose_match(const struct cls_rule *rule,
1172 const struct minimatch *criteria)
1173 {
1174 return (!minimask_has_extra(&rule->match.mask, &criteria->mask)
1175 && miniflow_equal_in_minimask(&rule->match.flow, &criteria->flow,
1176 &criteria->mask));
1177 }
1178 \f
1179 /* Iteration. */
1180
1181 static bool
1182 rule_matches(const struct cls_match *rule, const struct cls_rule *target)
1183 {
1184 return (!target
1185 || miniflow_equal_in_minimask(&rule->flow,
1186 &target->match.flow,
1187 &target->match.mask));
1188 }
1189
1190 static struct cls_match *
1191 search_subtable(const struct cls_subtable *subtable,
1192 struct cls_cursor *cursor)
1193 {
1194 if (!cursor->target
1195 || !minimask_has_extra(&subtable->mask, &cursor->target->match.mask)) {
1196 struct cls_match *rule;
1197
1198 CMAP_CURSOR_FOR_EACH (rule, cmap_node, &cursor->rules,
1199 &subtable->rules) {
1200 if (rule_matches(rule, cursor->target)) {
1201 return rule;
1202 }
1203 }
1204 }
1205 return NULL;
1206 }
1207
1208 /* Initializes 'cursor' for iterating through rules in 'cls', and returns the
1209 * first matching cls_rule via '*pnode', or NULL if there are no matches.
1210 *
1211 * - If 'target' is null, the cursor will visit every rule in 'cls'.
1212 *
1213 * - If 'target' is nonnull, the cursor will visit each 'rule' in 'cls'
1214 * such that cls_rule_is_loose_match(rule, target) returns true.
1215 *
1216 * Ignores target->priority. */
1217 struct cls_cursor cls_cursor_start(const struct classifier *cls,
1218 const struct cls_rule *target,
1219 bool safe)
1220 OVS_NO_THREAD_SAFETY_ANALYSIS
1221 {
1222 struct cls_cursor cursor;
1223 struct cls_subtable *subtable;
1224
1225 cursor.safe = safe;
1226 cursor.cls = cls;
1227 cursor.target = target && !cls_rule_is_catchall(target) ? target : NULL;
1228 cursor.rule = NULL;
1229
1230 /* Find first rule. */
1231 ovs_mutex_lock(&cursor.cls->mutex);
1232 CMAP_CURSOR_FOR_EACH (subtable, cmap_node, &cursor.subtables,
1233 &cursor.cls->subtables_map) {
1234 struct cls_match *rule = search_subtable(subtable, &cursor);
1235
1236 if (rule) {
1237 cursor.subtable = subtable;
1238 cursor.rule = rule->cls_rule;
1239 break;
1240 }
1241 }
1242
1243 /* Leave locked if requested and have a rule. */
1244 if (safe || !cursor.rule) {
1245 ovs_mutex_unlock(&cursor.cls->mutex);
1246 }
1247 return cursor;
1248 }
1249
1250 static struct cls_rule *
1251 cls_cursor_next(struct cls_cursor *cursor)
1252 OVS_NO_THREAD_SAFETY_ANALYSIS
1253 {
1254 struct cls_match *rule = cursor->rule->cls_match;
1255 const struct cls_subtable *subtable;
1256 struct cls_match *next;
1257
1258 next = next_rule_in_list__(rule);
1259 if (next->priority < rule->priority) {
1260 return next->cls_rule;
1261 }
1262
1263 /* 'next' is the head of the list, that is, the rule that is included in
1264 * the subtable's map. (This is important when the classifier contains
1265 * rules that differ only in priority.) */
1266 rule = next;
1267 CMAP_CURSOR_FOR_EACH_CONTINUE (rule, cmap_node, &cursor->rules) {
1268 if (rule_matches(rule, cursor->target)) {
1269 return rule->cls_rule;
1270 }
1271 }
1272
1273 subtable = cursor->subtable;
1274 CMAP_CURSOR_FOR_EACH_CONTINUE (subtable, cmap_node, &cursor->subtables) {
1275 rule = search_subtable(subtable, cursor);
1276 if (rule) {
1277 cursor->subtable = subtable;
1278 return rule->cls_rule;
1279 }
1280 }
1281
1282 return NULL;
1283 }
1284
1285 /* Sets 'cursor->rule' to the next matching cls_rule in 'cursor''s iteration,
1286 * or to null if all matching rules have been visited. */
1287 void
1288 cls_cursor_advance(struct cls_cursor *cursor)
1289 OVS_NO_THREAD_SAFETY_ANALYSIS
1290 {
1291 if (cursor->safe) {
1292 ovs_mutex_lock(&cursor->cls->mutex);
1293 }
1294 cursor->rule = cls_cursor_next(cursor);
1295 if (cursor->safe || !cursor->rule) {
1296 ovs_mutex_unlock(&cursor->cls->mutex);
1297 }
1298 }
1299 \f
1300 static struct cls_subtable *
1301 find_subtable(const struct classifier *cls, const struct minimask *mask)
1302 OVS_REQUIRES(cls->mutex)
1303 {
1304 struct cls_subtable *subtable;
1305
1306 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, minimask_hash(mask, 0),
1307 &cls->subtables_map) {
1308 if (minimask_equal(mask, &subtable->mask)) {
1309 return subtable;
1310 }
1311 }
1312 return NULL;
1313 }
1314
1315 /* The new subtable will be visible to the readers only after this. */
1316 static struct cls_subtable *
1317 insert_subtable(struct classifier *cls, const struct minimask *mask)
1318 OVS_REQUIRES(cls->mutex)
1319 {
1320 uint32_t hash = minimask_hash(mask, 0);
1321 struct cls_subtable *subtable;
1322 int i, index = 0;
1323 struct flow_wildcards old, new;
1324 uint8_t prev;
1325 int count = count_1bits(mask->masks.map);
1326
1327 subtable = xzalloc(sizeof *subtable - sizeof mask->masks.inline_values
1328 + MINIFLOW_VALUES_SIZE(count));
1329 cmap_init(&subtable->rules);
1330 miniflow_clone_inline(&subtable->mask.masks, &mask->masks, count);
1331
1332 /* Init indices for segmented lookup, if any. */
1333 flow_wildcards_init_catchall(&new);
1334 old = new;
1335 prev = 0;
1336 for (i = 0; i < cls->n_flow_segments; i++) {
1337 flow_wildcards_fold_minimask_range(&new, mask, prev,
1338 cls->flow_segments[i]);
1339 /* Add an index if it adds mask bits. */
1340 if (!flow_wildcards_equal(&new, &old)) {
1341 cmap_init(&subtable->indices[index]);
1342 subtable->index_ofs[index] = cls->flow_segments[i];
1343 index++;
1344 old = new;
1345 }
1346 prev = cls->flow_segments[i];
1347 }
1348 /* Check if the rest of the subtable's mask adds any bits,
1349 * and remove the last index if it doesn't. */
1350 if (index > 0) {
1351 flow_wildcards_fold_minimask_range(&new, mask, prev, FLOW_U32S);
1352 if (flow_wildcards_equal(&new, &old)) {
1353 --index;
1354 subtable->index_ofs[index] = 0;
1355 cmap_destroy(&subtable->indices[index]);
1356 }
1357 }
1358 subtable->n_indices = index;
1359
1360 subtable->tag = (minimask_get_metadata_mask(mask) == OVS_BE64_MAX
1361 ? tag_create_deterministic(hash)
1362 : TAG_ALL);
1363
1364 for (i = 0; i < cls->n_tries; i++) {
1365 subtable->trie_plen[i] = minimask_get_prefix_len(mask,
1366 cls->tries[i].field);
1367 }
1368
1369 /* Ports trie. */
1370 ovsrcu_set_hidden(&subtable->ports_trie, NULL);
1371 subtable->ports_mask_len
1372 = 32 - ctz32(ntohl(MINIFLOW_GET_BE32(&mask->masks, tp_src)));
1373
1374 cmap_insert(&cls->subtables_map, &subtable->cmap_node, hash);
1375
1376 return subtable;
1377 }
1378
1379 static void
1380 destroy_subtable(struct classifier *cls, struct cls_subtable *subtable)
1381 OVS_REQUIRES(cls->mutex)
1382 {
1383 int i;
1384
1385 pvector_remove(&cls->subtables, subtable);
1386 trie_destroy(&subtable->ports_trie);
1387
1388 for (i = 0; i < subtable->n_indices; i++) {
1389 cmap_destroy(&subtable->indices[i]);
1390 }
1391 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
1392 minimask_hash(&subtable->mask, 0));
1393 minimask_destroy(&subtable->mask);
1394 cmap_destroy(&subtable->rules);
1395 ovsrcu_postpone(free, subtable);
1396 }
1397
1398 struct range {
1399 uint8_t start;
1400 uint8_t end;
1401 };
1402
1403 static unsigned int be_get_bit_at(const ovs_be32 value[], unsigned int ofs);
1404
1405 /* Return 'true' if can skip rest of the subtable based on the prefix trie
1406 * lookup results. */
1407 static inline bool
1408 check_tries(struct trie_ctx trie_ctx[CLS_MAX_TRIES], unsigned int n_tries,
1409 const unsigned int field_plen[CLS_MAX_TRIES],
1410 const struct range ofs, const struct flow *flow,
1411 struct flow_wildcards *wc)
1412 {
1413 int j;
1414
1415 /* Check if we could avoid fully unwildcarding the next level of
1416 * fields using the prefix tries. The trie checks are done only as
1417 * needed to avoid folding in additional bits to the wildcards mask. */
1418 for (j = 0; j < n_tries; j++) {
1419 /* Is the trie field relevant for this subtable? */
1420 if (field_plen[j]) {
1421 struct trie_ctx *ctx = &trie_ctx[j];
1422 uint8_t be32ofs = ctx->be32ofs;
1423
1424 /* Is the trie field within the current range of fields? */
1425 if (be32ofs >= ofs.start && be32ofs < ofs.end) {
1426 /* On-demand trie lookup. */
1427 if (!ctx->lookup_done) {
1428 memset(&ctx->match_plens, 0, sizeof ctx->match_plens);
1429 ctx->maskbits = trie_lookup(ctx->trie, flow,
1430 &ctx->match_plens);
1431 ctx->lookup_done = true;
1432 }
1433 /* Possible to skip the rest of the subtable if subtable's
1434 * prefix on the field is not included in the lookup result. */
1435 if (!be_get_bit_at(&ctx->match_plens.be32, field_plen[j] - 1)) {
1436 /* We want the trie lookup to never result in unwildcarding
1437 * any bits that would not be unwildcarded otherwise.
1438 * Since the trie is shared by the whole classifier, it is
1439 * possible that the 'maskbits' contain bits that are
1440 * irrelevant for the partition relevant for the current
1441 * packet. Hence the checks below. */
1442
1443 /* Check that the trie result will not unwildcard more bits
1444 * than this subtable would otherwise. */
1445 if (ctx->maskbits <= field_plen[j]) {
1446 /* Unwildcard the bits and skip the rest. */
1447 mask_set_prefix_bits(wc, be32ofs, ctx->maskbits);
1448 /* Note: Prerequisite already unwildcarded, as the only
1449 * prerequisite of the supported trie lookup fields is
1450 * the ethertype, which is always unwildcarded. */
1451 return true;
1452 }
1453 /* Can skip if the field is already unwildcarded. */
1454 if (mask_prefix_bits_set(wc, be32ofs, ctx->maskbits)) {
1455 return true;
1456 }
1457 }
1458 }
1459 }
1460 }
1461 return false;
1462 }
1463
1464 /* Returns true if 'target' satisifies 'flow'/'mask', that is, if each bit
1465 * for which 'flow', for which 'mask' has a bit set, specifies a particular
1466 * value has the correct value in 'target'.
1467 *
1468 * This function is equivalent to miniflow_equal_flow_in_minimask(flow,
1469 * target, mask) but this is faster because of the invariant that
1470 * flow->map and mask->masks.map are the same, and that this version
1471 * takes the 'wc'. */
1472 static inline bool
1473 miniflow_and_mask_matches_flow(const struct miniflow *flow,
1474 const struct minimask *mask,
1475 const struct flow *target)
1476 {
1477 const uint32_t *flowp = miniflow_get_u32_values(flow);
1478 const uint32_t *maskp = miniflow_get_u32_values(&mask->masks);
1479 uint32_t idx;
1480
1481 MAP_FOR_EACH_INDEX(idx, mask->masks.map) {
1482 uint32_t diff = (*flowp++ ^ flow_u32_value(target, idx)) & *maskp++;
1483
1484 if (diff) {
1485 return false;
1486 }
1487 }
1488
1489 return true;
1490 }
1491
1492 static inline struct cls_match *
1493 find_match(const struct cls_subtable *subtable, const struct flow *flow,
1494 uint32_t hash)
1495 {
1496 struct cls_match *rule;
1497
1498 CMAP_FOR_EACH_WITH_HASH (rule, cmap_node, hash, &subtable->rules) {
1499 if (miniflow_and_mask_matches_flow(&rule->flow, &subtable->mask,
1500 flow)) {
1501 return rule;
1502 }
1503 }
1504
1505 return NULL;
1506 }
1507
1508 /* Returns true if 'target' satisifies 'flow'/'mask', that is, if each bit
1509 * for which 'flow', for which 'mask' has a bit set, specifies a particular
1510 * value has the correct value in 'target'.
1511 *
1512 * This function is equivalent to miniflow_and_mask_matches_flow() but this
1513 * version fills in the mask bits in 'wc'. */
1514 static inline bool
1515 miniflow_and_mask_matches_flow_wc(const struct miniflow *flow,
1516 const struct minimask *mask,
1517 const struct flow *target,
1518 struct flow_wildcards *wc)
1519 {
1520 const uint32_t *flowp = miniflow_get_u32_values(flow);
1521 const uint32_t *maskp = miniflow_get_u32_values(&mask->masks);
1522 uint32_t idx;
1523
1524 MAP_FOR_EACH_INDEX(idx, mask->masks.map) {
1525 uint32_t mask = *maskp++;
1526 uint32_t diff = (*flowp++ ^ flow_u32_value(target, idx)) & mask;
1527
1528 if (diff) {
1529 /* Only unwildcard if none of the differing bits is already
1530 * exact-matched. */
1531 if (!(flow_u32_value(&wc->masks, idx) & diff)) {
1532 /* Keep one bit of the difference. */
1533 *flow_u32_lvalue(&wc->masks, idx) |= rightmost_1bit(diff);
1534 }
1535 return false;
1536 }
1537 /* Fill in the bits that were looked at. */
1538 *flow_u32_lvalue(&wc->masks, idx) |= mask;
1539 }
1540
1541 return true;
1542 }
1543
1544 /* Unwildcard the fields looked up so far, if any. */
1545 static void
1546 fill_range_wc(const struct cls_subtable *subtable, struct flow_wildcards *wc,
1547 uint8_t to)
1548 {
1549 if (to) {
1550 flow_wildcards_fold_minimask_range(wc, &subtable->mask, 0, to);
1551 }
1552 }
1553
1554 static struct cls_match *
1555 find_match_wc(const struct cls_subtable *subtable, const struct flow *flow,
1556 struct trie_ctx trie_ctx[CLS_MAX_TRIES], unsigned int n_tries,
1557 struct flow_wildcards *wc)
1558 {
1559 uint32_t basis = 0, hash;
1560 struct cls_match *rule;
1561 int i;
1562 struct range ofs;
1563
1564 if (OVS_UNLIKELY(!wc)) {
1565 return find_match(subtable, flow,
1566 flow_hash_in_minimask(flow, &subtable->mask, 0));
1567 }
1568
1569 ofs.start = 0;
1570 /* Try to finish early by checking fields in segments. */
1571 for (i = 0; i < subtable->n_indices; i++) {
1572 struct cmap_node *inode;
1573
1574 ofs.end = subtable->index_ofs[i];
1575
1576 if (check_tries(trie_ctx, n_tries, subtable->trie_plen, ofs, flow,
1577 wc)) {
1578 /* 'wc' bits for the trie field set, now unwildcard the preceding
1579 * bits used so far. */
1580 fill_range_wc(subtable, wc, ofs.start);
1581 return NULL;
1582 }
1583 hash = flow_hash_in_minimask_range(flow, &subtable->mask, ofs.start,
1584 ofs.end, &basis);
1585 inode = cmap_find(&subtable->indices[i], hash);
1586 if (!inode) {
1587 /* No match, can stop immediately, but must fold in the bits
1588 * used in lookup so far. */
1589 fill_range_wc(subtable, wc, ofs.end);
1590 return NULL;
1591 }
1592
1593 /* If we have narrowed down to a single rule already, check whether
1594 * that rule matches. Either way, we're done.
1595 *
1596 * (Rare) hash collisions may cause us to miss the opportunity for this
1597 * optimization. */
1598 if (!cmap_node_next(inode)) {
1599 ASSIGN_CONTAINER(rule, inode - i, index_nodes);
1600 if (miniflow_and_mask_matches_flow_wc(&rule->flow, &subtable->mask,
1601 flow, wc)) {
1602 return rule;
1603 }
1604 return NULL;
1605 }
1606 ofs.start = ofs.end;
1607 }
1608 ofs.end = FLOW_U32S;
1609 /* Trie check for the final range. */
1610 if (check_tries(trie_ctx, n_tries, subtable->trie_plen, ofs, flow, wc)) {
1611 fill_range_wc(subtable, wc, ofs.start);
1612 return NULL;
1613 }
1614 hash = flow_hash_in_minimask_range(flow, &subtable->mask, ofs.start,
1615 ofs.end, &basis);
1616 rule = find_match(subtable, flow, hash);
1617 if (!rule && subtable->ports_mask_len) {
1618 /* Ports are always part of the final range, if any.
1619 * No match was found for the ports. Use the ports trie to figure out
1620 * which ports bits to unwildcard. */
1621 unsigned int mbits;
1622 ovs_be32 value, plens, mask;
1623
1624 mask = MINIFLOW_GET_BE32(&subtable->mask.masks, tp_src);
1625 value = ((OVS_FORCE ovs_be32 *)flow)[TP_PORTS_OFS32] & mask;
1626 mbits = trie_lookup_value(&subtable->ports_trie, &value, &plens, 32);
1627
1628 ((OVS_FORCE ovs_be32 *)&wc->masks)[TP_PORTS_OFS32] |=
1629 mask & htonl(~0 << (32 - mbits));
1630
1631 /* Unwildcard all bits in the mask upto the ports, as they were used
1632 * to determine there is no match. */
1633 fill_range_wc(subtable, wc, TP_PORTS_OFS32);
1634 return NULL;
1635 }
1636
1637 /* Must unwildcard all the fields, as they were looked at. */
1638 flow_wildcards_fold_minimask(wc, &subtable->mask);
1639 return rule;
1640 }
1641
1642 static struct cls_match *
1643 find_equal(struct cls_subtable *subtable, const struct miniflow *flow,
1644 uint32_t hash)
1645 {
1646 struct cls_match *head;
1647
1648 CMAP_FOR_EACH_WITH_HASH (head, cmap_node, hash, &subtable->rules) {
1649 if (miniflow_equal(&head->flow, flow)) {
1650 return head;
1651 }
1652 }
1653 return NULL;
1654 }
1655
1656 /*
1657 * As the readers are operating concurrently with the modifications, a
1658 * concurrent reader may or may not see the new rule, depending on how
1659 * the concurrent events overlap with each other. This is no
1660 * different from the former locked behavior, but there the visibility
1661 * of the new rule only depended on the timing of the locking
1662 * functions.
1663 *
1664 * The new rule is first added to the segment indices, so the readers
1665 * may find the rule in the indices before the rule is visible in the
1666 * subtables 'rules' map. This may result in us losing the
1667 * opportunity to quit lookups earlier, resulting in sub-optimal
1668 * wildcarding. This will be fixed by forthcoming revalidation always
1669 * scheduled after flow table changes.
1670 *
1671 * Similar behavior may happen due to us removing the overlapping rule
1672 * (if any) from the indices only after the new rule has been added.
1673 *
1674 * The subtable's max priority is updated only after the rule is
1675 * inserted, so the concurrent readers may not see the rule, as the
1676 * updated priority ordered subtable list will only be visible after
1677 * the subtable's max priority is updated.
1678 *
1679 * Similarly, the classifier's partitions for new rules are updated by
1680 * the caller after this function, so the readers may keep skipping
1681 * the subtable until they see the updated partitions.
1682 */
1683 static struct cls_match *
1684 insert_rule(struct classifier *cls, struct cls_subtable *subtable,
1685 struct cls_rule *new_rule)
1686 OVS_REQUIRES(cls->mutex)
1687 {
1688 struct cls_match *old = NULL;
1689 struct cls_match *new = cls_match_alloc(new_rule);
1690 struct cls_match *head;
1691 int i;
1692 uint32_t basis = 0, hash, ihash[CLS_MAX_INDICES];
1693 uint8_t prev_be32ofs = 0;
1694
1695 /* Add new node to segment indices. */
1696 for (i = 0; i < subtable->n_indices; i++) {
1697 ihash[i] = minimatch_hash_range(&new_rule->match, prev_be32ofs,
1698 subtable->index_ofs[i], &basis);
1699 cmap_insert(&subtable->indices[i], &new->index_nodes[i], ihash[i]);
1700 prev_be32ofs = subtable->index_ofs[i];
1701 }
1702 hash = minimatch_hash_range(&new_rule->match, prev_be32ofs, FLOW_U32S,
1703 &basis);
1704 head = find_equal(subtable, &new_rule->match.flow, hash);
1705 if (!head) {
1706 cmap_insert(&subtable->rules, &new->cmap_node, hash);
1707 list_init(&new->list);
1708 goto out;
1709 } else {
1710 /* Scan the list for the insertion point that will keep the list in
1711 * order of decreasing priority. */
1712 struct cls_match *rule;
1713
1714 FOR_EACH_RULE_IN_LIST (rule, head) {
1715 if (new->priority >= rule->priority) {
1716 if (rule == head) {
1717 /* 'new' is the new highest-priority flow in the list. */
1718 cmap_replace(&subtable->rules, &rule->cmap_node,
1719 &new->cmap_node, hash);
1720 }
1721
1722 if (new->priority == rule->priority) {
1723 list_replace(&new->list, &rule->list);
1724 old = rule;
1725 } else {
1726 list_insert(&rule->list, &new->list);
1727 }
1728 goto out;
1729 }
1730 }
1731
1732 /* Insert 'new' at the end of the list. */
1733 list_push_back(&head->list, &new->list);
1734 }
1735
1736 out:
1737 if (!old) {
1738 subtable->n_rules++;
1739
1740 /* Rule was added, not replaced. Update 'subtable's 'max_priority'
1741 * and 'max_count', if necessary. */
1742 if (subtable->n_rules == 1) {
1743 subtable->max_priority = new->priority;
1744 subtable->max_count = 1;
1745 pvector_insert(&cls->subtables, subtable, new->priority);
1746 } else if (subtable->max_priority == new->priority) {
1747 ++subtable->max_count;
1748 } else if (new->priority > subtable->max_priority) {
1749 subtable->max_priority = new->priority;
1750 subtable->max_count = 1;
1751 pvector_change_priority(&cls->subtables, subtable, new->priority);
1752 }
1753 } else {
1754 /* Remove old node from indices. */
1755 for (i = 0; i < subtable->n_indices; i++) {
1756 cmap_remove(&subtable->indices[i], &old->index_nodes[i], ihash[i]);
1757 }
1758 }
1759 return old;
1760 }
1761
1762 static struct cls_match *
1763 next_rule_in_list__(struct cls_match *rule)
1764 OVS_NO_THREAD_SAFETY_ANALYSIS
1765 {
1766 struct cls_match *next = OBJECT_CONTAINING(rule->list.next, next, list);
1767 return next;
1768 }
1769
1770 static struct cls_match *
1771 next_rule_in_list(struct cls_match *rule)
1772 {
1773 struct cls_match *next = next_rule_in_list__(rule);
1774 return next->priority < rule->priority ? next : NULL;
1775 }
1776 \f
1777 /* A longest-prefix match tree. */
1778 struct trie_node {
1779 uint32_t prefix; /* Prefix bits for this node, MSB first. */
1780 uint8_t n_bits; /* Never zero, except for the root node. */
1781 unsigned int n_rules; /* Number of rules that have this prefix. */
1782 rcu_trie_ptr edges[2]; /* Both NULL if leaf. */
1783 };
1784
1785 /* Max bits per node. Must fit in struct trie_node's 'prefix'.
1786 * Also tested with 16, 8, and 5 to stress the implementation. */
1787 #define TRIE_PREFIX_BITS 32
1788
1789 /* Return at least 'plen' bits of the 'prefix', starting at bit offset 'ofs'.
1790 * Prefixes are in the network byte order, and the offset 0 corresponds to
1791 * the most significant bit of the first byte. The offset can be read as
1792 * "how many bits to skip from the start of the prefix starting at 'pr'". */
1793 static uint32_t
1794 raw_get_prefix(const ovs_be32 pr[], unsigned int ofs, unsigned int plen)
1795 {
1796 uint32_t prefix;
1797
1798 pr += ofs / 32; /* Where to start. */
1799 ofs %= 32; /* How many bits to skip at 'pr'. */
1800
1801 prefix = ntohl(*pr) << ofs; /* Get the first 32 - ofs bits. */
1802 if (plen > 32 - ofs) { /* Need more than we have already? */
1803 prefix |= ntohl(*++pr) >> (32 - ofs);
1804 }
1805 /* Return with possible unwanted bits at the end. */
1806 return prefix;
1807 }
1808
1809 /* Return min(TRIE_PREFIX_BITS, plen) bits of the 'prefix', starting at bit
1810 * offset 'ofs'. Prefixes are in the network byte order, and the offset 0
1811 * corresponds to the most significant bit of the first byte. The offset can
1812 * be read as "how many bits to skip from the start of the prefix starting at
1813 * 'pr'". */
1814 static uint32_t
1815 trie_get_prefix(const ovs_be32 pr[], unsigned int ofs, unsigned int plen)
1816 {
1817 if (!plen) {
1818 return 0;
1819 }
1820 if (plen > TRIE_PREFIX_BITS) {
1821 plen = TRIE_PREFIX_BITS; /* Get at most TRIE_PREFIX_BITS. */
1822 }
1823 /* Return with unwanted bits cleared. */
1824 return raw_get_prefix(pr, ofs, plen) & ~0u << (32 - plen);
1825 }
1826
1827 /* Return the number of equal bits in 'n_bits' of 'prefix's MSBs and a 'value'
1828 * starting at "MSB 0"-based offset 'ofs'. */
1829 static unsigned int
1830 prefix_equal_bits(uint32_t prefix, unsigned int n_bits, const ovs_be32 value[],
1831 unsigned int ofs)
1832 {
1833 uint64_t diff = prefix ^ raw_get_prefix(value, ofs, n_bits);
1834 /* Set the bit after the relevant bits to limit the result. */
1835 return raw_clz64(diff << 32 | UINT64_C(1) << (63 - n_bits));
1836 }
1837
1838 /* Return the number of equal bits in 'node' prefix and a 'prefix' of length
1839 * 'plen', starting at "MSB 0"-based offset 'ofs'. */
1840 static unsigned int
1841 trie_prefix_equal_bits(const struct trie_node *node, const ovs_be32 prefix[],
1842 unsigned int ofs, unsigned int plen)
1843 {
1844 return prefix_equal_bits(node->prefix, MIN(node->n_bits, plen - ofs),
1845 prefix, ofs);
1846 }
1847
1848 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' can
1849 * be greater than 31. */
1850 static unsigned int
1851 be_get_bit_at(const ovs_be32 value[], unsigned int ofs)
1852 {
1853 return (((const uint8_t *)value)[ofs / 8] >> (7 - ofs % 8)) & 1u;
1854 }
1855
1856 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' must
1857 * be between 0 and 31, inclusive. */
1858 static unsigned int
1859 get_bit_at(const uint32_t prefix, unsigned int ofs)
1860 {
1861 return (prefix >> (31 - ofs)) & 1u;
1862 }
1863
1864 /* Create new branch. */
1865 static struct trie_node *
1866 trie_branch_create(const ovs_be32 *prefix, unsigned int ofs, unsigned int plen,
1867 unsigned int n_rules)
1868 {
1869 struct trie_node *node = xmalloc(sizeof *node);
1870
1871 node->prefix = trie_get_prefix(prefix, ofs, plen);
1872
1873 if (plen <= TRIE_PREFIX_BITS) {
1874 node->n_bits = plen;
1875 ovsrcu_set_hidden(&node->edges[0], NULL);
1876 ovsrcu_set_hidden(&node->edges[1], NULL);
1877 node->n_rules = n_rules;
1878 } else { /* Need intermediate nodes. */
1879 struct trie_node *subnode = trie_branch_create(prefix,
1880 ofs + TRIE_PREFIX_BITS,
1881 plen - TRIE_PREFIX_BITS,
1882 n_rules);
1883 int bit = get_bit_at(subnode->prefix, 0);
1884 node->n_bits = TRIE_PREFIX_BITS;
1885 ovsrcu_set_hidden(&node->edges[bit], subnode);
1886 ovsrcu_set_hidden(&node->edges[!bit], NULL);
1887 node->n_rules = 0;
1888 }
1889 return node;
1890 }
1891
1892 static void
1893 trie_node_destroy(const struct trie_node *node)
1894 {
1895 ovsrcu_postpone(free, CONST_CAST(struct trie_node *, node));
1896 }
1897
1898 /* Copy a trie node for modification and postpone delete the old one. */
1899 static struct trie_node *
1900 trie_node_rcu_realloc(const struct trie_node *node)
1901 {
1902 struct trie_node *new_node = xmalloc(sizeof *node);
1903
1904 *new_node = *node;
1905 trie_node_destroy(node);
1906
1907 return new_node;
1908 }
1909
1910 /* May only be called while holding the classifier mutex. */
1911 static void
1912 trie_destroy(rcu_trie_ptr *trie)
1913 {
1914 struct trie_node *node = ovsrcu_get_protected(struct trie_node *, trie);
1915
1916 if (node) {
1917 ovsrcu_set_hidden(trie, NULL);
1918 trie_destroy(&node->edges[0]);
1919 trie_destroy(&node->edges[1]);
1920 trie_node_destroy(node);
1921 }
1922 }
1923
1924 static bool
1925 trie_is_leaf(const struct trie_node *trie)
1926 {
1927 /* No children? */
1928 return !ovsrcu_get(struct trie_node *, &trie->edges[0])
1929 && !ovsrcu_get(struct trie_node *, &trie->edges[1]);
1930 }
1931
1932 static void
1933 mask_set_prefix_bits(struct flow_wildcards *wc, uint8_t be32ofs,
1934 unsigned int n_bits)
1935 {
1936 ovs_be32 *mask = &((ovs_be32 *)&wc->masks)[be32ofs];
1937 unsigned int i;
1938
1939 for (i = 0; i < n_bits / 32; i++) {
1940 mask[i] = OVS_BE32_MAX;
1941 }
1942 if (n_bits % 32) {
1943 mask[i] |= htonl(~0u << (32 - n_bits % 32));
1944 }
1945 }
1946
1947 static bool
1948 mask_prefix_bits_set(const struct flow_wildcards *wc, uint8_t be32ofs,
1949 unsigned int n_bits)
1950 {
1951 ovs_be32 *mask = &((ovs_be32 *)&wc->masks)[be32ofs];
1952 unsigned int i;
1953 ovs_be32 zeroes = 0;
1954
1955 for (i = 0; i < n_bits / 32; i++) {
1956 zeroes |= ~mask[i];
1957 }
1958 if (n_bits % 32) {
1959 zeroes |= ~mask[i] & htonl(~0u << (32 - n_bits % 32));
1960 }
1961
1962 return !zeroes; /* All 'n_bits' bits set. */
1963 }
1964
1965 static rcu_trie_ptr *
1966 trie_next_edge(struct trie_node *node, const ovs_be32 value[],
1967 unsigned int ofs)
1968 {
1969 return node->edges + be_get_bit_at(value, ofs);
1970 }
1971
1972 static const struct trie_node *
1973 trie_next_node(const struct trie_node *node, const ovs_be32 value[],
1974 unsigned int ofs)
1975 {
1976 return ovsrcu_get(struct trie_node *,
1977 &node->edges[be_get_bit_at(value, ofs)]);
1978 }
1979
1980 /* Set the bit at ("MSB 0"-based) offset 'ofs'. 'ofs' can be greater than 31.
1981 */
1982 static void
1983 be_set_bit_at(ovs_be32 value[], unsigned int ofs)
1984 {
1985 ((uint8_t *)value)[ofs / 8] |= 1u << (7 - ofs % 8);
1986 }
1987
1988 /* Returns the number of bits in the prefix mask necessary to determine a
1989 * mismatch, in case there are longer prefixes in the tree below the one that
1990 * matched.
1991 * '*plens' will have a bit set for each prefix length that may have matching
1992 * rules. The caller is responsible for clearing the '*plens' prior to
1993 * calling this.
1994 */
1995 static unsigned int
1996 trie_lookup_value(const rcu_trie_ptr *trie, const ovs_be32 value[],
1997 ovs_be32 plens[], unsigned int n_bits)
1998 {
1999 const struct trie_node *prev = NULL;
2000 const struct trie_node *node = ovsrcu_get(struct trie_node *, trie);
2001 unsigned int match_len = 0; /* Number of matching bits. */
2002
2003 for (; node; prev = node, node = trie_next_node(node, value, match_len)) {
2004 unsigned int eqbits;
2005 /* Check if this edge can be followed. */
2006 eqbits = prefix_equal_bits(node->prefix, node->n_bits, value,
2007 match_len);
2008 match_len += eqbits;
2009 if (eqbits < node->n_bits) { /* Mismatch, nothing more to be found. */
2010 /* Bit at offset 'match_len' differed. */
2011 return match_len + 1; /* Includes the first mismatching bit. */
2012 }
2013 /* Full match, check if rules exist at this prefix length. */
2014 if (node->n_rules > 0) {
2015 be_set_bit_at(plens, match_len - 1);
2016 }
2017 if (match_len >= n_bits) {
2018 return n_bits; /* Full prefix. */
2019 }
2020 }
2021 /* node == NULL. Full match so far, but we tried to follow an
2022 * non-existing branch. Need to exclude the other branch if it exists
2023 * (it does not if we were called on an empty trie or 'prev' is a leaf
2024 * node). */
2025 return !prev || trie_is_leaf(prev) ? match_len : match_len + 1;
2026 }
2027
2028 static unsigned int
2029 trie_lookup(const struct cls_trie *trie, const struct flow *flow,
2030 union mf_value *plens)
2031 {
2032 const struct mf_field *mf = trie->field;
2033
2034 /* Check that current flow matches the prerequisites for the trie
2035 * field. Some match fields are used for multiple purposes, so we
2036 * must check that the trie is relevant for this flow. */
2037 if (mf_are_prereqs_ok(mf, flow)) {
2038 return trie_lookup_value(&trie->root,
2039 &((ovs_be32 *)flow)[mf->flow_be32ofs],
2040 &plens->be32, mf->n_bits);
2041 }
2042 memset(plens, 0xff, sizeof *plens); /* All prefixes, no skipping. */
2043 return 0; /* Value not used in this case. */
2044 }
2045
2046 /* Returns the length of a prefix match mask for the field 'mf' in 'minimask'.
2047 * Returns the u32 offset to the miniflow data in '*miniflow_index', if
2048 * 'miniflow_index' is not NULL. */
2049 static unsigned int
2050 minimask_get_prefix_len(const struct minimask *minimask,
2051 const struct mf_field *mf)
2052 {
2053 unsigned int n_bits = 0, mask_tz = 0; /* Non-zero when end of mask seen. */
2054 uint8_t u32_ofs = mf->flow_be32ofs;
2055 uint8_t u32_end = u32_ofs + mf->n_bytes / 4;
2056
2057 for (; u32_ofs < u32_end; ++u32_ofs) {
2058 uint32_t mask;
2059 mask = ntohl((OVS_FORCE ovs_be32)minimask_get(minimask, u32_ofs));
2060
2061 /* Validate mask, count the mask length. */
2062 if (mask_tz) {
2063 if (mask) {
2064 return 0; /* No bits allowed after mask ended. */
2065 }
2066 } else {
2067 if (~mask & (~mask + 1)) {
2068 return 0; /* Mask not contiguous. */
2069 }
2070 mask_tz = ctz32(mask);
2071 n_bits += 32 - mask_tz;
2072 }
2073 }
2074
2075 return n_bits;
2076 }
2077
2078 /*
2079 * This is called only when mask prefix is known to be CIDR and non-zero.
2080 * Relies on the fact that the flow and mask have the same map, and since
2081 * the mask is CIDR, the storage for the flow field exists even if it
2082 * happened to be zeros.
2083 */
2084 static const ovs_be32 *
2085 minimatch_get_prefix(const struct minimatch *match, const struct mf_field *mf)
2086 {
2087 return miniflow_get_be32_values(&match->flow) +
2088 count_1bits(match->flow.map & ((UINT64_C(1) << mf->flow_be32ofs) - 1));
2089 }
2090
2091 /* Insert rule in to the prefix tree.
2092 * 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2093 * in 'rule'. */
2094 static void
2095 trie_insert(struct cls_trie *trie, const struct cls_rule *rule, int mlen)
2096 {
2097 trie_insert_prefix(&trie->root,
2098 minimatch_get_prefix(&rule->match, trie->field), mlen);
2099 }
2100
2101 static void
2102 trie_insert_prefix(rcu_trie_ptr *edge, const ovs_be32 *prefix, int mlen)
2103 {
2104 struct trie_node *node;
2105 int ofs = 0;
2106
2107 /* Walk the tree. */
2108 for (; (node = ovsrcu_get_protected(struct trie_node *, edge));
2109 edge = trie_next_edge(node, prefix, ofs)) {
2110 unsigned int eqbits = trie_prefix_equal_bits(node, prefix, ofs, mlen);
2111 ofs += eqbits;
2112 if (eqbits < node->n_bits) {
2113 /* Mismatch, new node needs to be inserted above. */
2114 int old_branch = get_bit_at(node->prefix, eqbits);
2115 struct trie_node *new_parent;
2116
2117 new_parent = trie_branch_create(prefix, ofs - eqbits, eqbits,
2118 ofs == mlen ? 1 : 0);
2119 /* Copy the node to modify it. */
2120 node = trie_node_rcu_realloc(node);
2121 /* Adjust the new node for its new position in the tree. */
2122 node->prefix <<= eqbits;
2123 node->n_bits -= eqbits;
2124 ovsrcu_set_hidden(&new_parent->edges[old_branch], node);
2125
2126 /* Check if need a new branch for the new rule. */
2127 if (ofs < mlen) {
2128 ovsrcu_set_hidden(&new_parent->edges[!old_branch],
2129 trie_branch_create(prefix, ofs, mlen - ofs,
2130 1));
2131 }
2132 ovsrcu_set(edge, new_parent); /* Publish changes. */
2133 return;
2134 }
2135 /* Full match so far. */
2136
2137 if (ofs == mlen) {
2138 /* Full match at the current node, rule needs to be added here. */
2139 node->n_rules++;
2140 return;
2141 }
2142 }
2143 /* Must insert a new tree branch for the new rule. */
2144 ovsrcu_set(edge, trie_branch_create(prefix, ofs, mlen - ofs, 1));
2145 }
2146
2147 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2148 * in 'rule'. */
2149 static void
2150 trie_remove(struct cls_trie *trie, const struct cls_rule *rule, int mlen)
2151 {
2152 trie_remove_prefix(&trie->root,
2153 minimatch_get_prefix(&rule->match, trie->field), mlen);
2154 }
2155
2156 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2157 * in 'rule'. */
2158 static void
2159 trie_remove_prefix(rcu_trie_ptr *root, const ovs_be32 *prefix, int mlen)
2160 {
2161 struct trie_node *node;
2162 rcu_trie_ptr *edges[sizeof(union mf_value) * 8];
2163 int depth = 0, ofs = 0;
2164
2165 /* Walk the tree. */
2166 for (edges[0] = root;
2167 (node = ovsrcu_get_protected(struct trie_node *, edges[depth]));
2168 edges[++depth] = trie_next_edge(node, prefix, ofs)) {
2169 unsigned int eqbits = trie_prefix_equal_bits(node, prefix, ofs, mlen);
2170
2171 if (eqbits < node->n_bits) {
2172 /* Mismatch, nothing to be removed. This should never happen, as
2173 * only rules in the classifier are ever removed. */
2174 break; /* Log a warning. */
2175 }
2176 /* Full match so far. */
2177 ofs += eqbits;
2178
2179 if (ofs == mlen) {
2180 /* Full prefix match at the current node, remove rule here. */
2181 if (!node->n_rules) {
2182 break; /* Log a warning. */
2183 }
2184 node->n_rules--;
2185
2186 /* Check if can prune the tree. */
2187 while (!node->n_rules) {
2188 struct trie_node *next,
2189 *edge0 = ovsrcu_get_protected(struct trie_node *,
2190 &node->edges[0]),
2191 *edge1 = ovsrcu_get_protected(struct trie_node *,
2192 &node->edges[1]);
2193
2194 if (edge0 && edge1) {
2195 break; /* A branching point, cannot prune. */
2196 }
2197
2198 /* Else have at most one child node, remove this node. */
2199 next = edge0 ? edge0 : edge1;
2200
2201 if (next) {
2202 if (node->n_bits + next->n_bits > TRIE_PREFIX_BITS) {
2203 break; /* Cannot combine. */
2204 }
2205 next = trie_node_rcu_realloc(next); /* Modify. */
2206
2207 /* Combine node with next. */
2208 next->prefix = node->prefix | next->prefix >> node->n_bits;
2209 next->n_bits += node->n_bits;
2210 }
2211 /* Update the parent's edge. */
2212 ovsrcu_set(edges[depth], next); /* Publish changes. */
2213 trie_node_destroy(node);
2214
2215 if (next || !depth) {
2216 /* Branch not pruned or at root, nothing more to do. */
2217 break;
2218 }
2219 node = ovsrcu_get_protected(struct trie_node *,
2220 edges[--depth]);
2221 }
2222 return;
2223 }
2224 }
2225 /* Cannot go deeper. This should never happen, since only rules
2226 * that actually exist in the classifier are ever removed. */
2227 VLOG_WARN("Trying to remove non-existing rule from a prefix trie.");
2228 }