xref: /openbmc/linux/net/openvswitch/flow_table.c (revision 76ce0265)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2007-2014 Nicira, Inc.
4  */
5 
6 #include "flow.h"
7 #include "datapath.h"
8 #include "flow_netlink.h"
9 #include <linux/uaccess.h>
10 #include <linux/netdevice.h>
11 #include <linux/etherdevice.h>
12 #include <linux/if_ether.h>
13 #include <linux/if_vlan.h>
14 #include <net/llc_pdu.h>
15 #include <linux/kernel.h>
16 #include <linux/jhash.h>
17 #include <linux/jiffies.h>
18 #include <linux/llc.h>
19 #include <linux/module.h>
20 #include <linux/in.h>
21 #include <linux/rcupdate.h>
22 #include <linux/cpumask.h>
23 #include <linux/if_arp.h>
24 #include <linux/ip.h>
25 #include <linux/ipv6.h>
26 #include <linux/sctp.h>
27 #include <linux/tcp.h>
28 #include <linux/udp.h>
29 #include <linux/icmp.h>
30 #include <linux/icmpv6.h>
31 #include <linux/rculist.h>
32 #include <net/ip.h>
33 #include <net/ipv6.h>
34 #include <net/ndisc.h>
35 
36 #define TBL_MIN_BUCKETS		1024
37 #define MASK_ARRAY_SIZE_MIN	16
38 #define REHASH_INTERVAL		(10 * 60 * HZ)
39 
40 #define MC_HASH_SHIFT		8
41 #define MC_HASH_ENTRIES		(1u << MC_HASH_SHIFT)
42 #define MC_HASH_SEGS		((sizeof(uint32_t) * 8) / MC_HASH_SHIFT)
43 
44 static struct kmem_cache *flow_cache;
45 struct kmem_cache *flow_stats_cache __read_mostly;
46 
47 static u16 range_n_bytes(const struct sw_flow_key_range *range)
48 {
49 	return range->end - range->start;
50 }
51 
52 void ovs_flow_mask_key(struct sw_flow_key *dst, const struct sw_flow_key *src,
53 		       bool full, const struct sw_flow_mask *mask)
54 {
55 	int start = full ? 0 : mask->range.start;
56 	int len = full ? sizeof *dst : range_n_bytes(&mask->range);
57 	const long *m = (const long *)((const u8 *)&mask->key + start);
58 	const long *s = (const long *)((const u8 *)src + start);
59 	long *d = (long *)((u8 *)dst + start);
60 	int i;
61 
62 	/* If 'full' is true then all of 'dst' is fully initialized. Otherwise,
63 	 * if 'full' is false the memory outside of the 'mask->range' is left
64 	 * uninitialized. This can be used as an optimization when further
65 	 * operations on 'dst' only use contents within 'mask->range'.
66 	 */
67 	for (i = 0; i < len; i += sizeof(long))
68 		*d++ = *s++ & *m++;
69 }
70 
71 struct sw_flow *ovs_flow_alloc(void)
72 {
73 	struct sw_flow *flow;
74 	struct sw_flow_stats *stats;
75 
76 	flow = kmem_cache_zalloc(flow_cache, GFP_KERNEL);
77 	if (!flow)
78 		return ERR_PTR(-ENOMEM);
79 
80 	flow->stats_last_writer = -1;
81 
82 	/* Initialize the default stat node. */
83 	stats = kmem_cache_alloc_node(flow_stats_cache,
84 				      GFP_KERNEL | __GFP_ZERO,
85 				      node_online(0) ? 0 : NUMA_NO_NODE);
86 	if (!stats)
87 		goto err;
88 
89 	spin_lock_init(&stats->lock);
90 
91 	RCU_INIT_POINTER(flow->stats[0], stats);
92 
93 	cpumask_set_cpu(0, &flow->cpu_used_mask);
94 
95 	return flow;
96 err:
97 	kmem_cache_free(flow_cache, flow);
98 	return ERR_PTR(-ENOMEM);
99 }
100 
101 int ovs_flow_tbl_count(const struct flow_table *table)
102 {
103 	return table->count;
104 }
105 
106 static void flow_free(struct sw_flow *flow)
107 {
108 	int cpu;
109 
110 	if (ovs_identifier_is_key(&flow->id))
111 		kfree(flow->id.unmasked_key);
112 	if (flow->sf_acts)
113 		ovs_nla_free_flow_actions((struct sw_flow_actions __force *)flow->sf_acts);
114 	/* We open code this to make sure cpu 0 is always considered */
115 	for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, &flow->cpu_used_mask))
116 		if (flow->stats[cpu])
117 			kmem_cache_free(flow_stats_cache,
118 					(struct sw_flow_stats __force *)flow->stats[cpu]);
119 	kmem_cache_free(flow_cache, flow);
120 }
121 
122 static void rcu_free_flow_callback(struct rcu_head *rcu)
123 {
124 	struct sw_flow *flow = container_of(rcu, struct sw_flow, rcu);
125 
126 	flow_free(flow);
127 }
128 
129 void ovs_flow_free(struct sw_flow *flow, bool deferred)
130 {
131 	if (!flow)
132 		return;
133 
134 	if (deferred)
135 		call_rcu(&flow->rcu, rcu_free_flow_callback);
136 	else
137 		flow_free(flow);
138 }
139 
140 static void __table_instance_destroy(struct table_instance *ti)
141 {
142 	kvfree(ti->buckets);
143 	kfree(ti);
144 }
145 
146 static struct table_instance *table_instance_alloc(int new_size)
147 {
148 	struct table_instance *ti = kmalloc(sizeof(*ti), GFP_KERNEL);
149 	int i;
150 
151 	if (!ti)
152 		return NULL;
153 
154 	ti->buckets = kvmalloc_array(new_size, sizeof(struct hlist_head),
155 				     GFP_KERNEL);
156 	if (!ti->buckets) {
157 		kfree(ti);
158 		return NULL;
159 	}
160 
161 	for (i = 0; i < new_size; i++)
162 		INIT_HLIST_HEAD(&ti->buckets[i]);
163 
164 	ti->n_buckets = new_size;
165 	ti->node_ver = 0;
166 	ti->keep_flows = false;
167 	get_random_bytes(&ti->hash_seed, sizeof(u32));
168 
169 	return ti;
170 }
171 
172 static struct mask_array *tbl_mask_array_alloc(int size)
173 {
174 	struct mask_array *new;
175 
176 	size = max(MASK_ARRAY_SIZE_MIN, size);
177 	new = kzalloc(sizeof(struct mask_array) +
178 		      sizeof(struct sw_flow_mask *) * size, GFP_KERNEL);
179 	if (!new)
180 		return NULL;
181 
182 	new->count = 0;
183 	new->max = size;
184 
185 	return new;
186 }
187 
188 static int tbl_mask_array_realloc(struct flow_table *tbl, int size)
189 {
190 	struct mask_array *old;
191 	struct mask_array *new;
192 
193 	new = tbl_mask_array_alloc(size);
194 	if (!new)
195 		return -ENOMEM;
196 
197 	old = ovsl_dereference(tbl->mask_array);
198 	if (old) {
199 		int i;
200 
201 		for (i = 0; i < old->max; i++) {
202 			if (ovsl_dereference(old->masks[i]))
203 				new->masks[new->count++] = old->masks[i];
204 		}
205 	}
206 
207 	rcu_assign_pointer(tbl->mask_array, new);
208 	kfree_rcu(old, rcu);
209 
210 	return 0;
211 }
212 
213 static int tbl_mask_array_add_mask(struct flow_table *tbl,
214 				   struct sw_flow_mask *new)
215 {
216 	struct mask_array *ma = ovsl_dereference(tbl->mask_array);
217 	int err, ma_count = READ_ONCE(ma->count);
218 
219 	if (ma_count >= ma->max) {
220 		err = tbl_mask_array_realloc(tbl, ma->max +
221 					      MASK_ARRAY_SIZE_MIN);
222 		if (err)
223 			return err;
224 
225 		ma = ovsl_dereference(tbl->mask_array);
226 	}
227 
228 	BUG_ON(ovsl_dereference(ma->masks[ma_count]));
229 
230 	rcu_assign_pointer(ma->masks[ma_count], new);
231 	WRITE_ONCE(ma->count, ma_count +1);
232 
233 	return 0;
234 }
235 
236 static void tbl_mask_array_del_mask(struct flow_table *tbl,
237 				    struct sw_flow_mask *mask)
238 {
239 	struct mask_array *ma = ovsl_dereference(tbl->mask_array);
240 	int i, ma_count = READ_ONCE(ma->count);
241 
242 	/* Remove the deleted mask pointers from the array */
243 	for (i = 0; i < ma_count; i++) {
244 		if (mask == ovsl_dereference(ma->masks[i]))
245 			goto found;
246 	}
247 
248 	BUG();
249 	return;
250 
251 found:
252 	WRITE_ONCE(ma->count, ma_count -1);
253 
254 	rcu_assign_pointer(ma->masks[i], ma->masks[ma_count -1]);
255 	RCU_INIT_POINTER(ma->masks[ma_count -1], NULL);
256 
257 	kfree_rcu(mask, rcu);
258 
259 	/* Shrink the mask array if necessary. */
260 	if (ma->max >= (MASK_ARRAY_SIZE_MIN * 2) &&
261 	    ma_count <= (ma->max / 3))
262 		tbl_mask_array_realloc(tbl, ma->max / 2);
263 }
264 
265 /* Remove 'mask' from the mask list, if it is not needed any more. */
266 static void flow_mask_remove(struct flow_table *tbl, struct sw_flow_mask *mask)
267 {
268 	if (mask) {
269 		/* ovs-lock is required to protect mask-refcount and
270 		 * mask list.
271 		 */
272 		ASSERT_OVSL();
273 		BUG_ON(!mask->ref_count);
274 		mask->ref_count--;
275 
276 		if (!mask->ref_count)
277 			tbl_mask_array_del_mask(tbl, mask);
278 	}
279 }
280 
281 int ovs_flow_tbl_init(struct flow_table *table)
282 {
283 	struct table_instance *ti, *ufid_ti;
284 	struct mask_array *ma;
285 
286 	table->mask_cache = __alloc_percpu(sizeof(struct mask_cache_entry) *
287 					   MC_HASH_ENTRIES,
288 					   __alignof__(struct mask_cache_entry));
289 	if (!table->mask_cache)
290 		return -ENOMEM;
291 
292 	ma = tbl_mask_array_alloc(MASK_ARRAY_SIZE_MIN);
293 	if (!ma)
294 		goto free_mask_cache;
295 
296 	ti = table_instance_alloc(TBL_MIN_BUCKETS);
297 	if (!ti)
298 		goto free_mask_array;
299 
300 	ufid_ti = table_instance_alloc(TBL_MIN_BUCKETS);
301 	if (!ufid_ti)
302 		goto free_ti;
303 
304 	rcu_assign_pointer(table->ti, ti);
305 	rcu_assign_pointer(table->ufid_ti, ufid_ti);
306 	rcu_assign_pointer(table->mask_array, ma);
307 	table->last_rehash = jiffies;
308 	table->count = 0;
309 	table->ufid_count = 0;
310 	return 0;
311 
312 free_ti:
313 	__table_instance_destroy(ti);
314 free_mask_array:
315 	kfree(ma);
316 free_mask_cache:
317 	free_percpu(table->mask_cache);
318 	return -ENOMEM;
319 }
320 
321 static void flow_tbl_destroy_rcu_cb(struct rcu_head *rcu)
322 {
323 	struct table_instance *ti = container_of(rcu, struct table_instance, rcu);
324 
325 	__table_instance_destroy(ti);
326 }
327 
328 static void table_instance_flow_free(struct flow_table *table,
329 				  struct table_instance *ti,
330 				  struct table_instance *ufid_ti,
331 				  struct sw_flow *flow,
332 				  bool count)
333 {
334 	hlist_del_rcu(&flow->flow_table.node[ti->node_ver]);
335 	if (count)
336 		table->count--;
337 
338 	if (ovs_identifier_is_ufid(&flow->id)) {
339 		hlist_del_rcu(&flow->ufid_table.node[ufid_ti->node_ver]);
340 
341 		if (count)
342 			table->ufid_count--;
343 	}
344 
345 	flow_mask_remove(table, flow->mask);
346 }
347 
348 static void table_instance_destroy(struct flow_table *table,
349 				   struct table_instance *ti,
350 				   struct table_instance *ufid_ti,
351 				   bool deferred)
352 {
353 	int i;
354 
355 	if (!ti)
356 		return;
357 
358 	BUG_ON(!ufid_ti);
359 	if (ti->keep_flows)
360 		goto skip_flows;
361 
362 	for (i = 0; i < ti->n_buckets; i++) {
363 		struct sw_flow *flow;
364 		struct hlist_head *head = &ti->buckets[i];
365 		struct hlist_node *n;
366 
367 		hlist_for_each_entry_safe(flow, n, head,
368 					  flow_table.node[ti->node_ver]) {
369 
370 			table_instance_flow_free(table, ti, ufid_ti,
371 						 flow, false);
372 			ovs_flow_free(flow, deferred);
373 		}
374 	}
375 
376 skip_flows:
377 	if (deferred) {
378 		call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
379 		call_rcu(&ufid_ti->rcu, flow_tbl_destroy_rcu_cb);
380 	} else {
381 		__table_instance_destroy(ti);
382 		__table_instance_destroy(ufid_ti);
383 	}
384 }
385 
386 /* No need for locking this function is called from RCU callback or
387  * error path.
388  */
389 void ovs_flow_tbl_destroy(struct flow_table *table)
390 {
391 	struct table_instance *ti = rcu_dereference_raw(table->ti);
392 	struct table_instance *ufid_ti = rcu_dereference_raw(table->ufid_ti);
393 
394 	free_percpu(table->mask_cache);
395 	kfree_rcu(rcu_dereference_raw(table->mask_array), rcu);
396 	table_instance_destroy(table, ti, ufid_ti, false);
397 }
398 
399 struct sw_flow *ovs_flow_tbl_dump_next(struct table_instance *ti,
400 				       u32 *bucket, u32 *last)
401 {
402 	struct sw_flow *flow;
403 	struct hlist_head *head;
404 	int ver;
405 	int i;
406 
407 	ver = ti->node_ver;
408 	while (*bucket < ti->n_buckets) {
409 		i = 0;
410 		head = &ti->buckets[*bucket];
411 		hlist_for_each_entry_rcu(flow, head, flow_table.node[ver]) {
412 			if (i < *last) {
413 				i++;
414 				continue;
415 			}
416 			*last = i + 1;
417 			return flow;
418 		}
419 		(*bucket)++;
420 		*last = 0;
421 	}
422 
423 	return NULL;
424 }
425 
426 static struct hlist_head *find_bucket(struct table_instance *ti, u32 hash)
427 {
428 	hash = jhash_1word(hash, ti->hash_seed);
429 	return &ti->buckets[hash & (ti->n_buckets - 1)];
430 }
431 
432 static void table_instance_insert(struct table_instance *ti,
433 				  struct sw_flow *flow)
434 {
435 	struct hlist_head *head;
436 
437 	head = find_bucket(ti, flow->flow_table.hash);
438 	hlist_add_head_rcu(&flow->flow_table.node[ti->node_ver], head);
439 }
440 
441 static void ufid_table_instance_insert(struct table_instance *ti,
442 				       struct sw_flow *flow)
443 {
444 	struct hlist_head *head;
445 
446 	head = find_bucket(ti, flow->ufid_table.hash);
447 	hlist_add_head_rcu(&flow->ufid_table.node[ti->node_ver], head);
448 }
449 
450 static void flow_table_copy_flows(struct table_instance *old,
451 				  struct table_instance *new, bool ufid)
452 {
453 	int old_ver;
454 	int i;
455 
456 	old_ver = old->node_ver;
457 	new->node_ver = !old_ver;
458 
459 	/* Insert in new table. */
460 	for (i = 0; i < old->n_buckets; i++) {
461 		struct sw_flow *flow;
462 		struct hlist_head *head = &old->buckets[i];
463 
464 		if (ufid)
465 			hlist_for_each_entry(flow, head,
466 					     ufid_table.node[old_ver])
467 				ufid_table_instance_insert(new, flow);
468 		else
469 			hlist_for_each_entry(flow, head,
470 					     flow_table.node[old_ver])
471 				table_instance_insert(new, flow);
472 	}
473 
474 	old->keep_flows = true;
475 }
476 
477 static struct table_instance *table_instance_rehash(struct table_instance *ti,
478 						    int n_buckets, bool ufid)
479 {
480 	struct table_instance *new_ti;
481 
482 	new_ti = table_instance_alloc(n_buckets);
483 	if (!new_ti)
484 		return NULL;
485 
486 	flow_table_copy_flows(ti, new_ti, ufid);
487 
488 	return new_ti;
489 }
490 
491 int ovs_flow_tbl_flush(struct flow_table *flow_table)
492 {
493 	struct table_instance *old_ti, *new_ti;
494 	struct table_instance *old_ufid_ti, *new_ufid_ti;
495 
496 	new_ti = table_instance_alloc(TBL_MIN_BUCKETS);
497 	if (!new_ti)
498 		return -ENOMEM;
499 	new_ufid_ti = table_instance_alloc(TBL_MIN_BUCKETS);
500 	if (!new_ufid_ti)
501 		goto err_free_ti;
502 
503 	old_ti = ovsl_dereference(flow_table->ti);
504 	old_ufid_ti = ovsl_dereference(flow_table->ufid_ti);
505 
506 	rcu_assign_pointer(flow_table->ti, new_ti);
507 	rcu_assign_pointer(flow_table->ufid_ti, new_ufid_ti);
508 	flow_table->last_rehash = jiffies;
509 	flow_table->count = 0;
510 	flow_table->ufid_count = 0;
511 
512 	table_instance_destroy(flow_table, old_ti, old_ufid_ti, true);
513 	return 0;
514 
515 err_free_ti:
516 	__table_instance_destroy(new_ti);
517 	return -ENOMEM;
518 }
519 
520 static u32 flow_hash(const struct sw_flow_key *key,
521 		     const struct sw_flow_key_range *range)
522 {
523 	const u32 *hash_key = (const u32 *)((const u8 *)key + range->start);
524 
525 	/* Make sure number of hash bytes are multiple of u32. */
526 	int hash_u32s = range_n_bytes(range) >> 2;
527 
528 	return jhash2(hash_key, hash_u32s, 0);
529 }
530 
531 static int flow_key_start(const struct sw_flow_key *key)
532 {
533 	if (key->tun_proto)
534 		return 0;
535 	else
536 		return rounddown(offsetof(struct sw_flow_key, phy),
537 					  sizeof(long));
538 }
539 
540 static bool cmp_key(const struct sw_flow_key *key1,
541 		    const struct sw_flow_key *key2,
542 		    int key_start, int key_end)
543 {
544 	const long *cp1 = (const long *)((const u8 *)key1 + key_start);
545 	const long *cp2 = (const long *)((const u8 *)key2 + key_start);
546 	long diffs = 0;
547 	int i;
548 
549 	for (i = key_start; i < key_end;  i += sizeof(long))
550 		diffs |= *cp1++ ^ *cp2++;
551 
552 	return diffs == 0;
553 }
554 
555 static bool flow_cmp_masked_key(const struct sw_flow *flow,
556 				const struct sw_flow_key *key,
557 				const struct sw_flow_key_range *range)
558 {
559 	return cmp_key(&flow->key, key, range->start, range->end);
560 }
561 
562 static bool ovs_flow_cmp_unmasked_key(const struct sw_flow *flow,
563 				      const struct sw_flow_match *match)
564 {
565 	struct sw_flow_key *key = match->key;
566 	int key_start = flow_key_start(key);
567 	int key_end = match->range.end;
568 
569 	BUG_ON(ovs_identifier_is_ufid(&flow->id));
570 	return cmp_key(flow->id.unmasked_key, key, key_start, key_end);
571 }
572 
573 static struct sw_flow *masked_flow_lookup(struct table_instance *ti,
574 					  const struct sw_flow_key *unmasked,
575 					  const struct sw_flow_mask *mask,
576 					  u32 *n_mask_hit)
577 {
578 	struct sw_flow *flow;
579 	struct hlist_head *head;
580 	u32 hash;
581 	struct sw_flow_key masked_key;
582 
583 	ovs_flow_mask_key(&masked_key, unmasked, false, mask);
584 	hash = flow_hash(&masked_key, &mask->range);
585 	head = find_bucket(ti, hash);
586 	(*n_mask_hit)++;
587 
588 	hlist_for_each_entry_rcu(flow, head, flow_table.node[ti->node_ver],
589 				lockdep_ovsl_is_held()) {
590 		if (flow->mask == mask && flow->flow_table.hash == hash &&
591 		    flow_cmp_masked_key(flow, &masked_key, &mask->range))
592 			return flow;
593 	}
594 	return NULL;
595 }
596 
597 /* Flow lookup does full lookup on flow table. It starts with
598  * mask from index passed in *index.
599  */
600 static struct sw_flow *flow_lookup(struct flow_table *tbl,
601 				   struct table_instance *ti,
602 				   struct mask_array *ma,
603 				   const struct sw_flow_key *key,
604 				   u32 *n_mask_hit,
605 				   u32 *index)
606 {
607 	struct sw_flow *flow;
608 	struct sw_flow_mask *mask;
609 	int i;
610 
611 	if (likely(*index < ma->max)) {
612 		mask = rcu_dereference_ovsl(ma->masks[*index]);
613 		if (mask) {
614 			flow = masked_flow_lookup(ti, key, mask, n_mask_hit);
615 			if (flow)
616 				return flow;
617 		}
618 	}
619 
620 	for (i = 0; i < ma->max; i++)  {
621 
622 		if (i == *index)
623 			continue;
624 
625 		mask = rcu_dereference_ovsl(ma->masks[i]);
626 		if (unlikely(!mask))
627 			break;
628 
629 		flow = masked_flow_lookup(ti, key, mask, n_mask_hit);
630 		if (flow) { /* Found */
631 			*index = i;
632 			return flow;
633 		}
634 	}
635 
636 	return NULL;
637 }
638 
639 /*
640  * mask_cache maps flow to probable mask. This cache is not tightly
641  * coupled cache, It means updates to  mask list can result in inconsistent
642  * cache entry in mask cache.
643  * This is per cpu cache and is divided in MC_HASH_SEGS segments.
644  * In case of a hash collision the entry is hashed in next segment.
645  * */
646 struct sw_flow *ovs_flow_tbl_lookup_stats(struct flow_table *tbl,
647 					  const struct sw_flow_key *key,
648 					  u32 skb_hash,
649 					  u32 *n_mask_hit)
650 {
651 	struct mask_array *ma = rcu_dereference(tbl->mask_array);
652 	struct table_instance *ti = rcu_dereference(tbl->ti);
653 	struct mask_cache_entry *entries, *ce;
654 	struct sw_flow *flow;
655 	u32 hash;
656 	int seg;
657 
658 	*n_mask_hit = 0;
659 	if (unlikely(!skb_hash)) {
660 		u32 mask_index = 0;
661 
662 		return flow_lookup(tbl, ti, ma, key, n_mask_hit, &mask_index);
663 	}
664 
665 	/* Pre and post recirulation flows usually have the same skb_hash
666 	 * value. To avoid hash collisions, rehash the 'skb_hash' with
667 	 * 'recirc_id'.  */
668 	if (key->recirc_id)
669 		skb_hash = jhash_1word(skb_hash, key->recirc_id);
670 
671 	ce = NULL;
672 	hash = skb_hash;
673 	entries = this_cpu_ptr(tbl->mask_cache);
674 
675 	/* Find the cache entry 'ce' to operate on. */
676 	for (seg = 0; seg < MC_HASH_SEGS; seg++) {
677 		int index = hash & (MC_HASH_ENTRIES - 1);
678 		struct mask_cache_entry *e;
679 
680 		e = &entries[index];
681 		if (e->skb_hash == skb_hash) {
682 			flow = flow_lookup(tbl, ti, ma, key, n_mask_hit,
683 					   &e->mask_index);
684 			if (!flow)
685 				e->skb_hash = 0;
686 			return flow;
687 		}
688 
689 		if (!ce || e->skb_hash < ce->skb_hash)
690 			ce = e;  /* A better replacement cache candidate. */
691 
692 		hash >>= MC_HASH_SHIFT;
693 	}
694 
695 	/* Cache miss, do full lookup. */
696 	flow = flow_lookup(tbl, ti, ma, key, n_mask_hit, &ce->mask_index);
697 	if (flow)
698 		ce->skb_hash = skb_hash;
699 
700 	return flow;
701 }
702 
703 struct sw_flow *ovs_flow_tbl_lookup(struct flow_table *tbl,
704 				    const struct sw_flow_key *key)
705 {
706 	struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
707 	struct mask_array *ma = rcu_dereference_ovsl(tbl->mask_array);
708 	u32 __always_unused n_mask_hit;
709 	u32 index = 0;
710 
711 	return flow_lookup(tbl, ti, ma, key, &n_mask_hit, &index);
712 }
713 
714 struct sw_flow *ovs_flow_tbl_lookup_exact(struct flow_table *tbl,
715 					  const struct sw_flow_match *match)
716 {
717 	struct mask_array *ma = ovsl_dereference(tbl->mask_array);
718 	int i;
719 
720 	/* Always called under ovs-mutex. */
721 	for (i = 0; i < ma->max; i++) {
722 		struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
723 		u32 __always_unused n_mask_hit;
724 		struct sw_flow_mask *mask;
725 		struct sw_flow *flow;
726 
727 		mask = ovsl_dereference(ma->masks[i]);
728 		if (!mask)
729 			continue;
730 
731 		flow = masked_flow_lookup(ti, match->key, mask, &n_mask_hit);
732 		if (flow && ovs_identifier_is_key(&flow->id) &&
733 		    ovs_flow_cmp_unmasked_key(flow, match)) {
734 			return flow;
735 		}
736 	}
737 
738 	return NULL;
739 }
740 
741 static u32 ufid_hash(const struct sw_flow_id *sfid)
742 {
743 	return jhash(sfid->ufid, sfid->ufid_len, 0);
744 }
745 
746 static bool ovs_flow_cmp_ufid(const struct sw_flow *flow,
747 			      const struct sw_flow_id *sfid)
748 {
749 	if (flow->id.ufid_len != sfid->ufid_len)
750 		return false;
751 
752 	return !memcmp(flow->id.ufid, sfid->ufid, sfid->ufid_len);
753 }
754 
755 bool ovs_flow_cmp(const struct sw_flow *flow, const struct sw_flow_match *match)
756 {
757 	if (ovs_identifier_is_ufid(&flow->id))
758 		return flow_cmp_masked_key(flow, match->key, &match->range);
759 
760 	return ovs_flow_cmp_unmasked_key(flow, match);
761 }
762 
763 struct sw_flow *ovs_flow_tbl_lookup_ufid(struct flow_table *tbl,
764 					 const struct sw_flow_id *ufid)
765 {
766 	struct table_instance *ti = rcu_dereference_ovsl(tbl->ufid_ti);
767 	struct sw_flow *flow;
768 	struct hlist_head *head;
769 	u32 hash;
770 
771 	hash = ufid_hash(ufid);
772 	head = find_bucket(ti, hash);
773 	hlist_for_each_entry_rcu(flow, head, ufid_table.node[ti->node_ver],
774 				lockdep_ovsl_is_held()) {
775 		if (flow->ufid_table.hash == hash &&
776 		    ovs_flow_cmp_ufid(flow, ufid))
777 			return flow;
778 	}
779 	return NULL;
780 }
781 
782 int ovs_flow_tbl_num_masks(const struct flow_table *table)
783 {
784 	struct mask_array *ma = rcu_dereference_ovsl(table->mask_array);
785 	return READ_ONCE(ma->count);
786 }
787 
788 static struct table_instance *table_instance_expand(struct table_instance *ti,
789 						    bool ufid)
790 {
791 	return table_instance_rehash(ti, ti->n_buckets * 2, ufid);
792 }
793 
794 /* Must be called with OVS mutex held. */
795 void ovs_flow_tbl_remove(struct flow_table *table, struct sw_flow *flow)
796 {
797 	struct table_instance *ti = ovsl_dereference(table->ti);
798 	struct table_instance *ufid_ti = ovsl_dereference(table->ufid_ti);
799 
800 	BUG_ON(table->count == 0);
801 	table_instance_flow_free(table, ti, ufid_ti, flow, true);
802 }
803 
804 static struct sw_flow_mask *mask_alloc(void)
805 {
806 	struct sw_flow_mask *mask;
807 
808 	mask = kmalloc(sizeof(*mask), GFP_KERNEL);
809 	if (mask)
810 		mask->ref_count = 1;
811 
812 	return mask;
813 }
814 
815 static bool mask_equal(const struct sw_flow_mask *a,
816 		       const struct sw_flow_mask *b)
817 {
818 	const u8 *a_ = (const u8 *)&a->key + a->range.start;
819 	const u8 *b_ = (const u8 *)&b->key + b->range.start;
820 
821 	return  (a->range.end == b->range.end)
822 		&& (a->range.start == b->range.start)
823 		&& (memcmp(a_, b_, range_n_bytes(&a->range)) == 0);
824 }
825 
826 static struct sw_flow_mask *flow_mask_find(const struct flow_table *tbl,
827 					   const struct sw_flow_mask *mask)
828 {
829 	struct mask_array *ma;
830 	int i;
831 
832 	ma = ovsl_dereference(tbl->mask_array);
833 	for (i = 0; i < ma->max; i++) {
834 		struct sw_flow_mask *t;
835 		t = ovsl_dereference(ma->masks[i]);
836 
837 		if (t && mask_equal(mask, t))
838 			return t;
839 	}
840 
841 	return NULL;
842 }
843 
844 /* Add 'mask' into the mask list, if it is not already there. */
845 static int flow_mask_insert(struct flow_table *tbl, struct sw_flow *flow,
846 			    const struct sw_flow_mask *new)
847 {
848 	struct sw_flow_mask *mask;
849 
850 	mask = flow_mask_find(tbl, new);
851 	if (!mask) {
852 		/* Allocate a new mask if none exsits. */
853 		mask = mask_alloc();
854 		if (!mask)
855 			return -ENOMEM;
856 		mask->key = new->key;
857 		mask->range = new->range;
858 
859 		/* Add mask to mask-list. */
860 		if (tbl_mask_array_add_mask(tbl, mask)) {
861 			kfree(mask);
862 			return -ENOMEM;
863 		}
864 	} else {
865 		BUG_ON(!mask->ref_count);
866 		mask->ref_count++;
867 	}
868 
869 	flow->mask = mask;
870 	return 0;
871 }
872 
873 /* Must be called with OVS mutex held. */
874 static void flow_key_insert(struct flow_table *table, struct sw_flow *flow)
875 {
876 	struct table_instance *new_ti = NULL;
877 	struct table_instance *ti;
878 
879 	flow->flow_table.hash = flow_hash(&flow->key, &flow->mask->range);
880 	ti = ovsl_dereference(table->ti);
881 	table_instance_insert(ti, flow);
882 	table->count++;
883 
884 	/* Expand table, if necessary, to make room. */
885 	if (table->count > ti->n_buckets)
886 		new_ti = table_instance_expand(ti, false);
887 	else if (time_after(jiffies, table->last_rehash + REHASH_INTERVAL))
888 		new_ti = table_instance_rehash(ti, ti->n_buckets, false);
889 
890 	if (new_ti) {
891 		rcu_assign_pointer(table->ti, new_ti);
892 		call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
893 		table->last_rehash = jiffies;
894 	}
895 }
896 
897 /* Must be called with OVS mutex held. */
898 static void flow_ufid_insert(struct flow_table *table, struct sw_flow *flow)
899 {
900 	struct table_instance *ti;
901 
902 	flow->ufid_table.hash = ufid_hash(&flow->id);
903 	ti = ovsl_dereference(table->ufid_ti);
904 	ufid_table_instance_insert(ti, flow);
905 	table->ufid_count++;
906 
907 	/* Expand table, if necessary, to make room. */
908 	if (table->ufid_count > ti->n_buckets) {
909 		struct table_instance *new_ti;
910 
911 		new_ti = table_instance_expand(ti, true);
912 		if (new_ti) {
913 			rcu_assign_pointer(table->ufid_ti, new_ti);
914 			call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
915 		}
916 	}
917 }
918 
919 /* Must be called with OVS mutex held. */
920 int ovs_flow_tbl_insert(struct flow_table *table, struct sw_flow *flow,
921 			const struct sw_flow_mask *mask)
922 {
923 	int err;
924 
925 	err = flow_mask_insert(table, flow, mask);
926 	if (err)
927 		return err;
928 	flow_key_insert(table, flow);
929 	if (ovs_identifier_is_ufid(&flow->id))
930 		flow_ufid_insert(table, flow);
931 
932 	return 0;
933 }
934 
935 /* Initializes the flow module.
936  * Returns zero if successful or a negative error code. */
937 int ovs_flow_init(void)
938 {
939 	BUILD_BUG_ON(__alignof__(struct sw_flow_key) % __alignof__(long));
940 	BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
941 
942 	flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
943 				       + (nr_cpu_ids
944 					  * sizeof(struct sw_flow_stats *)),
945 				       0, 0, NULL);
946 	if (flow_cache == NULL)
947 		return -ENOMEM;
948 
949 	flow_stats_cache
950 		= kmem_cache_create("sw_flow_stats", sizeof(struct sw_flow_stats),
951 				    0, SLAB_HWCACHE_ALIGN, NULL);
952 	if (flow_stats_cache == NULL) {
953 		kmem_cache_destroy(flow_cache);
954 		flow_cache = NULL;
955 		return -ENOMEM;
956 	}
957 
958 	return 0;
959 }
960 
961 /* Uninitializes the flow module. */
962 void ovs_flow_exit(void)
963 {
964 	kmem_cache_destroy(flow_stats_cache);
965 	kmem_cache_destroy(flow_cache);
966 }
967