2 * Copyright (c) 2007-2013 Nicira, Inc.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * You should have received a copy of the GNU General Public License
14 * along with this program; if not, write to the Free Software
15 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
21 #include <linux/uaccess.h>
22 #include <linux/netdevice.h>
23 #include <linux/etherdevice.h>
24 #include <linux/if_ether.h>
25 #include <linux/if_vlan.h>
26 #include <net/llc_pdu.h>
27 #include <linux/kernel.h>
28 #include <linux/hash.h>
29 #include <linux/jiffies.h>
30 #include <linux/llc.h>
31 #include <linux/module.h>
33 #include <linux/rcupdate.h>
34 #include <linux/if_arp.h>
36 #include <linux/ipv6.h>
37 #include <linux/sctp.h>
38 #include <linux/tcp.h>
39 #include <linux/udp.h>
40 #include <linux/icmp.h>
41 #include <linux/icmpv6.h>
42 #include <linux/rculist.h>
45 #include <net/ndisc.h>
49 #define TBL_MIN_BUCKETS 1024
50 #define MASK_ARRAY_SIZE_MIN 16
51 #define REHASH_INTERVAL (10 * 60 * HZ)
53 #define MC_HASH_SHIFT 8
54 #define MC_HASH_ENTRIES (1u << MC_HASH_SHIFT)
55 #define MC_HASH_SEGS ((sizeof(uint32_t) * 8) / MC_HASH_SHIFT)
57 static struct kmem_cache *flow_cache;
58 struct kmem_cache *flow_stats_cache __read_mostly;
60 static u16 range_n_bytes(const struct sw_flow_key_range *range)
62 return range->end - range->start;
65 void ovs_flow_mask_key(struct sw_flow_key *dst, const struct sw_flow_key *src,
66 const struct sw_flow_mask *mask)
68 const long *m = (const long *)((const u8 *)&mask->key +
70 const long *s = (const long *)((const u8 *)src +
72 long *d = (long *)((u8 *)dst + mask->range.start);
75 /* The memory outside of the 'mask->range' are not set since
76 * further operations on 'dst' only uses contents within
79 for (i = 0; i < range_n_bytes(&mask->range); i += sizeof(long))
83 struct sw_flow *ovs_flow_alloc(void)
86 struct flow_stats *stats;
89 flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
91 return ERR_PTR(-ENOMEM);
95 flow->stats_last_writer = NUMA_NO_NODE;
97 /* Initialize the default stat node. */
98 stats = kmem_cache_alloc_node(flow_stats_cache,
99 GFP_KERNEL | __GFP_ZERO, 0);
103 spin_lock_init(&stats->lock);
105 RCU_INIT_POINTER(flow->stats[0], stats);
109 RCU_INIT_POINTER(flow->stats[node], NULL);
113 kmem_cache_free(flow_cache, flow);
114 return ERR_PTR(-ENOMEM);
117 int ovs_flow_tbl_count(struct flow_table *table)
122 static struct flex_array *alloc_buckets(unsigned int n_buckets)
124 struct flex_array *buckets;
127 buckets = flex_array_alloc(sizeof(struct hlist_head),
128 n_buckets, GFP_KERNEL);
132 err = flex_array_prealloc(buckets, 0, n_buckets, GFP_KERNEL);
134 flex_array_free(buckets);
138 for (i = 0; i < n_buckets; i++)
139 INIT_HLIST_HEAD((struct hlist_head *)
140 flex_array_get(buckets, i));
145 static void flow_free(struct sw_flow *flow)
149 kfree((struct sw_flow_actions __force *)flow->sf_acts);
151 if (flow->stats[node])
152 kmem_cache_free(flow_stats_cache,
153 (struct flow_stats __force *)flow->stats[node]);
154 kmem_cache_free(flow_cache, flow);
157 static void rcu_free_flow_callback(struct rcu_head *rcu)
159 struct sw_flow *flow = container_of(rcu, struct sw_flow, rcu);
164 static void rcu_free_sw_flow_mask_cb(struct rcu_head *rcu)
166 struct sw_flow_mask *mask = container_of(rcu, struct sw_flow_mask, rcu);
171 void ovs_flow_free(struct sw_flow *flow, bool deferred)
177 call_rcu(&flow->rcu, rcu_free_flow_callback);
182 static void free_buckets(struct flex_array *buckets)
184 flex_array_free(buckets);
188 static void __table_instance_destroy(struct table_instance *ti)
190 free_buckets(ti->buckets);
194 static struct table_instance *table_instance_alloc(int new_size)
196 struct table_instance *ti = kmalloc(sizeof(*ti), GFP_KERNEL);
201 ti->buckets = alloc_buckets(new_size);
207 ti->n_buckets = new_size;
209 ti->keep_flows = false;
210 get_random_bytes(&ti->hash_seed, sizeof(u32));
215 static void mask_array_rcu_cb(struct rcu_head *rcu)
217 struct mask_array *ma = container_of(rcu, struct mask_array, rcu);
222 static struct mask_array *tbl_mask_array_alloc(int size)
224 struct mask_array *new;
226 new = kzalloc(sizeof(struct mask_array) +
227 sizeof(struct sw_flow_mask *) * size, GFP_KERNEL);
237 static int tbl_mask_array_realloc(struct flow_table *tbl, int size)
239 struct mask_array *old;
240 struct mask_array *new;
242 new = tbl_mask_array_alloc(size);
246 old = ovsl_dereference(tbl->mask_array);
250 for (i = 0; i < old->max; i++)
251 new->masks[i] = old->masks[i];
253 new->count = old->count;
255 rcu_assign_pointer(tbl->mask_array, new);
258 call_rcu(&old->rcu, mask_array_rcu_cb);
263 static void tbl_mask_array_delete_mask(struct mask_array *ma,
264 const struct sw_flow_mask *mask)
268 /* Delete a mask pointer from the valid section.
270 * Also move the last entry in its place, so there is no
271 * whole in the valid section.
273 * Notice the last entry still points to the original mask.
275 * <Note>: there is a small race window that may cause a mask
276 * to be missed in a search. Imaging a core is
277 * walking through the array, passing the index of deleting mask.
278 * But before reaching the last entry, it is overwritten,
279 * by another core that is adding a new mask, now the last entry
280 * will point to the new mask. In this case, the moved up last
281 * entry can be missed by the core walking the mask array.
283 * In case this missed mask would have led to successful
284 * lookup, Hitting the race window could cause a packet to miss
285 * kernel flow cache, and be sent to the user space.
288 for (i = 0; i < ma->count; i++)
289 if (mask == ovsl_dereference(ma->masks[i])) {
290 struct sw_flow_mask *last;
292 last = ovsl_dereference(ma->masks[ma->count - 1]);
293 rcu_assign_pointer(ma->masks[i], last);
298 /* Remove the deleted mask pointers from the invalid section. */
299 for (i = ma->count; i < ma->max; i++)
300 if (mask == ovsl_dereference(ma->masks[i]))
301 RCU_INIT_POINTER(ma->masks[i], NULL);
304 int ovs_flow_tbl_init(struct flow_table *table)
306 struct table_instance *ti;
307 struct mask_array *ma;
309 table->mask_cache = __alloc_percpu(sizeof(struct mask_cache_entry) *
310 MC_HASH_ENTRIES, __alignof__(struct mask_cache_entry));
311 if (!table->mask_cache)
314 ma = tbl_mask_array_alloc(MASK_ARRAY_SIZE_MIN);
316 goto free_mask_cache;
318 ti = table_instance_alloc(TBL_MIN_BUCKETS);
320 goto free_mask_array;
322 rcu_assign_pointer(table->ti, ti);
323 rcu_assign_pointer(table->mask_array, ma);
324 table->last_rehash = jiffies;
329 kfree((struct mask_array __force *)table->mask_array);
331 free_percpu(table->mask_cache);
335 static void flow_tbl_destroy_rcu_cb(struct rcu_head *rcu)
337 struct table_instance *ti = container_of(rcu, struct table_instance, rcu);
339 __table_instance_destroy(ti);
342 static void table_instance_destroy(struct table_instance *ti, bool deferred)
352 for (i = 0; i < ti->n_buckets; i++) {
353 struct sw_flow *flow;
354 struct hlist_head *head = flex_array_get(ti->buckets, i);
355 struct hlist_node *n;
356 int ver = ti->node_ver;
358 hlist_for_each_entry_safe(flow, n, head, hash_node[ver]) {
359 hlist_del_rcu(&flow->hash_node[ver]);
360 ovs_flow_free(flow, deferred);
366 call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
368 __table_instance_destroy(ti);
371 /* No need for locking this function is called from RCU callback or
373 void ovs_flow_tbl_destroy(struct flow_table *table)
375 struct table_instance *ti = (struct table_instance __force *)table->ti;
377 free_percpu(table->mask_cache);
378 kfree((struct mask_array __force *)table->mask_array);
379 table_instance_destroy(ti, false);
382 struct sw_flow *ovs_flow_tbl_dump_next(struct table_instance *ti,
383 u32 *bucket, u32 *last)
385 struct sw_flow *flow;
386 struct hlist_head *head;
391 while (*bucket < ti->n_buckets) {
393 head = flex_array_get(ti->buckets, *bucket);
394 hlist_for_each_entry_rcu(flow, head, hash_node[ver]) {
409 static struct hlist_head *find_bucket(struct table_instance *ti, u32 hash)
411 hash = jhash_1word(hash, ti->hash_seed);
412 return flex_array_get(ti->buckets,
413 (hash & (ti->n_buckets - 1)));
416 static void table_instance_insert(struct table_instance *ti, struct sw_flow *flow)
418 struct hlist_head *head;
420 head = find_bucket(ti, flow->hash);
421 hlist_add_head_rcu(&flow->hash_node[ti->node_ver], head);
424 static void flow_table_copy_flows(struct table_instance *old,
425 struct table_instance *new)
430 old_ver = old->node_ver;
431 new->node_ver = !old_ver;
433 /* Insert in new table. */
434 for (i = 0; i < old->n_buckets; i++) {
435 struct sw_flow *flow;
436 struct hlist_head *head;
438 head = flex_array_get(old->buckets, i);
440 hlist_for_each_entry(flow, head, hash_node[old_ver])
441 table_instance_insert(new, flow);
444 old->keep_flows = true;
447 static struct table_instance *table_instance_rehash(struct table_instance *ti,
450 struct table_instance *new_ti;
452 new_ti = table_instance_alloc(n_buckets);
456 flow_table_copy_flows(ti, new_ti);
461 int ovs_flow_tbl_flush(struct flow_table *flow_table)
463 struct table_instance *old_ti;
464 struct table_instance *new_ti;
466 old_ti = ovsl_dereference(flow_table->ti);
467 new_ti = table_instance_alloc(TBL_MIN_BUCKETS);
471 rcu_assign_pointer(flow_table->ti, new_ti);
472 flow_table->last_rehash = jiffies;
473 flow_table->count = 0;
475 table_instance_destroy(old_ti, true);
479 static u32 flow_hash(const struct sw_flow_key *key, int key_start,
482 const u32 *hash_key = (const u32 *)((const u8 *)key + key_start);
483 int hash_u32s = (key_end - key_start) >> 2;
485 /* Make sure number of hash bytes are multiple of u32. */
486 BUILD_BUG_ON(sizeof(long) % sizeof(u32));
488 return arch_fast_hash2(hash_key, hash_u32s, 0);
491 static int flow_key_start(const struct sw_flow_key *key)
493 if (key->tun_key.ipv4_dst)
496 return rounddown(offsetof(struct sw_flow_key, phy),
500 static bool cmp_key(const struct sw_flow_key *key1,
501 const struct sw_flow_key *key2,
502 int key_start, int key_end)
504 const long *cp1 = (const long *)((const u8 *)key1 + key_start);
505 const long *cp2 = (const long *)((const u8 *)key2 + key_start);
509 for (i = key_start; i < key_end; i += sizeof(long))
510 diffs |= *cp1++ ^ *cp2++;
515 static bool flow_cmp_masked_key(const struct sw_flow *flow,
516 const struct sw_flow_key *key,
517 int key_start, int key_end)
519 return cmp_key(&flow->key, key, key_start, key_end);
522 bool ovs_flow_cmp_unmasked_key(const struct sw_flow *flow,
523 struct sw_flow_match *match)
525 struct sw_flow_key *key = match->key;
526 int key_start = flow_key_start(key);
527 int key_end = match->range.end;
529 return cmp_key(&flow->unmasked_key, key, key_start, key_end);
532 static struct sw_flow *masked_flow_lookup(struct table_instance *ti,
533 const struct sw_flow_key *unmasked,
534 struct sw_flow_mask *mask,
537 struct sw_flow *flow;
538 struct hlist_head *head;
539 int key_start = mask->range.start;
540 int key_end = mask->range.end;
542 struct sw_flow_key masked_key;
544 ovs_flow_mask_key(&masked_key, unmasked, mask);
545 hash = flow_hash(&masked_key, key_start, key_end);
546 head = find_bucket(ti, hash);
548 hlist_for_each_entry_rcu(flow, head, hash_node[ti->node_ver]) {
549 if (flow->mask == mask && flow->hash == hash &&
550 flow_cmp_masked_key(flow, &masked_key,
557 static struct sw_flow *flow_lookup(struct flow_table *tbl,
558 struct table_instance *ti,
559 struct mask_array *ma,
560 const struct sw_flow_key *key,
564 struct sw_flow *flow;
567 for (i = 0; i < ma->max; i++) {
568 struct sw_flow_mask *mask;
570 mask = rcu_dereference_ovsl(ma->masks[i]);
574 flow = masked_flow_lookup(ti, key, mask, n_mask_hit);
575 if (flow) { /* Found */
585 * mask_cache maps flow to probable mask. This cache is not tightly
586 * coupled cache, It means updates to mask list can result in inconsistent
587 * cache entry in mask cache.
588 * This is per cpu cache and is divided in MC_HASH_SEGS segments.
589 * In case of a hash collision the entry is hashed in next segment.
591 struct sw_flow *ovs_flow_tbl_lookup_stats(struct flow_table *tbl,
592 const struct sw_flow_key *key,
596 struct mask_array *ma = rcu_dereference(tbl->mask_array);
597 struct table_instance *ti = rcu_dereference(tbl->ti);
598 struct mask_cache_entry *entries, *ce;
599 struct sw_flow *flow;
604 if (unlikely(!skb_hash)) {
605 u32 __always_unused mask_index;
607 return flow_lookup(tbl, ti, ma, key, n_mask_hit, &mask_index);
611 entries = this_cpu_ptr(tbl->mask_cache);
613 /* Find the cache entry 'ce' to operate on. */
614 for (seg = 0; seg < MC_HASH_SEGS; seg++) {
615 int index = hash & (MC_HASH_ENTRIES - 1);
616 struct mask_cache_entry *e;
619 if (e->skb_hash == skb_hash) {
620 struct sw_flow_mask *cache;
621 int i = e->mask_index;
623 if (likely(i < ma->max)) {
624 cache = rcu_dereference(ma->masks[i]);
626 flow = masked_flow_lookup(ti, key,
633 /* Cache miss. This is the best cache
634 * replacement candidate. */
640 if (!ce || e->skb_hash < ce->skb_hash)
641 ce = e; /* A better replacement cache candidate. */
643 hash >>= MC_HASH_SHIFT;
646 /* Cache miss, do full lookup. */
647 flow = flow_lookup(tbl, ti, ma, key, n_mask_hit, &ce->mask_index);
649 ce->skb_hash = skb_hash;
654 struct sw_flow *ovs_flow_tbl_lookup(struct flow_table *tbl,
655 const struct sw_flow_key *key)
657 struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
658 struct mask_array *ma = rcu_dereference_ovsl(tbl->mask_array);
659 u32 __always_unused n_mask_hit;
660 u32 __always_unused index;
662 return flow_lookup(tbl, ti, ma, key, &n_mask_hit, &index);
665 struct sw_flow *ovs_flow_tbl_lookup_exact(struct flow_table *tbl,
666 struct sw_flow_match *match)
668 struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
669 struct mask_array *ma = rcu_dereference_ovsl(tbl->mask_array);
670 struct sw_flow *flow;
671 u32 __always_unused n_mask_hit;
674 /* Always called under ovs-mutex. */
675 for (i = 0; i < ma->count; i++) {
676 struct sw_flow_mask *mask;
678 mask = ovsl_dereference(ma->masks[i]);
680 flow = masked_flow_lookup(ti, match->key, mask, &n_mask_hit);
681 if (flow && ovs_flow_cmp_unmasked_key(flow, match)) { /* Found */
689 int ovs_flow_tbl_num_masks(const struct flow_table *table)
691 struct mask_array *ma;
693 ma = rcu_dereference_ovsl(table->mask_array);
697 static struct table_instance *table_instance_expand(struct table_instance *ti)
699 return table_instance_rehash(ti, ti->n_buckets * 2);
702 /* Remove 'mask' from the mask list, if it is not needed any more. */
703 static void flow_mask_remove(struct flow_table *tbl, struct sw_flow_mask *mask)
706 /* ovs-lock is required to protect mask-refcount and
710 BUG_ON(!mask->ref_count);
713 if (!mask->ref_count) {
714 struct mask_array *ma;
716 ma = ovsl_dereference(tbl->mask_array);
717 /* Shrink the mask array if necessary. */
718 if (ma->max > MASK_ARRAY_SIZE_MIN * 2
719 && ma->count <= ma->max / 4) {
721 tbl_mask_array_realloc(tbl, ma->max / 2);
722 ma = ovsl_dereference(tbl->mask_array);
725 tbl_mask_array_delete_mask(ma, mask);
726 call_rcu(&mask->rcu, rcu_free_sw_flow_mask_cb);
731 /* Must be called with OVS mutex held. */
732 void ovs_flow_tbl_remove(struct flow_table *table, struct sw_flow *flow)
734 struct table_instance *ti = ovsl_dereference(table->ti);
736 BUG_ON(table->count == 0);
737 hlist_del_rcu(&flow->hash_node[ti->node_ver]);
740 /* RCU delete the mask. 'flow->mask' is not NULLed, as it should be
741 * accessible as long as the RCU read lock is held. */
742 flow_mask_remove(table, flow->mask);
745 static struct sw_flow_mask *mask_alloc(void)
747 struct sw_flow_mask *mask;
749 mask = kmalloc(sizeof(*mask), GFP_KERNEL);
756 static bool mask_equal(const struct sw_flow_mask *a,
757 const struct sw_flow_mask *b)
759 const u8 *a_ = (const u8 *)&a->key + a->range.start;
760 const u8 *b_ = (const u8 *)&b->key + b->range.start;
762 return (a->range.end == b->range.end)
763 && (a->range.start == b->range.start)
764 && (memcmp(a_, b_, range_n_bytes(&a->range)) == 0);
767 static struct sw_flow_mask *flow_mask_find(const struct flow_table *tbl,
768 const struct sw_flow_mask *mask)
770 struct mask_array *ma;
773 ma = ovsl_dereference(tbl->mask_array);
774 for (i = 0; i < ma->count; i++) {
775 struct sw_flow_mask *t;
777 t = ovsl_dereference(ma->masks[i]);
778 if (t && mask_equal(mask, t))
785 /* Add 'mask' into the mask list, if it is not already there. */
786 static int flow_mask_insert(struct flow_table *tbl, struct sw_flow *flow,
787 struct sw_flow_mask *new)
789 struct sw_flow_mask *mask;
791 mask = flow_mask_find(tbl, new);
793 struct mask_array *ma;
795 /* Allocate a new mask if none exsits. */
800 mask->key = new->key;
801 mask->range = new->range;
803 /* Add mask to mask-list. */
804 ma = ovsl_dereference(tbl->mask_array);
805 if (ma->count >= ma->max) {
808 err = tbl_mask_array_realloc(tbl, ma->max +
809 MASK_ARRAY_SIZE_MIN);
814 ma = ovsl_dereference(tbl->mask_array);
817 rcu_assign_pointer(ma->masks[ma->count], mask);
820 BUG_ON(!mask->ref_count);
828 /* Must be called with OVS mutex held. */
829 int ovs_flow_tbl_insert(struct flow_table *table, struct sw_flow *flow,
830 struct sw_flow_mask *mask)
832 struct table_instance *new_ti = NULL;
833 struct table_instance *ti;
836 err = flow_mask_insert(table, flow, mask);
840 flow->hash = flow_hash(&flow->key, flow->mask->range.start,
841 flow->mask->range.end);
842 ti = ovsl_dereference(table->ti);
843 table_instance_insert(ti, flow);
846 /* Expand table, if necessary, to make room. */
847 if (table->count > ti->n_buckets)
848 new_ti = table_instance_expand(ti);
849 else if (time_after(jiffies, table->last_rehash + REHASH_INTERVAL))
850 new_ti = table_instance_rehash(ti, ti->n_buckets);
853 rcu_assign_pointer(table->ti, new_ti);
854 table_instance_destroy(ti, true);
855 table->last_rehash = jiffies;
860 /* Initializes the flow module.
861 * Returns zero if successful or a negative error code. */
862 int ovs_flow_init(void)
864 BUILD_BUG_ON(__alignof__(struct sw_flow_key) % __alignof__(long));
865 BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
867 flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
868 + (num_possible_nodes()
869 * sizeof(struct flow_stats *)),
871 if (flow_cache == NULL)
875 = kmem_cache_create("sw_flow_stats", sizeof(struct flow_stats),
876 0, SLAB_HWCACHE_ALIGN, NULL);
877 if (flow_stats_cache == NULL) {
878 kmem_cache_destroy(flow_cache);
886 /* Uninitializes the flow module. */
887 void ovs_flow_exit(void)
889 kmem_cache_destroy(flow_stats_cache);
890 kmem_cache_destroy(flow_cache);