2 * Copyright (c) 2010 Nicira Networks.
3 * Distributed under the terms of the GNU GPL version 2.
5 * Significant portions of this file may be copied from parts of the Linux
6 * kernel, by Linus Torvalds and others.
9 #include <linux/if_arp.h>
10 #include <linux/if_ether.h>
12 #include <linux/if_vlan.h>
14 #include <linux/in_route.h>
15 #include <linux/jhash.h>
16 #include <linux/kernel.h>
17 #include <linux/version.h>
18 #include <linux/workqueue.h>
20 #include <net/dsfield.h>
23 #include <net/inet_ecn.h>
25 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
28 #include <net/route.h>
37 #include "vport-generic.h"
38 #include "vport-internal_dev.h"
40 #ifdef NEED_CACHE_TIMEOUT
42 * On kernels where we can't quickly detect changes in the rest of the system
43 * we use an expiration time to invalidate the cache. A shorter expiration
44 * reduces the length of time that we may potentially blackhole packets while
45 * a longer time increases performance by reducing the frequency that the
46 * cache needs to be rebuilt. A variety of factors may cause the cache to be
47 * invalidated before the expiration time but this is the maximum. The time
48 * is expressed in jiffies.
50 #define MAX_CACHE_EXP HZ
54 * Interval to check for and remove caches that are no longer valid. Caches
55 * are checked for validity before they are used for packet encapsulation and
56 * old caches are removed at that time. However, if no packets are sent through
57 * the tunnel then the cache will never be destroyed. Since it holds
58 * references to a number of system objects, the cache will continue to use
59 * system resources by not allowing those objects to be destroyed. The cache
60 * cleaner is periodically run to free invalid caches. It does not
61 * significantly affect system performance. A lower interval will release
62 * resources faster but will itself consume resources by requiring more frequent
63 * checks. A longer interval may result in messages being printed to the kernel
64 * message buffer about unreleased resources. The interval is expressed in
67 #define CACHE_CLEANER_INTERVAL (5 * HZ)
69 #define CACHE_DATA_ALIGN 16
71 /* Protected by RCU. */
72 static struct tbl *port_table __read_mostly;
74 static void cache_cleaner(struct work_struct *work);
75 static DECLARE_DELAYED_WORK(cache_cleaner_wq, cache_cleaner);
78 * These are just used as an optimization: they don't require any kind of
79 * synchronization because we could have just as easily read the value before
80 * the port change happened.
82 static unsigned int key_local_remote_ports __read_mostly;
83 static unsigned int key_remote_ports __read_mostly;
84 static unsigned int local_remote_ports __read_mostly;
85 static unsigned int remote_ports __read_mostly;
87 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,36)
88 #define rt_dst(rt) (rt->dst)
90 #define rt_dst(rt) (rt->u.dst)
93 static inline struct vport *tnl_vport_to_vport(const struct tnl_vport *tnl_vport)
95 return vport_from_priv(tnl_vport);
98 static inline struct tnl_vport *tnl_vport_table_cast(const struct tbl_node *node)
100 return container_of(node, struct tnl_vport, tbl_node);
103 static inline void schedule_cache_cleaner(void)
105 schedule_delayed_work(&cache_cleaner_wq, CACHE_CLEANER_INTERVAL);
108 static void free_cache(struct tnl_cache *cache)
113 flow_put(cache->flow);
114 ip_rt_put(cache->rt);
118 static void free_config_rcu(struct rcu_head *rcu)
120 struct tnl_mutable_config *c = container_of(rcu, struct tnl_mutable_config, rcu);
124 static void free_cache_rcu(struct rcu_head *rcu)
126 struct tnl_cache *c = container_of(rcu, struct tnl_cache, rcu);
130 static void assign_config_rcu(struct vport *vport,
131 struct tnl_mutable_config *new_config)
133 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
134 struct tnl_mutable_config *old_config;
136 old_config = tnl_vport->mutable;
137 rcu_assign_pointer(tnl_vport->mutable, new_config);
138 call_rcu(&old_config->rcu, free_config_rcu);
141 static void assign_cache_rcu(struct vport *vport, struct tnl_cache *new_cache)
143 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
144 struct tnl_cache *old_cache;
146 old_cache = tnl_vport->cache;
147 rcu_assign_pointer(tnl_vport->cache, new_cache);
150 call_rcu(&old_cache->rcu, free_cache_rcu);
153 static unsigned int *find_port_pool(const struct tnl_mutable_config *mutable)
155 if (mutable->port_config.flags & TNL_F_IN_KEY_MATCH) {
156 if (mutable->port_config.saddr)
157 return &local_remote_ports;
159 return &remote_ports;
161 if (mutable->port_config.saddr)
162 return &key_local_remote_ports;
164 return &key_remote_ports;
168 struct port_lookup_key {
169 const struct tnl_mutable_config *mutable;
177 * Modifies 'target' to store the rcu_dereferenced pointer that was used to do
180 static int port_cmp(const struct tbl_node *node, void *target)
182 const struct tnl_vport *tnl_vport = tnl_vport_table_cast(node);
183 struct port_lookup_key *lookup = target;
185 lookup->mutable = rcu_dereference(tnl_vport->mutable);
187 return (lookup->mutable->tunnel_type == lookup->tunnel_type &&
188 lookup->mutable->port_config.daddr == lookup->daddr &&
189 lookup->mutable->port_config.in_key == lookup->key &&
190 lookup->mutable->port_config.saddr == lookup->saddr);
193 static u32 port_hash(struct port_lookup_key *k)
195 u32 x = jhash_3words(k->saddr, k->daddr, k->tunnel_type, 0);
196 return jhash_2words(k->key >> 32, k->key, x);
199 static u32 mutable_hash(const struct tnl_mutable_config *mutable)
201 struct port_lookup_key lookup;
203 lookup.saddr = mutable->port_config.saddr;
204 lookup.daddr = mutable->port_config.daddr;
205 lookup.key = mutable->port_config.in_key;
206 lookup.tunnel_type = mutable->tunnel_type;
208 return port_hash(&lookup);
211 static void check_table_empty(void)
213 if (tbl_count(port_table) == 0) {
214 struct tbl *old_table = port_table;
216 cancel_delayed_work_sync(&cache_cleaner_wq);
217 rcu_assign_pointer(port_table, NULL);
218 tbl_deferred_destroy(old_table, NULL);
222 static int add_port(struct vport *vport)
224 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
228 struct tbl *new_table;
230 new_table = tbl_create(0);
234 rcu_assign_pointer(port_table, new_table);
235 schedule_cache_cleaner();
237 } else if (tbl_count(port_table) > tbl_n_buckets(port_table)) {
238 struct tbl *old_table = port_table;
239 struct tbl *new_table;
241 new_table = tbl_expand(old_table);
242 if (IS_ERR(new_table))
243 return PTR_ERR(new_table);
245 rcu_assign_pointer(port_table, new_table);
246 tbl_deferred_destroy(old_table, NULL);
249 err = tbl_insert(port_table, &tnl_vport->tbl_node, mutable_hash(tnl_vport->mutable));
255 (*find_port_pool(tnl_vport->mutable))++;
260 static int move_port(struct vport *vport, struct tnl_mutable_config *new_mutable)
263 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
266 hash = mutable_hash(new_mutable);
267 if (hash == tnl_vport->tbl_node.hash)
271 * Ideally we should make this move atomic to avoid having gaps in
272 * finding tunnels or the possibility of failure. However, if we do
273 * find a tunnel it will always be consistent.
275 err = tbl_remove(port_table, &tnl_vport->tbl_node);
279 err = tbl_insert(port_table, &tnl_vport->tbl_node, hash);
281 (*find_port_pool(tnl_vport->mutable))--;
287 (*find_port_pool(tnl_vport->mutable))--;
288 assign_config_rcu(vport, new_mutable);
289 (*find_port_pool(tnl_vport->mutable))++;
294 static int del_port(struct vport *vport)
296 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
299 err = tbl_remove(port_table, &tnl_vport->tbl_node);
304 (*find_port_pool(tnl_vport->mutable))--;
309 struct vport *tnl_find_port(__be32 saddr, __be32 daddr, __be64 key,
311 const struct tnl_mutable_config **mutable)
313 struct port_lookup_key lookup;
314 struct tbl *table = rcu_dereference(port_table);
315 struct tbl_node *tbl_node;
317 if (unlikely(!table))
320 lookup.saddr = saddr;
321 lookup.daddr = daddr;
323 if (tunnel_type & TNL_T_KEY_EXACT) {
325 lookup.tunnel_type = tunnel_type & ~TNL_T_KEY_MATCH;
327 if (key_local_remote_ports) {
328 tbl_node = tbl_lookup(table, &lookup, port_hash(&lookup), port_cmp);
333 if (key_remote_ports) {
336 tbl_node = tbl_lookup(table, &lookup, port_hash(&lookup), port_cmp);
340 lookup.saddr = saddr;
344 if (tunnel_type & TNL_T_KEY_MATCH) {
346 lookup.tunnel_type = tunnel_type & ~TNL_T_KEY_EXACT;
348 if (local_remote_ports) {
349 tbl_node = tbl_lookup(table, &lookup, port_hash(&lookup), port_cmp);
357 tbl_node = tbl_lookup(table, &lookup, port_hash(&lookup), port_cmp);
366 *mutable = lookup.mutable;
367 return tnl_vport_to_vport(tnl_vport_table_cast(tbl_node));
370 static inline void ecn_decapsulate(struct sk_buff *skb)
372 /* This is accessing the outer IP header of the tunnel, which we've
373 * already validated to be OK. skb->data is currently set to the start
374 * of the inner Ethernet header, and we've validated ETH_HLEN.
376 if (unlikely(INET_ECN_is_ce(ip_hdr(skb)->tos))) {
377 __be16 protocol = skb->protocol;
379 skb_set_network_header(skb, ETH_HLEN);
381 if (skb->protocol == htons(ETH_P_8021Q)) {
382 if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
385 protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
386 skb_set_network_header(skb, VLAN_ETH_HLEN);
389 if (protocol == htons(ETH_P_IP)) {
390 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
391 + sizeof(struct iphdr))))
394 IP_ECN_set_ce(ip_hdr(skb));
396 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
397 else if (protocol == htons(ETH_P_IPV6)) {
398 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
399 + sizeof(struct ipv6hdr))))
402 IP6_ECN_set_ce(ipv6_hdr(skb));
408 /* Called with rcu_read_lock. */
409 void tnl_rcv(struct vport *vport, struct sk_buff *skb)
411 /* Packets received by this function are in the following state:
412 * - skb->data points to the inner Ethernet header.
413 * - The inner Ethernet header is in the linear data area.
414 * - skb->csum does not include the inner Ethernet header.
415 * - The layer pointers point at the outer headers.
418 struct ethhdr *eh = (struct ethhdr *)skb->data;
420 if (likely(ntohs(eh->h_proto) >= 1536))
421 skb->protocol = eh->h_proto;
423 skb->protocol = htons(ETH_P_802_2);
429 ecn_decapsulate(skb);
430 compute_ip_summed(skb, false);
432 vport_receive(vport, skb);
435 static bool check_ipv4_address(__be32 addr)
437 if (ipv4_is_multicast(addr) || ipv4_is_lbcast(addr)
438 || ipv4_is_loopback(addr) || ipv4_is_zeronet(addr))
444 static bool ipv4_should_icmp(struct sk_buff *skb)
446 struct iphdr *old_iph = ip_hdr(skb);
448 /* Don't respond to L2 broadcast. */
449 if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
452 /* Don't respond to L3 broadcast or invalid addresses. */
453 if (!check_ipv4_address(old_iph->daddr) ||
454 !check_ipv4_address(old_iph->saddr))
457 /* Only respond to the first fragment. */
458 if (old_iph->frag_off & htons(IP_OFFSET))
461 /* Don't respond to ICMP error messages. */
462 if (old_iph->protocol == IPPROTO_ICMP) {
463 u8 icmp_type, *icmp_typep;
465 icmp_typep = skb_header_pointer(skb, (u8 *)old_iph +
466 (old_iph->ihl << 2) +
467 offsetof(struct icmphdr, type) -
468 skb->data, sizeof(icmp_type),
474 if (*icmp_typep > NR_ICMP_TYPES
475 || (*icmp_typep <= ICMP_PARAMETERPROB
476 && *icmp_typep != ICMP_ECHOREPLY
477 && *icmp_typep != ICMP_ECHO))
484 static void ipv4_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
485 unsigned int mtu, unsigned int payload_length)
487 struct iphdr *iph, *old_iph = ip_hdr(skb);
488 struct icmphdr *icmph;
491 iph = (struct iphdr *)skb_put(nskb, sizeof(struct iphdr));
492 icmph = (struct icmphdr *)skb_put(nskb, sizeof(struct icmphdr));
493 payload = skb_put(nskb, payload_length);
497 iph->ihl = sizeof(struct iphdr) >> 2;
498 iph->tos = (old_iph->tos & IPTOS_TOS_MASK) |
499 IPTOS_PREC_INTERNETCONTROL;
500 iph->tot_len = htons(sizeof(struct iphdr)
501 + sizeof(struct icmphdr)
503 get_random_bytes(&iph->id, sizeof(iph->id));
506 iph->protocol = IPPROTO_ICMP;
507 iph->daddr = old_iph->saddr;
508 iph->saddr = old_iph->daddr;
513 icmph->type = ICMP_DEST_UNREACH;
514 icmph->code = ICMP_FRAG_NEEDED;
515 icmph->un.gateway = htonl(mtu);
518 nskb->csum = csum_partial((u8 *)icmph, sizeof(struct icmphdr), 0);
519 nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_iph - skb->data,
520 payload, payload_length,
522 icmph->checksum = csum_fold(nskb->csum);
525 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
526 static bool ipv6_should_icmp(struct sk_buff *skb)
528 struct ipv6hdr *old_ipv6h = ipv6_hdr(skb);
530 int payload_off = (u8 *)(old_ipv6h + 1) - skb->data;
531 u8 nexthdr = ipv6_hdr(skb)->nexthdr;
533 /* Check source address is valid. */
534 addr_type = ipv6_addr_type(&old_ipv6h->saddr);
535 if (addr_type & IPV6_ADDR_MULTICAST || addr_type == IPV6_ADDR_ANY)
538 /* Don't reply to unspecified addresses. */
539 if (ipv6_addr_type(&old_ipv6h->daddr) == IPV6_ADDR_ANY)
542 /* Don't respond to ICMP error messages. */
543 payload_off = ipv6_skip_exthdr(skb, payload_off, &nexthdr);
547 if (nexthdr == NEXTHDR_ICMP) {
548 u8 icmp_type, *icmp_typep;
550 icmp_typep = skb_header_pointer(skb, payload_off +
551 offsetof(struct icmp6hdr,
553 sizeof(icmp_type), &icmp_type);
555 if (!icmp_typep || !(*icmp_typep & ICMPV6_INFOMSG_MASK))
562 static void ipv6_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
563 unsigned int mtu, unsigned int payload_length)
565 struct ipv6hdr *ipv6h, *old_ipv6h = ipv6_hdr(skb);
566 struct icmp6hdr *icmp6h;
569 ipv6h = (struct ipv6hdr *)skb_put(nskb, sizeof(struct ipv6hdr));
570 icmp6h = (struct icmp6hdr *)skb_put(nskb, sizeof(struct icmp6hdr));
571 payload = skb_put(nskb, payload_length);
576 memset(&ipv6h->flow_lbl, 0, sizeof(ipv6h->flow_lbl));
577 ipv6h->payload_len = htons(sizeof(struct icmp6hdr)
579 ipv6h->nexthdr = NEXTHDR_ICMP;
580 ipv6h->hop_limit = IPV6_DEFAULT_HOPLIMIT;
581 ipv6_addr_copy(&ipv6h->daddr, &old_ipv6h->saddr);
582 ipv6_addr_copy(&ipv6h->saddr, &old_ipv6h->daddr);
585 icmp6h->icmp6_type = ICMPV6_PKT_TOOBIG;
586 icmp6h->icmp6_code = 0;
587 icmp6h->icmp6_cksum = 0;
588 icmp6h->icmp6_mtu = htonl(mtu);
590 nskb->csum = csum_partial((u8 *)icmp6h, sizeof(struct icmp6hdr), 0);
591 nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_ipv6h - skb->data,
592 payload, payload_length,
594 icmp6h->icmp6_cksum = csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
595 sizeof(struct icmp6hdr)
597 ipv6h->nexthdr, nskb->csum);
601 bool tnl_frag_needed(struct vport *vport, const struct tnl_mutable_config *mutable,
602 struct sk_buff *skb, unsigned int mtu, __be64 flow_key)
604 unsigned int eth_hdr_len = ETH_HLEN;
605 unsigned int total_length = 0, header_length = 0, payload_length;
606 struct ethhdr *eh, *old_eh = eth_hdr(skb);
607 struct sk_buff *nskb;
610 if (skb->protocol == htons(ETH_P_IP)) {
611 if (mtu < IP_MIN_MTU)
614 if (!ipv4_should_icmp(skb))
617 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
618 else if (skb->protocol == htons(ETH_P_IPV6)) {
619 if (mtu < IPV6_MIN_MTU)
623 * In theory we should do PMTUD on IPv6 multicast messages but
624 * we don't have an address to send from so just fragment.
626 if (ipv6_addr_type(&ipv6_hdr(skb)->daddr) & IPV6_ADDR_MULTICAST)
629 if (!ipv6_should_icmp(skb))
637 if (old_eh->h_proto == htons(ETH_P_8021Q))
638 eth_hdr_len = VLAN_ETH_HLEN;
640 payload_length = skb->len - eth_hdr_len;
641 if (skb->protocol == htons(ETH_P_IP)) {
642 header_length = sizeof(struct iphdr) + sizeof(struct icmphdr);
643 total_length = min_t(unsigned int, header_length +
644 payload_length, 576);
646 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
648 header_length = sizeof(struct ipv6hdr) +
649 sizeof(struct icmp6hdr);
650 total_length = min_t(unsigned int, header_length +
651 payload_length, IPV6_MIN_MTU);
655 total_length = min(total_length, mutable->mtu);
656 payload_length = total_length - header_length;
658 nskb = dev_alloc_skb(NET_IP_ALIGN + eth_hdr_len + header_length +
663 skb_reserve(nskb, NET_IP_ALIGN);
665 /* Ethernet / VLAN */
666 eh = (struct ethhdr *)skb_put(nskb, eth_hdr_len);
667 memcpy(eh->h_dest, old_eh->h_source, ETH_ALEN);
668 memcpy(eh->h_source, mutable->eth_addr, ETH_ALEN);
669 nskb->protocol = eh->h_proto = old_eh->h_proto;
670 if (old_eh->h_proto == htons(ETH_P_8021Q)) {
671 struct vlan_ethhdr *vh = (struct vlan_ethhdr *)eh;
673 vh->h_vlan_TCI = vlan_eth_hdr(skb)->h_vlan_TCI;
674 vh->h_vlan_encapsulated_proto = skb->protocol;
676 skb_reset_mac_header(nskb);
679 if (skb->protocol == htons(ETH_P_IP))
680 ipv4_build_icmp(skb, nskb, mtu, payload_length);
681 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
683 ipv6_build_icmp(skb, nskb, mtu, payload_length);
687 * Assume that flow based keys are symmetric with respect to input
688 * and output and use the key that we were going to put on the
689 * outgoing packet for the fake received packet. If the keys are
690 * not symmetric then PMTUD needs to be disabled since we won't have
691 * any way of synthesizing packets.
693 if ((mutable->port_config.flags & (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION)) ==
694 (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION))
695 OVS_CB(nskb)->tun_id = flow_key;
697 compute_ip_summed(nskb, false);
698 vport_receive(vport, nskb);
703 static bool check_mtu(struct sk_buff *skb,
705 const struct tnl_mutable_config *mutable,
706 const struct rtable *rt, __be16 *frag_offp)
711 frag_off = (mutable->port_config.flags & TNL_F_PMTUD) ? htons(IP_DF) : 0;
713 mtu = dst_mtu(&rt_dst(rt))
715 - mutable->tunnel_hlen
716 - (eth_hdr(skb)->h_proto == htons(ETH_P_8021Q) ? VLAN_HLEN : 0);
720 if (skb->protocol == htons(ETH_P_IP)) {
721 struct iphdr *old_iph = ip_hdr(skb);
723 frag_off |= old_iph->frag_off & htons(IP_DF);
724 mtu = max(mtu, IP_MIN_MTU);
726 if ((old_iph->frag_off & htons(IP_DF)) &&
727 mtu < ntohs(old_iph->tot_len)) {
728 if (tnl_frag_needed(vport, mutable, skb, mtu, OVS_CB(skb)->tun_id))
732 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
733 else if (skb->protocol == htons(ETH_P_IPV6)) {
734 unsigned int packet_length = skb->len - ETH_HLEN
735 - (eth_hdr(skb)->h_proto == htons(ETH_P_8021Q) ? VLAN_HLEN : 0);
737 mtu = max(mtu, IPV6_MIN_MTU);
739 /* IPv6 requires PMTUD if the packet is above the minimum MTU. */
740 if (packet_length > IPV6_MIN_MTU)
741 frag_off = htons(IP_DF);
743 if (mtu < packet_length) {
744 if (tnl_frag_needed(vport, mutable, skb, mtu, OVS_CB(skb)->tun_id))
750 *frag_offp = frag_off;
758 static void create_tunnel_header(const struct vport *vport,
759 const struct tnl_mutable_config *mutable,
760 const struct rtable *rt, void *header)
762 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
763 struct iphdr *iph = header;
766 iph->ihl = sizeof(struct iphdr) >> 2;
767 iph->frag_off = htons(IP_DF);
768 iph->protocol = tnl_vport->tnl_ops->ipproto;
769 iph->tos = mutable->port_config.tos;
770 iph->daddr = rt->rt_dst;
771 iph->saddr = rt->rt_src;
772 iph->ttl = mutable->port_config.ttl;
774 iph->ttl = dst_metric(&rt_dst(rt), RTAX_HOPLIMIT);
776 tnl_vport->tnl_ops->build_header(vport, mutable, iph + 1);
779 static inline void *get_cached_header(const struct tnl_cache *cache)
781 return (void *)cache + ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN);
784 static inline bool check_cache_valid(const struct tnl_cache *cache,
785 const struct tnl_mutable_config *mutable)
788 #ifdef NEED_CACHE_TIMEOUT
789 time_before(jiffies, cache->expiration) &&
792 atomic_read(&init_net.ipv4.rt_genid) == cache->rt->rt_genid &&
795 rt_dst(cache->rt).hh->hh_lock.sequence == cache->hh_seq &&
797 mutable->seq == cache->mutable_seq &&
798 (!is_internal_dev(rt_dst(cache->rt).dev) ||
799 (cache->flow && !cache->flow->dead));
802 static int cache_cleaner_cb(struct tbl_node *tbl_node, void *aux)
804 struct tnl_vport *tnl_vport = tnl_vport_table_cast(tbl_node);
805 const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
806 const struct tnl_cache *cache = rcu_dereference(tnl_vport->cache);
808 if (cache && !check_cache_valid(cache, mutable) &&
809 spin_trylock_bh(&tnl_vport->cache_lock)) {
810 assign_cache_rcu(tnl_vport_to_vport(tnl_vport), NULL);
811 spin_unlock_bh(&tnl_vport->cache_lock);
817 static void cache_cleaner(struct work_struct *work)
819 schedule_cache_cleaner();
822 tbl_foreach(rcu_dereference(port_table), cache_cleaner_cb, NULL);
826 static inline void create_eth_hdr(struct tnl_cache *cache,
827 const struct rtable *rt)
829 void *cache_data = get_cached_header(cache);
830 int hh_len = rt_dst(rt).hh->hh_len;
831 int hh_off = HH_DATA_ALIGN(rt_dst(rt).hh->hh_len) - hh_len;
837 hh_seq = read_seqbegin(&rt_dst(rt).hh->hh_lock);
838 memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
839 } while (read_seqretry(&rt_dst(rt).hh->hh_lock, hh_seq));
841 cache->hh_seq = hh_seq;
843 read_lock_bh(&rt_dst(rt).hh->hh_lock);
844 memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
845 read_unlock_bh(&rt_dst(rt).hh->hh_lock);
849 static struct tnl_cache *build_cache(struct vport *vport,
850 const struct tnl_mutable_config *mutable,
853 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
854 struct tnl_cache *cache;
858 if (!(mutable->port_config.flags & TNL_F_HDR_CACHE))
862 * If there is no entry in the ARP cache or if this device does not
863 * support hard header caching just fall back to the IP stack.
869 * If lock is contended fall back to directly building the header.
870 * We're not going to help performance by sitting here spinning.
872 if (!spin_trylock_bh(&tnl_vport->cache_lock))
875 cache = tnl_vport->cache;
876 if (check_cache_valid(cache, mutable))
881 cache_len = rt_dst(rt).hh->hh_len + mutable->tunnel_hlen;
883 cache = kzalloc(ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN) +
884 cache_len, GFP_ATOMIC);
888 cache->len = cache_len;
890 create_eth_hdr(cache, rt);
891 cache_data = get_cached_header(cache) + rt_dst(rt).hh->hh_len;
893 create_tunnel_header(vport, mutable, rt, cache_data);
895 cache->mutable_seq = mutable->seq;
897 #ifdef NEED_CACHE_TIMEOUT
898 cache->expiration = jiffies + tnl_vport->cache_exp_interval;
901 if (is_internal_dev(rt_dst(rt).dev)) {
902 struct odp_flow_key flow_key;
903 struct tbl_node *flow_node;
909 vport = internal_dev_get_vport(rt_dst(rt).dev);
913 skb = alloc_skb(cache->len, GFP_ATOMIC);
917 __skb_put(skb, cache->len);
918 memcpy(skb->data, get_cached_header(cache), cache->len);
920 err = flow_extract(skb, vport->port_no, &flow_key, &is_frag);
926 flow_node = tbl_lookup(rcu_dereference(vport->dp->table),
927 &flow_key, flow_hash(&flow_key),
930 struct sw_flow *flow = flow_cast(flow_node);
938 assign_cache_rcu(vport, cache);
941 spin_unlock_bh(&tnl_vport->cache_lock);
946 static struct rtable *find_route(struct vport *vport,
947 const struct tnl_mutable_config *mutable,
948 u8 tos, struct tnl_cache **cache)
950 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
951 struct tnl_cache *cur_cache = rcu_dereference(tnl_vport->cache);
956 if (likely(tos == mutable->port_config.tos &&
957 check_cache_valid(cur_cache, mutable))) {
959 return cur_cache->rt;
962 struct flowi fl = { .nl_u = { .ip4_u =
963 { .daddr = mutable->port_config.daddr,
964 .saddr = mutable->port_config.saddr,
966 .proto = tnl_vport->tnl_ops->ipproto };
968 if (unlikely(ip_route_output_key(&init_net, &rt, &fl)))
971 if (likely(tos == mutable->port_config.tos))
972 *cache = build_cache(vport, mutable, rt);
978 static struct sk_buff *check_headroom(struct sk_buff *skb, int headroom)
980 if (skb_headroom(skb) < headroom || skb_header_cloned(skb)) {
981 struct sk_buff *nskb = skb_realloc_headroom(skb, headroom + 16);
982 if (unlikely(!nskb)) {
984 return ERR_PTR(-ENOMEM);
987 set_skb_csum_bits(skb, nskb);
990 skb_set_owner_w(nskb, skb->sk);
999 static inline bool need_linearize(const struct sk_buff *skb)
1003 if (unlikely(skb_shinfo(skb)->frag_list))
1007 * Generally speaking we should linearize if there are paged frags.
1008 * However, if all of the refcounts are 1 we know nobody else can
1009 * change them from underneath us and we can skip the linearization.
1011 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
1012 if (unlikely(page_count(skb_shinfo(skb)->frags[0].page) > 1))
1018 static struct sk_buff *handle_offloads(struct sk_buff *skb,
1019 const struct tnl_mutable_config *mutable,
1020 const struct rtable *rt)
1025 forward_ip_summed(skb);
1027 err = vswitch_skb_checksum_setup(skb);
1031 min_headroom = LL_RESERVED_SPACE(rt_dst(rt).dev) + rt_dst(rt).header_len
1032 + mutable->tunnel_hlen;
1034 if (skb_is_gso(skb)) {
1035 struct sk_buff *nskb;
1038 * If we are doing GSO on a pskb it is better to make sure that
1039 * the headroom is correct now. We will only have to copy the
1040 * portion in the linear data area and GSO will preserve
1041 * headroom when it creates the segments. This is particularly
1042 * beneficial on Xen where we get a lot of GSO pskbs.
1043 * Conversely, we avoid copying if it is just to get our own
1044 * writable clone because GSO will do the copy for us.
1046 if (skb_headroom(skb) < min_headroom) {
1047 skb = check_headroom(skb, min_headroom);
1054 nskb = skb_gso_segment(skb, 0);
1057 err = PTR_ERR(nskb);
1063 skb = check_headroom(skb, min_headroom);
1069 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1071 * Pages aren't locked and could change at any time.
1072 * If this happens after we compute the checksum, the
1073 * checksum will be wrong. We linearize now to avoid
1076 if (unlikely(need_linearize(skb))) {
1077 err = __skb_linearize(skb);
1082 err = skb_checksum_help(skb);
1085 } else if (skb->ip_summed == CHECKSUM_COMPLETE)
1086 skb->ip_summed = CHECKSUM_NONE;
1094 return ERR_PTR(err);
1097 static int send_frags(struct sk_buff *skb,
1098 const struct tnl_mutable_config *mutable)
1105 struct sk_buff *next = skb->next;
1106 int frag_len = skb->len - mutable->tunnel_hlen;
1109 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1111 err = ip_local_out(skb);
1112 if (likely(net_xmit_eval(err) == 0))
1113 sent_len += frag_len;
1126 * There's no point in continuing to send fragments once one has been
1127 * dropped so just free the rest. This may help improve the congestion
1128 * that caused the first packet to be dropped.
1130 tnl_free_linked_skbs(skb);
1134 int tnl_send(struct vport *vport, struct sk_buff *skb)
1136 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1137 const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
1139 enum vport_err_type err = VPORT_E_TX_ERROR;
1141 struct dst_entry *unattached_dst = NULL;
1142 struct tnl_cache *cache;
1149 /* Validate the protocol headers before we try to use them. */
1150 if (skb->protocol == htons(ETH_P_8021Q)) {
1151 if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
1154 skb->protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
1155 skb_set_network_header(skb, VLAN_ETH_HLEN);
1158 if (skb->protocol == htons(ETH_P_IP)) {
1159 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1160 + sizeof(struct iphdr))))
1163 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1164 else if (skb->protocol == htons(ETH_P_IPV6)) {
1165 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1166 + sizeof(struct ipv6hdr))))
1172 if (skb->protocol == htons(ETH_P_IP))
1173 inner_tos = ip_hdr(skb)->tos;
1174 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1175 else if (skb->protocol == htons(ETH_P_IPV6))
1176 inner_tos = ipv6_get_dsfield(ipv6_hdr(skb));
1181 if (mutable->port_config.flags & TNL_F_TOS_INHERIT)
1184 tos = mutable->port_config.tos;
1186 tos = INET_ECN_encapsulate(tos, inner_tos);
1189 rt = find_route(vport, mutable, tos, &cache);
1192 if (unlikely(!cache))
1193 unattached_dst = &rt_dst(rt);
1201 skb = handle_offloads(skb, mutable, rt);
1206 if (unlikely(!check_mtu(skb, vport, mutable, rt, &frag_off))) {
1207 err = VPORT_E_TX_DROPPED;
1212 * If we are over the MTU, allow the IP stack to handle fragmentation.
1213 * Fragmentation is a slow path anyways.
1215 if (unlikely(skb->len + mutable->tunnel_hlen > dst_mtu(&rt_dst(rt)) &&
1217 unattached_dst = &rt_dst(rt);
1218 dst_hold(unattached_dst);
1223 ttl = mutable->port_config.ttl;
1225 ttl = dst_metric(&rt_dst(rt), RTAX_HOPLIMIT);
1227 if (mutable->port_config.flags & TNL_F_TTL_INHERIT) {
1228 if (skb->protocol == htons(ETH_P_IP))
1229 ttl = ip_hdr(skb)->ttl;
1230 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1231 else if (skb->protocol == htons(ETH_P_IPV6))
1232 ttl = ipv6_hdr(skb)->hop_limit;
1238 struct sk_buff *next_skb = skb->next;
1241 if (likely(cache)) {
1242 skb_push(skb, cache->len);
1243 memcpy(skb->data, get_cached_header(cache), cache->len);
1244 skb_reset_mac_header(skb);
1245 skb_set_network_header(skb, rt_dst(rt).hh->hh_len);
1248 skb_push(skb, mutable->tunnel_hlen);
1249 create_tunnel_header(vport, mutable, rt, skb->data);
1250 skb_reset_network_header(skb);
1253 skb_dst_set(skb, dst_clone(unattached_dst));
1255 skb_dst_set(skb, unattached_dst);
1256 unattached_dst = NULL;
1259 skb_set_transport_header(skb, skb_network_offset(skb) + sizeof(struct iphdr));
1264 iph->frag_off = frag_off;
1265 ip_select_ident(iph, &rt_dst(rt), NULL);
1267 skb = tnl_vport->tnl_ops->update_header(vport, mutable, &rt_dst(rt), skb);
1271 if (likely(cache)) {
1272 int orig_len = skb->len - cache->len;
1273 struct vport *cache_vport = internal_dev_get_vport(rt_dst(rt).dev);
1275 skb->protocol = htons(ETH_P_IP);
1276 iph->tot_len = htons(skb->len - skb_network_offset(skb));
1280 OVS_CB(skb)->flow = cache->flow;
1281 compute_ip_summed(skb, true);
1282 vport_receive(cache_vport, skb);
1283 sent_len += orig_len;
1287 skb->dev = rt_dst(rt).dev;
1288 err = dev_queue_xmit(skb);
1290 if (likely(net_xmit_eval(err) == 0))
1291 sent_len += orig_len;
1294 sent_len += send_frags(skb, mutable);
1300 if (unlikely(sent_len == 0))
1301 vport_record_error(vport, VPORT_E_TX_DROPPED);
1306 tnl_free_linked_skbs(skb);
1308 dst_release(unattached_dst);
1309 vport_record_error(vport, err);
1314 static int set_config(const void *config, const struct tnl_ops *tnl_ops,
1315 const struct vport *cur_vport,
1316 struct tnl_mutable_config *mutable)
1318 const struct vport *old_vport;
1319 const struct tnl_mutable_config *old_mutable;
1321 mutable->port_config = *(struct tnl_port_config *)config;
1323 if (mutable->port_config.daddr == 0)
1326 if (mutable->port_config.tos != RT_TOS(mutable->port_config.tos))
1329 mutable->tunnel_hlen = tnl_ops->hdr_len(&mutable->port_config);
1330 if (mutable->tunnel_hlen < 0)
1331 return mutable->tunnel_hlen;
1333 mutable->tunnel_hlen += sizeof(struct iphdr);
1335 mutable->tunnel_type = tnl_ops->tunnel_type;
1336 if (mutable->port_config.flags & TNL_F_IN_KEY_MATCH) {
1337 mutable->tunnel_type |= TNL_T_KEY_MATCH;
1338 mutable->port_config.in_key = 0;
1340 mutable->tunnel_type |= TNL_T_KEY_EXACT;
1342 old_vport = tnl_find_port(mutable->port_config.saddr,
1343 mutable->port_config.daddr,
1344 mutable->port_config.in_key,
1345 mutable->tunnel_type,
1348 if (old_vport && old_vport != cur_vport)
1351 if (mutable->port_config.flags & TNL_F_OUT_KEY_ACTION)
1352 mutable->port_config.out_key = 0;
1357 struct vport *tnl_create(const struct vport_parms *parms,
1358 const struct vport_ops *vport_ops,
1359 const struct tnl_ops *tnl_ops)
1361 struct vport *vport;
1362 struct tnl_vport *tnl_vport;
1363 int initial_frag_id;
1366 vport = vport_alloc(sizeof(struct tnl_vport), vport_ops, parms);
1367 if (IS_ERR(vport)) {
1368 err = PTR_ERR(vport);
1372 tnl_vport = tnl_vport_priv(vport);
1374 strcpy(tnl_vport->name, parms->name);
1375 tnl_vport->tnl_ops = tnl_ops;
1377 tnl_vport->mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1378 if (!tnl_vport->mutable) {
1380 goto error_free_vport;
1383 vport_gen_rand_ether_addr(tnl_vport->mutable->eth_addr);
1384 tnl_vport->mutable->mtu = ETH_DATA_LEN;
1386 get_random_bytes(&initial_frag_id, sizeof(int));
1387 atomic_set(&tnl_vport->frag_id, initial_frag_id);
1389 err = set_config(parms->config, tnl_ops, NULL, tnl_vport->mutable);
1391 goto error_free_mutable;
1393 spin_lock_init(&tnl_vport->cache_lock);
1395 #ifdef NEED_CACHE_TIMEOUT
1396 tnl_vport->cache_exp_interval = MAX_CACHE_EXP -
1397 (net_random() % (MAX_CACHE_EXP / 2));
1400 err = add_port(vport);
1402 goto error_free_mutable;
1407 kfree(tnl_vport->mutable);
1411 return ERR_PTR(err);
1414 int tnl_modify(struct vport *vport, struct odp_port *port)
1416 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1417 struct tnl_mutable_config *mutable;
1420 mutable = kmemdup(tnl_vport->mutable, sizeof(struct tnl_mutable_config), GFP_KERNEL);
1426 err = set_config(port->config, tnl_vport->tnl_ops, vport, mutable);
1432 err = move_port(vport, mutable);
1444 static void free_port_rcu(struct rcu_head *rcu)
1446 struct tnl_vport *tnl_vport = container_of(rcu, struct tnl_vport, rcu);
1448 spin_lock_bh(&tnl_vport->cache_lock);
1449 free_cache(tnl_vport->cache);
1450 spin_unlock_bh(&tnl_vport->cache_lock);
1452 kfree(tnl_vport->mutable);
1453 vport_free(tnl_vport_to_vport(tnl_vport));
1456 int tnl_destroy(struct vport *vport)
1458 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1459 const struct tnl_mutable_config *old_mutable;
1461 if (vport == tnl_find_port(tnl_vport->mutable->port_config.saddr,
1462 tnl_vport->mutable->port_config.daddr,
1463 tnl_vport->mutable->port_config.in_key,
1464 tnl_vport->mutable->tunnel_type,
1468 call_rcu(&tnl_vport->rcu, free_port_rcu);
1473 int tnl_set_mtu(struct vport *vport, int mtu)
1475 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1476 struct tnl_mutable_config *mutable;
1478 mutable = kmemdup(tnl_vport->mutable, sizeof(struct tnl_mutable_config), GFP_KERNEL);
1483 assign_config_rcu(vport, mutable);
1488 int tnl_set_addr(struct vport *vport, const unsigned char *addr)
1490 struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1491 struct tnl_mutable_config *mutable;
1493 mutable = kmemdup(tnl_vport->mutable, sizeof(struct tnl_mutable_config), GFP_KERNEL);
1497 memcpy(mutable->eth_addr, addr, ETH_ALEN);
1498 assign_config_rcu(vport, mutable);
1503 const char *tnl_get_name(const struct vport *vport)
1505 const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1506 return tnl_vport->name;
1509 const unsigned char *tnl_get_addr(const struct vport *vport)
1511 const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1512 return rcu_dereference(tnl_vport->mutable)->eth_addr;
1515 int tnl_get_mtu(const struct vport *vport)
1517 const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1518 return rcu_dereference(tnl_vport->mutable)->mtu;
1521 void tnl_free_linked_skbs(struct sk_buff *skb)
1527 struct sk_buff *next = skb->next;