datapath: Clarify locking.
[sliver-openvswitch.git] / datapath / flow.c
1 /*
2  * Copyright (c) 2007-2014 Nicira, Inc.
3  *
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.
7  *
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.
12  *
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
16  * 02110-1301, USA
17  */
18
19 #include "flow.h"
20 #include "datapath.h"
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/jhash.h>
29 #include <linux/jiffies.h>
30 #include <linux/llc.h>
31 #include <linux/module.h>
32 #include <linux/in.h>
33 #include <linux/rcupdate.h>
34 #include <linux/if_arp.h>
35 #include <linux/ip.h>
36 #include <linux/ipv6.h>
37 #include <linux/sctp.h>
38 #include <linux/smp.h>
39 #include <linux/tcp.h>
40 #include <linux/udp.h>
41 #include <linux/icmp.h>
42 #include <linux/icmpv6.h>
43 #include <linux/rculist.h>
44 #include <net/ip.h>
45 #include <net/ipv6.h>
46 #include <net/ndisc.h>
47
48 #include "vlan.h"
49
50 u64 ovs_flow_used_time(unsigned long flow_jiffies)
51 {
52         struct timespec cur_ts;
53         u64 cur_ms, idle_ms;
54
55         ktime_get_ts(&cur_ts);
56         idle_ms = jiffies_to_msecs(jiffies - flow_jiffies);
57         cur_ms = (u64)cur_ts.tv_sec * MSEC_PER_SEC +
58                  cur_ts.tv_nsec / NSEC_PER_MSEC;
59
60         return cur_ms - idle_ms;
61 }
62
63 #define TCP_FLAGS_BE16(tp) (*(__be16 *)&tcp_flag_word(tp) & htons(0x0FFF))
64
65 void ovs_flow_stats_update(struct sw_flow *flow, struct sk_buff *skb)
66 {
67         struct flow_stats *stats;
68         __be16 tcp_flags = flow->key.tp.flags;
69         int node = numa_node_id();
70
71         stats = rcu_dereference(flow->stats[node]);
72
73         /* Check if already have node-specific stats. */
74         if (likely(stats)) {
75                 spin_lock(&stats->lock);
76                 /* Mark if we write on the pre-allocated stats. */
77                 if (node == 0 && unlikely(flow->stats_last_writer != node))
78                         flow->stats_last_writer = node;
79         } else {
80                 stats = rcu_dereference(flow->stats[0]); /* Pre-allocated. */
81                 spin_lock(&stats->lock);
82
83                 /* If the current NUMA-node is the only writer on the
84                  * pre-allocated stats keep using them.
85                  */
86                 if (unlikely(flow->stats_last_writer != node)) {
87                         /* A previous locker may have already allocated the
88                          * stats, so we need to check again.  If node-specific
89                          * stats were already allocated, we update the pre-
90                          * allocated stats as we have already locked them.
91                          */
92                         if (likely(flow->stats_last_writer != NUMA_NO_NODE)
93                             && likely(!rcu_dereference(flow->stats[node]))) {
94                                 /* Try to allocate node-specific stats. */
95                                 struct flow_stats *new_stats;
96
97                                 new_stats =
98                                         kmem_cache_alloc_node(flow_stats_cache,
99                                                               GFP_THISNODE |
100                                                               __GFP_NOMEMALLOC,
101                                                               node);
102                                 if (likely(new_stats)) {
103                                         new_stats->used = jiffies;
104                                         new_stats->packet_count = 1;
105                                         new_stats->byte_count = skb->len;
106                                         new_stats->tcp_flags = tcp_flags;
107                                         spin_lock_init(&new_stats->lock);
108
109                                         rcu_assign_pointer(flow->stats[node],
110                                                            new_stats);
111                                         goto unlock;
112                                 }
113                         }
114                         flow->stats_last_writer = node;
115                 }
116         }
117
118         stats->used = jiffies;
119         stats->packet_count++;
120         stats->byte_count += skb->len;
121         stats->tcp_flags |= tcp_flags;
122 unlock:
123         spin_unlock(&stats->lock);
124 }
125
126 /* Called with ovs_mutex. */
127 void ovs_flow_stats_get(struct sw_flow *flow, struct ovs_flow_stats *ovs_stats,
128                         unsigned long *used, __be16 *tcp_flags)
129 {
130         int node;
131
132         *used = 0;
133         *tcp_flags = 0;
134         memset(ovs_stats, 0, sizeof(*ovs_stats));
135
136         for_each_node(node) {
137                 struct flow_stats *stats = ovsl_dereference(flow->stats[node]);
138
139                 if (stats) {
140                         /* Local CPU may write on non-local stats, so we must
141                          * block bottom-halves here.
142                          */
143                         spin_lock_bh(&stats->lock);
144                         if (!*used || time_after(stats->used, *used))
145                                 *used = stats->used;
146                         *tcp_flags |= stats->tcp_flags;
147                         ovs_stats->n_packets += stats->packet_count;
148                         ovs_stats->n_bytes += stats->byte_count;
149                         spin_unlock_bh(&stats->lock);
150                 }
151         }
152 }
153
154 void ovs_flow_stats_clear(struct sw_flow *flow)
155 {
156         int node;
157
158         for_each_node(node) {
159                 struct flow_stats *stats = rcu_dereference(flow->stats[node]);
160
161                 if (stats) {
162                         spin_lock_bh(&stats->lock);
163                         stats->used = 0;
164                         stats->packet_count = 0;
165                         stats->byte_count = 0;
166                         stats->tcp_flags = 0;
167                         spin_unlock_bh(&stats->lock);
168                 }
169         }
170 }
171
172 static int check_header(struct sk_buff *skb, int len)
173 {
174         if (unlikely(skb->len < len))
175                 return -EINVAL;
176         if (unlikely(!pskb_may_pull(skb, len)))
177                 return -ENOMEM;
178         return 0;
179 }
180
181 static bool arphdr_ok(struct sk_buff *skb)
182 {
183         return pskb_may_pull(skb, skb_network_offset(skb) +
184                                   sizeof(struct arp_eth_header));
185 }
186
187 static int check_iphdr(struct sk_buff *skb)
188 {
189         unsigned int nh_ofs = skb_network_offset(skb);
190         unsigned int ip_len;
191         int err;
192
193         err = check_header(skb, nh_ofs + sizeof(struct iphdr));
194         if (unlikely(err))
195                 return err;
196
197         ip_len = ip_hdrlen(skb);
198         if (unlikely(ip_len < sizeof(struct iphdr) ||
199                      skb->len < nh_ofs + ip_len))
200                 return -EINVAL;
201
202         skb_set_transport_header(skb, nh_ofs + ip_len);
203         return 0;
204 }
205
206 static bool tcphdr_ok(struct sk_buff *skb)
207 {
208         int th_ofs = skb_transport_offset(skb);
209         int tcp_len;
210
211         if (unlikely(!pskb_may_pull(skb, th_ofs + sizeof(struct tcphdr))))
212                 return false;
213
214         tcp_len = tcp_hdrlen(skb);
215         if (unlikely(tcp_len < sizeof(struct tcphdr) ||
216                      skb->len < th_ofs + tcp_len))
217                 return false;
218
219         return true;
220 }
221
222 static bool udphdr_ok(struct sk_buff *skb)
223 {
224         return pskb_may_pull(skb, skb_transport_offset(skb) +
225                                   sizeof(struct udphdr));
226 }
227
228 static bool sctphdr_ok(struct sk_buff *skb)
229 {
230         return pskb_may_pull(skb, skb_transport_offset(skb) +
231                                   sizeof(struct sctphdr));
232 }
233
234 static bool icmphdr_ok(struct sk_buff *skb)
235 {
236         return pskb_may_pull(skb, skb_transport_offset(skb) +
237                                   sizeof(struct icmphdr));
238 }
239
240 static int parse_ipv6hdr(struct sk_buff *skb, struct sw_flow_key *key)
241 {
242         unsigned int nh_ofs = skb_network_offset(skb);
243         unsigned int nh_len;
244         int payload_ofs;
245         struct ipv6hdr *nh;
246         uint8_t nexthdr;
247         __be16 frag_off;
248         int err;
249
250         err = check_header(skb, nh_ofs + sizeof(*nh));
251         if (unlikely(err))
252                 return err;
253
254         nh = ipv6_hdr(skb);
255         nexthdr = nh->nexthdr;
256         payload_ofs = (u8 *)(nh + 1) - skb->data;
257
258         key->ip.proto = NEXTHDR_NONE;
259         key->ip.tos = ipv6_get_dsfield(nh);
260         key->ip.ttl = nh->hop_limit;
261         key->ipv6.label = *(__be32 *)nh & htonl(IPV6_FLOWINFO_FLOWLABEL);
262         key->ipv6.addr.src = nh->saddr;
263         key->ipv6.addr.dst = nh->daddr;
264
265         payload_ofs = ipv6_skip_exthdr(skb, payload_ofs, &nexthdr, &frag_off);
266         if (unlikely(payload_ofs < 0))
267                 return -EINVAL;
268
269         if (frag_off) {
270                 if (frag_off & htons(~0x7))
271                         key->ip.frag = OVS_FRAG_TYPE_LATER;
272                 else
273                         key->ip.frag = OVS_FRAG_TYPE_FIRST;
274         }
275
276         nh_len = payload_ofs - nh_ofs;
277         skb_set_transport_header(skb, nh_ofs + nh_len);
278         key->ip.proto = nexthdr;
279         return nh_len;
280 }
281
282 static bool icmp6hdr_ok(struct sk_buff *skb)
283 {
284         return pskb_may_pull(skb, skb_transport_offset(skb) +
285                                   sizeof(struct icmp6hdr));
286 }
287
288 static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
289 {
290         struct qtag_prefix {
291                 __be16 eth_type; /* ETH_P_8021Q */
292                 __be16 tci;
293         };
294         struct qtag_prefix *qp;
295
296         if (unlikely(skb->len < sizeof(struct qtag_prefix) + sizeof(__be16)))
297                 return 0;
298
299         if (unlikely(!pskb_may_pull(skb, sizeof(struct qtag_prefix) +
300                                          sizeof(__be16))))
301                 return -ENOMEM;
302
303         qp = (struct qtag_prefix *) skb->data;
304         key->eth.tci = qp->tci | htons(VLAN_TAG_PRESENT);
305         __skb_pull(skb, sizeof(struct qtag_prefix));
306
307         return 0;
308 }
309
310 static __be16 parse_ethertype(struct sk_buff *skb)
311 {
312         struct llc_snap_hdr {
313                 u8  dsap;  /* Always 0xAA */
314                 u8  ssap;  /* Always 0xAA */
315                 u8  ctrl;
316                 u8  oui[3];
317                 __be16 ethertype;
318         };
319         struct llc_snap_hdr *llc;
320         __be16 proto;
321
322         proto = *(__be16 *) skb->data;
323         __skb_pull(skb, sizeof(__be16));
324
325         if (ntohs(proto) >= ETH_P_802_3_MIN)
326                 return proto;
327
328         if (skb->len < sizeof(struct llc_snap_hdr))
329                 return htons(ETH_P_802_2);
330
331         if (unlikely(!pskb_may_pull(skb, sizeof(struct llc_snap_hdr))))
332                 return htons(0);
333
334         llc = (struct llc_snap_hdr *) skb->data;
335         if (llc->dsap != LLC_SAP_SNAP ||
336             llc->ssap != LLC_SAP_SNAP ||
337             (llc->oui[0] | llc->oui[1] | llc->oui[2]) != 0)
338                 return htons(ETH_P_802_2);
339
340         __skb_pull(skb, sizeof(struct llc_snap_hdr));
341
342         if (ntohs(llc->ethertype) >= ETH_P_802_3_MIN)
343                 return llc->ethertype;
344
345         return htons(ETH_P_802_2);
346 }
347
348 static int parse_icmpv6(struct sk_buff *skb, struct sw_flow_key *key,
349                         int nh_len)
350 {
351         struct icmp6hdr *icmp = icmp6_hdr(skb);
352
353         /* The ICMPv6 type and code fields use the 16-bit transport port
354          * fields, so we need to store them in 16-bit network byte order.
355          */
356         key->tp.src = htons(icmp->icmp6_type);
357         key->tp.dst = htons(icmp->icmp6_code);
358
359         if (icmp->icmp6_code == 0 &&
360             (icmp->icmp6_type == NDISC_NEIGHBOUR_SOLICITATION ||
361              icmp->icmp6_type == NDISC_NEIGHBOUR_ADVERTISEMENT)) {
362                 int icmp_len = skb->len - skb_transport_offset(skb);
363                 struct nd_msg *nd;
364                 int offset;
365
366                 /* In order to process neighbor discovery options, we need the
367                  * entire packet.
368                  */
369                 if (unlikely(icmp_len < sizeof(*nd)))
370                         return 0;
371
372                 if (unlikely(skb_linearize(skb)))
373                         return -ENOMEM;
374
375                 nd = (struct nd_msg *)skb_transport_header(skb);
376                 key->ipv6.nd.target = nd->target;
377
378                 icmp_len -= sizeof(*nd);
379                 offset = 0;
380                 while (icmp_len >= 8) {
381                         struct nd_opt_hdr *nd_opt =
382                                  (struct nd_opt_hdr *)(nd->opt + offset);
383                         int opt_len = nd_opt->nd_opt_len * 8;
384
385                         if (unlikely(!opt_len || opt_len > icmp_len))
386                                 return 0;
387
388                         /* Store the link layer address if the appropriate
389                          * option is provided.  It is considered an error if
390                          * the same link layer option is specified twice.
391                          */
392                         if (nd_opt->nd_opt_type == ND_OPT_SOURCE_LL_ADDR
393                             && opt_len == 8) {
394                                 if (unlikely(!is_zero_ether_addr(key->ipv6.nd.sll)))
395                                         goto invalid;
396                                 ether_addr_copy(key->ipv6.nd.sll,
397                                     &nd->opt[offset+sizeof(*nd_opt)]);
398                         } else if (nd_opt->nd_opt_type == ND_OPT_TARGET_LL_ADDR
399                                    && opt_len == 8) {
400                                 if (unlikely(!is_zero_ether_addr(key->ipv6.nd.tll)))
401                                         goto invalid;
402                                 ether_addr_copy(key->ipv6.nd.tll,
403                                     &nd->opt[offset+sizeof(*nd_opt)]);
404                         }
405
406                         icmp_len -= opt_len;
407                         offset += opt_len;
408                 }
409         }
410
411         return 0;
412
413 invalid:
414         memset(&key->ipv6.nd.target, 0, sizeof(key->ipv6.nd.target));
415         memset(key->ipv6.nd.sll, 0, sizeof(key->ipv6.nd.sll));
416         memset(key->ipv6.nd.tll, 0, sizeof(key->ipv6.nd.tll));
417
418         return 0;
419 }
420
421 /**
422  * ovs_flow_extract - extracts a flow key from an Ethernet frame.
423  * @skb: sk_buff that contains the frame, with skb->data pointing to the
424  * Ethernet header
425  * @in_port: port number on which @skb was received.
426  * @key: output flow key
427  *
428  * The caller must ensure that skb->len >= ETH_HLEN.
429  *
430  * Returns 0 if successful, otherwise a negative errno value.
431  *
432  * Initializes @skb header pointers as follows:
433  *
434  *    - skb->mac_header: the Ethernet header.
435  *
436  *    - skb->network_header: just past the Ethernet header, or just past the
437  *      VLAN header, to the first byte of the Ethernet payload.
438  *
439  *    - skb->transport_header: If key->eth.type is ETH_P_IP or ETH_P_IPV6
440  *      on output, then just past the IP header, if one is present and
441  *      of a correct length, otherwise the same as skb->network_header.
442  *      For other key->eth.type values it is left untouched.
443  */
444 int ovs_flow_extract(struct sk_buff *skb, u16 in_port, struct sw_flow_key *key)
445 {
446         int error;
447         struct ethhdr *eth;
448
449         memset(key, 0, sizeof(*key));
450
451         key->phy.priority = skb->priority;
452         if (OVS_CB(skb)->tun_key)
453                 memcpy(&key->tun_key, OVS_CB(skb)->tun_key, sizeof(key->tun_key));
454         key->phy.in_port = in_port;
455         key->phy.skb_mark = skb->mark;
456
457         skb_reset_mac_header(skb);
458
459         /* Link layer.  We are guaranteed to have at least the 14 byte Ethernet
460          * header in the linear data area.
461          */
462         eth = eth_hdr(skb);
463         ether_addr_copy(key->eth.src, eth->h_source);
464         ether_addr_copy(key->eth.dst, eth->h_dest);
465
466         __skb_pull(skb, 2 * ETH_ALEN);
467         /* We are going to push all headers that we pull, so no need to
468          * update skb->csum here. */
469
470         if (vlan_tx_tag_present(skb))
471                 key->eth.tci = htons(vlan_get_tci(skb));
472         else if (eth->h_proto == htons(ETH_P_8021Q))
473                 if (unlikely(parse_vlan(skb, key)))
474                         return -ENOMEM;
475
476         key->eth.type = parse_ethertype(skb);
477         if (unlikely(key->eth.type == htons(0)))
478                 return -ENOMEM;
479
480         skb_reset_network_header(skb);
481         __skb_push(skb, skb->data - skb_mac_header(skb));
482
483         /* Network layer. */
484         if (key->eth.type == htons(ETH_P_IP)) {
485                 struct iphdr *nh;
486                 __be16 offset;
487
488                 error = check_iphdr(skb);
489                 if (unlikely(error)) {
490                         if (error == -EINVAL) {
491                                 skb->transport_header = skb->network_header;
492                                 error = 0;
493                         }
494                         return error;
495                 }
496
497                 nh = ip_hdr(skb);
498                 key->ipv4.addr.src = nh->saddr;
499                 key->ipv4.addr.dst = nh->daddr;
500
501                 key->ip.proto = nh->protocol;
502                 key->ip.tos = nh->tos;
503                 key->ip.ttl = nh->ttl;
504
505                 offset = nh->frag_off & htons(IP_OFFSET);
506                 if (offset) {
507                         key->ip.frag = OVS_FRAG_TYPE_LATER;
508                         return 0;
509                 }
510                 if (nh->frag_off & htons(IP_MF) ||
511                          skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
512                         key->ip.frag = OVS_FRAG_TYPE_FIRST;
513
514                 /* Transport layer. */
515                 if (key->ip.proto == IPPROTO_TCP) {
516                         if (tcphdr_ok(skb)) {
517                                 struct tcphdr *tcp = tcp_hdr(skb);
518                                 key->tp.src = tcp->source;
519                                 key->tp.dst = tcp->dest;
520                                 key->tp.flags = TCP_FLAGS_BE16(tcp);
521                         }
522                 } else if (key->ip.proto == IPPROTO_UDP) {
523                         if (udphdr_ok(skb)) {
524                                 struct udphdr *udp = udp_hdr(skb);
525                                 key->tp.src = udp->source;
526                                 key->tp.dst = udp->dest;
527                         }
528                 } else if (key->ip.proto == IPPROTO_SCTP) {
529                         if (sctphdr_ok(skb)) {
530                                 struct sctphdr *sctp = sctp_hdr(skb);
531                                 key->tp.src = sctp->source;
532                                 key->tp.dst = sctp->dest;
533                         }
534                 } else if (key->ip.proto == IPPROTO_ICMP) {
535                         if (icmphdr_ok(skb)) {
536                                 struct icmphdr *icmp = icmp_hdr(skb);
537                                 /* The ICMP type and code fields use the 16-bit
538                                  * transport port fields, so we need to store
539                                  * them in 16-bit network byte order. */
540                                 key->tp.src = htons(icmp->type);
541                                 key->tp.dst = htons(icmp->code);
542                         }
543                 }
544
545         } else if ((key->eth.type == htons(ETH_P_ARP) ||
546                    key->eth.type == htons(ETH_P_RARP)) && arphdr_ok(skb)) {
547                 struct arp_eth_header *arp;
548
549                 arp = (struct arp_eth_header *)skb_network_header(skb);
550
551                 if (arp->ar_hrd == htons(ARPHRD_ETHER)
552                                 && arp->ar_pro == htons(ETH_P_IP)
553                                 && arp->ar_hln == ETH_ALEN
554                                 && arp->ar_pln == 4) {
555
556                         /* We only match on the lower 8 bits of the opcode. */
557                         if (ntohs(arp->ar_op) <= 0xff)
558                                 key->ip.proto = ntohs(arp->ar_op);
559                         memcpy(&key->ipv4.addr.src, arp->ar_sip, sizeof(key->ipv4.addr.src));
560                         memcpy(&key->ipv4.addr.dst, arp->ar_tip, sizeof(key->ipv4.addr.dst));
561                         ether_addr_copy(key->ipv4.arp.sha, arp->ar_sha);
562                         ether_addr_copy(key->ipv4.arp.tha, arp->ar_tha);
563                 }
564         } else if (key->eth.type == htons(ETH_P_IPV6)) {
565                 int nh_len;             /* IPv6 Header + Extensions */
566
567                 nh_len = parse_ipv6hdr(skb, key);
568                 if (unlikely(nh_len < 0)) {
569                         if (nh_len == -EINVAL) {
570                                 skb->transport_header = skb->network_header;
571                                 error = 0;
572                         } else {
573                                 error = nh_len;
574                         }
575                         return error;
576                 }
577
578                 if (key->ip.frag == OVS_FRAG_TYPE_LATER)
579                         return 0;
580                 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
581                         key->ip.frag = OVS_FRAG_TYPE_FIRST;
582
583                 /* Transport layer. */
584                 if (key->ip.proto == NEXTHDR_TCP) {
585                         if (tcphdr_ok(skb)) {
586                                 struct tcphdr *tcp = tcp_hdr(skb);
587                                 key->tp.src = tcp->source;
588                                 key->tp.dst = tcp->dest;
589                                 key->tp.flags = TCP_FLAGS_BE16(tcp);
590                         }
591                 } else if (key->ip.proto == NEXTHDR_UDP) {
592                         if (udphdr_ok(skb)) {
593                                 struct udphdr *udp = udp_hdr(skb);
594                                 key->tp.src = udp->source;
595                                 key->tp.dst = udp->dest;
596                         }
597                 } else if (key->ip.proto == NEXTHDR_SCTP) {
598                         if (sctphdr_ok(skb)) {
599                                 struct sctphdr *sctp = sctp_hdr(skb);
600                                 key->tp.src = sctp->source;
601                                 key->tp.dst = sctp->dest;
602                         }
603                 } else if (key->ip.proto == NEXTHDR_ICMP) {
604                         if (icmp6hdr_ok(skb)) {
605                                 error = parse_icmpv6(skb, key, nh_len);
606                                 if (error)
607                                         return error;
608                         }
609                 }
610         }
611
612         return 0;
613 }