Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[pandora-kernel.git] / net / core / flow_dissector.c
1 #include <linux/skbuff.h>
2 #include <linux/export.h>
3 #include <linux/ip.h>
4 #include <linux/ipv6.h>
5 #include <linux/if_vlan.h>
6 #include <net/ip.h>
7 #include <net/ipv6.h>
8 #include <linux/igmp.h>
9 #include <linux/icmp.h>
10 #include <linux/sctp.h>
11 #include <linux/dccp.h>
12 #include <linux/if_tunnel.h>
13 #include <linux/if_pppox.h>
14 #include <linux/ppp_defs.h>
15 #include <net/flow_keys.h>
16 #include <scsi/fc/fc_fcoe.h>
17
18 /* copy saddr & daddr, possibly using 64bit load/store
19  * Equivalent to :      flow->src = iph->saddr;
20  *                      flow->dst = iph->daddr;
21  */
22 static void iph_to_flow_copy_addrs(struct flow_keys *flow, const struct iphdr *iph)
23 {
24         BUILD_BUG_ON(offsetof(typeof(*flow), dst) !=
25                      offsetof(typeof(*flow), src) + sizeof(flow->src));
26         memcpy(&flow->src, &iph->saddr, sizeof(flow->src) + sizeof(flow->dst));
27 }
28
29 /**
30  * __skb_flow_get_ports - extract the upper layer ports and return them
31  * @skb: sk_buff to extract the ports from
32  * @thoff: transport header offset
33  * @ip_proto: protocol for which to get port offset
34  * @data: raw buffer pointer to the packet, if NULL use skb->data
35  * @hlen: packet header length, if @data is NULL use skb_headlen(skb)
36  *
37  * The function will try to retrieve the ports at offset thoff + poff where poff
38  * is the protocol port offset returned from proto_ports_offset
39  */
40 __be32 __skb_flow_get_ports(const struct sk_buff *skb, int thoff, u8 ip_proto,
41                             void *data, int hlen)
42 {
43         int poff = proto_ports_offset(ip_proto);
44
45         if (!data) {
46                 data = skb->data;
47                 hlen = skb_headlen(skb);
48         }
49
50         if (poff >= 0) {
51                 __be32 *ports, _ports;
52
53                 ports = __skb_header_pointer(skb, thoff + poff,
54                                              sizeof(_ports), data, hlen, &_ports);
55                 if (ports)
56                         return *ports;
57         }
58
59         return 0;
60 }
61 EXPORT_SYMBOL(__skb_flow_get_ports);
62
63 /**
64  * __skb_flow_dissect - extract the flow_keys struct and return it
65  * @skb: sk_buff to extract the flow from, can be NULL if the rest are specified
66  * @data: raw buffer pointer to the packet, if NULL use skb->data
67  * @proto: protocol for which to get the flow, if @data is NULL use skb->protocol
68  * @nhoff: network header offset, if @data is NULL use skb_network_offset(skb)
69  * @hlen: packet header length, if @data is NULL use skb_headlen(skb)
70  *
71  * The function will try to retrieve the struct flow_keys from either the skbuff
72  * or a raw buffer specified by the rest parameters
73  */
74 bool __skb_flow_dissect(const struct sk_buff *skb, struct flow_keys *flow,
75                         void *data, __be16 proto, int nhoff, int hlen)
76 {
77         u8 ip_proto;
78
79         if (!data) {
80                 data = skb->data;
81                 proto = skb->protocol;
82                 nhoff = skb_network_offset(skb);
83                 hlen = skb_headlen(skb);
84         }
85
86         memset(flow, 0, sizeof(*flow));
87
88 again:
89         switch (proto) {
90         case htons(ETH_P_IP): {
91                 const struct iphdr *iph;
92                 struct iphdr _iph;
93 ip:
94                 iph = __skb_header_pointer(skb, nhoff, sizeof(_iph), data, hlen, &_iph);
95                 if (!iph || iph->ihl < 5)
96                         return false;
97                 nhoff += iph->ihl * 4;
98
99                 ip_proto = iph->protocol;
100                 if (ip_is_fragment(iph))
101                         ip_proto = 0;
102
103                 /* skip the address processing if skb is NULL.  The assumption
104                  * here is that if there is no skb we are not looking for flow
105                  * info but lengths and protocols.
106                  */
107                 if (!skb)
108                         break;
109
110                 iph_to_flow_copy_addrs(flow, iph);
111                 break;
112         }
113         case htons(ETH_P_IPV6): {
114                 const struct ipv6hdr *iph;
115                 struct ipv6hdr _iph;
116                 __be32 flow_label;
117
118 ipv6:
119                 iph = __skb_header_pointer(skb, nhoff, sizeof(_iph), data, hlen, &_iph);
120                 if (!iph)
121                         return false;
122
123                 ip_proto = iph->nexthdr;
124                 nhoff += sizeof(struct ipv6hdr);
125
126                 /* see comment above in IPv4 section */
127                 if (!skb)
128                         break;
129
130                 flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr);
131                 flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr);
132
133                 flow_label = ip6_flowlabel(iph);
134                 if (flow_label) {
135                         /* Awesome, IPv6 packet has a flow label so we can
136                          * use that to represent the ports without any
137                          * further dissection.
138                          */
139                         flow->n_proto = proto;
140                         flow->ip_proto = ip_proto;
141                         flow->ports = flow_label;
142                         flow->thoff = (u16)nhoff;
143
144                         return true;
145                 }
146
147                 break;
148         }
149         case htons(ETH_P_8021AD):
150         case htons(ETH_P_8021Q): {
151                 const struct vlan_hdr *vlan;
152                 struct vlan_hdr _vlan;
153
154                 vlan = __skb_header_pointer(skb, nhoff, sizeof(_vlan), data, hlen, &_vlan);
155                 if (!vlan)
156                         return false;
157
158                 proto = vlan->h_vlan_encapsulated_proto;
159                 nhoff += sizeof(*vlan);
160                 goto again;
161         }
162         case htons(ETH_P_PPP_SES): {
163                 struct {
164                         struct pppoe_hdr hdr;
165                         __be16 proto;
166                 } *hdr, _hdr;
167                 hdr = __skb_header_pointer(skb, nhoff, sizeof(_hdr), data, hlen, &_hdr);
168                 if (!hdr)
169                         return false;
170                 proto = hdr->proto;
171                 nhoff += PPPOE_SES_HLEN;
172                 switch (proto) {
173                 case htons(PPP_IP):
174                         goto ip;
175                 case htons(PPP_IPV6):
176                         goto ipv6;
177                 default:
178                         return false;
179                 }
180         }
181         case htons(ETH_P_FCOE):
182                 flow->thoff = (u16)(nhoff + FCOE_HEADER_LEN);
183                 /* fall through */
184         default:
185                 return false;
186         }
187
188         switch (ip_proto) {
189         case IPPROTO_GRE: {
190                 struct gre_hdr {
191                         __be16 flags;
192                         __be16 proto;
193                 } *hdr, _hdr;
194
195                 hdr = __skb_header_pointer(skb, nhoff, sizeof(_hdr), data, hlen, &_hdr);
196                 if (!hdr)
197                         return false;
198                 /*
199                  * Only look inside GRE if version zero and no
200                  * routing
201                  */
202                 if (!(hdr->flags & (GRE_VERSION|GRE_ROUTING))) {
203                         proto = hdr->proto;
204                         nhoff += 4;
205                         if (hdr->flags & GRE_CSUM)
206                                 nhoff += 4;
207                         if (hdr->flags & GRE_KEY)
208                                 nhoff += 4;
209                         if (hdr->flags & GRE_SEQ)
210                                 nhoff += 4;
211                         if (proto == htons(ETH_P_TEB)) {
212                                 const struct ethhdr *eth;
213                                 struct ethhdr _eth;
214
215                                 eth = __skb_header_pointer(skb, nhoff,
216                                                            sizeof(_eth),
217                                                            data, hlen, &_eth);
218                                 if (!eth)
219                                         return false;
220                                 proto = eth->h_proto;
221                                 nhoff += sizeof(*eth);
222                         }
223                         goto again;
224                 }
225                 break;
226         }
227         case IPPROTO_IPIP:
228                 proto = htons(ETH_P_IP);
229                 goto ip;
230         case IPPROTO_IPV6:
231                 proto = htons(ETH_P_IPV6);
232                 goto ipv6;
233         default:
234                 break;
235         }
236
237         flow->n_proto = proto;
238         flow->ip_proto = ip_proto;
239         flow->thoff = (u16) nhoff;
240
241         /* unless skb is set we don't need to record port info */
242         if (skb)
243                 flow->ports = __skb_flow_get_ports(skb, nhoff, ip_proto,
244                                                    data, hlen);
245
246         return true;
247 }
248 EXPORT_SYMBOL(__skb_flow_dissect);
249
250 static u32 hashrnd __read_mostly;
251 static __always_inline void __flow_hash_secret_init(void)
252 {
253         net_get_random_once(&hashrnd, sizeof(hashrnd));
254 }
255
256 static __always_inline u32 __flow_hash_3words(u32 a, u32 b, u32 c)
257 {
258         __flow_hash_secret_init();
259         return jhash_3words(a, b, c, hashrnd);
260 }
261
262 static inline u32 __flow_hash_from_keys(struct flow_keys *keys)
263 {
264         u32 hash;
265
266         /* get a consistent hash (same value on both flow directions) */
267         if (((__force u32)keys->dst < (__force u32)keys->src) ||
268             (((__force u32)keys->dst == (__force u32)keys->src) &&
269              ((__force u16)keys->port16[1] < (__force u16)keys->port16[0]))) {
270                 swap(keys->dst, keys->src);
271                 swap(keys->port16[0], keys->port16[1]);
272         }
273
274         hash = __flow_hash_3words((__force u32)keys->dst,
275                                   (__force u32)keys->src,
276                                   (__force u32)keys->ports);
277         if (!hash)
278                 hash = 1;
279
280         return hash;
281 }
282
283 u32 flow_hash_from_keys(struct flow_keys *keys)
284 {
285         return __flow_hash_from_keys(keys);
286 }
287 EXPORT_SYMBOL(flow_hash_from_keys);
288
289 /*
290  * __skb_get_hash: calculate a flow hash based on src/dst addresses
291  * and src/dst port numbers.  Sets hash in skb to non-zero hash value
292  * on success, zero indicates no valid hash.  Also, sets l4_hash in skb
293  * if hash is a canonical 4-tuple hash over transport ports.
294  */
295 void __skb_get_hash(struct sk_buff *skb)
296 {
297         struct flow_keys keys;
298
299         if (!skb_flow_dissect(skb, &keys))
300                 return;
301
302         if (keys.ports)
303                 skb->l4_hash = 1;
304
305         skb->sw_hash = 1;
306
307         skb->hash = __flow_hash_from_keys(&keys);
308 }
309 EXPORT_SYMBOL(__skb_get_hash);
310
311 /*
312  * Returns a Tx hash based on the given packet descriptor a Tx queues' number
313  * to be used as a distribution range.
314  */
315 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
316                   unsigned int num_tx_queues)
317 {
318         u32 hash;
319         u16 qoffset = 0;
320         u16 qcount = num_tx_queues;
321
322         if (skb_rx_queue_recorded(skb)) {
323                 hash = skb_get_rx_queue(skb);
324                 while (unlikely(hash >= num_tx_queues))
325                         hash -= num_tx_queues;
326                 return hash;
327         }
328
329         if (dev->num_tc) {
330                 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
331                 qoffset = dev->tc_to_txq[tc].offset;
332                 qcount = dev->tc_to_txq[tc].count;
333         }
334
335         return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
336 }
337 EXPORT_SYMBOL(__skb_tx_hash);
338
339 u32 __skb_get_poff(const struct sk_buff *skb, void *data,
340                    const struct flow_keys *keys, int hlen)
341 {
342         u32 poff = keys->thoff;
343
344         switch (keys->ip_proto) {
345         case IPPROTO_TCP: {
346                 /* access doff as u8 to avoid unaligned access */
347                 const u8 *doff;
348                 u8 _doff;
349
350                 doff = __skb_header_pointer(skb, poff + 12, sizeof(_doff),
351                                             data, hlen, &_doff);
352                 if (!doff)
353                         return poff;
354
355                 poff += max_t(u32, sizeof(struct tcphdr), (*doff & 0xF0) >> 2);
356                 break;
357         }
358         case IPPROTO_UDP:
359         case IPPROTO_UDPLITE:
360                 poff += sizeof(struct udphdr);
361                 break;
362         /* For the rest, we do not really care about header
363          * extensions at this point for now.
364          */
365         case IPPROTO_ICMP:
366                 poff += sizeof(struct icmphdr);
367                 break;
368         case IPPROTO_ICMPV6:
369                 poff += sizeof(struct icmp6hdr);
370                 break;
371         case IPPROTO_IGMP:
372                 poff += sizeof(struct igmphdr);
373                 break;
374         case IPPROTO_DCCP:
375                 poff += sizeof(struct dccp_hdr);
376                 break;
377         case IPPROTO_SCTP:
378                 poff += sizeof(struct sctphdr);
379                 break;
380         }
381
382         return poff;
383 }
384
385 /* skb_get_poff() returns the offset to the payload as far as it could
386  * be dissected. The main user is currently BPF, so that we can dynamically
387  * truncate packets without needing to push actual payload to the user
388  * space and can analyze headers only, instead.
389  */
390 u32 skb_get_poff(const struct sk_buff *skb)
391 {
392         struct flow_keys keys;
393
394         if (!skb_flow_dissect(skb, &keys))
395                 return 0;
396
397         return __skb_get_poff(skb, skb->data, &keys, skb_headlen(skb));
398 }
399
400 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
401 {
402 #ifdef CONFIG_XPS
403         struct xps_dev_maps *dev_maps;
404         struct xps_map *map;
405         int queue_index = -1;
406
407         rcu_read_lock();
408         dev_maps = rcu_dereference(dev->xps_maps);
409         if (dev_maps) {
410                 map = rcu_dereference(
411                     dev_maps->cpu_map[raw_smp_processor_id()]);
412                 if (map) {
413                         if (map->len == 1)
414                                 queue_index = map->queues[0];
415                         else
416                                 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
417                                                                            map->len)];
418                         if (unlikely(queue_index >= dev->real_num_tx_queues))
419                                 queue_index = -1;
420                 }
421         }
422         rcu_read_unlock();
423
424         return queue_index;
425 #else
426         return -1;
427 #endif
428 }
429
430 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
431 {
432         struct sock *sk = skb->sk;
433         int queue_index = sk_tx_queue_get(sk);
434
435         if (queue_index < 0 || skb->ooo_okay ||
436             queue_index >= dev->real_num_tx_queues) {
437                 int new_index = get_xps_queue(dev, skb);
438                 if (new_index < 0)
439                         new_index = skb_tx_hash(dev, skb);
440
441                 if (queue_index != new_index && sk &&
442                     rcu_access_pointer(sk->sk_dst_cache))
443                         sk_tx_queue_set(sk, new_index);
444
445                 queue_index = new_index;
446         }
447
448         return queue_index;
449 }
450
451 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
452                                     struct sk_buff *skb,
453                                     void *accel_priv)
454 {
455         int queue_index = 0;
456
457         if (dev->real_num_tx_queues != 1) {
458                 const struct net_device_ops *ops = dev->netdev_ops;
459                 if (ops->ndo_select_queue)
460                         queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
461                                                             __netdev_pick_tx);
462                 else
463                         queue_index = __netdev_pick_tx(dev, skb);
464
465                 if (!accel_priv)
466                         queue_index = netdev_cap_txqueue(dev, queue_index);
467         }
468
469         skb_set_queue_mapping(skb, queue_index);
470         return netdev_get_tx_queue(dev, queue_index);
471 }