flow_dissector.c revision b9507bdaf40e91fea2b1c0c1ee7dc627c8ee6fd6 
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
17/* copy saddr & daddr, possibly using 64bit load/store
18 * Equivalent to :	flow->src = iph->saddr;
19 *			flow->dst = iph->daddr;
20 */
21static void iph_to_flow_copy_addrs(struct flow_keys *flow, const struct iphdr *iph)
22{
23	BUILD_BUG_ON(offsetof(typeof(*flow), dst) !=
24		     offsetof(typeof(*flow), src) + sizeof(flow->src));
25	memcpy(&flow->src, &iph->saddr, sizeof(flow->src) + sizeof(flow->dst));
26}
27
28/**
29 * skb_flow_get_ports - extract the upper layer ports and return them
30 * @skb: buffer to extract the ports from
31 * @thoff: transport header offset
32 * @ip_proto: protocol for which to get port offset
33 *
34 * The function will try to retrieve the ports at offset thoff + poff where poff
35 * is the protocol port offset returned from proto_ports_offset
36 */
37__be32 skb_flow_get_ports(const struct sk_buff *skb, int thoff, u8 ip_proto)
38{
39	int poff = proto_ports_offset(ip_proto);
40
41	if (poff >= 0) {
42		__be32 *ports, _ports;
43
44		ports = skb_header_pointer(skb, thoff + poff,
45					   sizeof(_ports), &_ports);
46		if (ports)
47			return *ports;
48	}
49
50	return 0;
51}
52EXPORT_SYMBOL(skb_flow_get_ports);
53
54bool skb_flow_dissect(const struct sk_buff *skb, struct flow_keys *flow)
55{
56	int nhoff = skb_network_offset(skb);
57	u8 ip_proto;
58	__be16 proto = skb->protocol;
59
60	memset(flow, 0, sizeof(*flow));
61
62again:
63	switch (proto) {
64	case __constant_htons(ETH_P_IP): {
65		const struct iphdr *iph;
66		struct iphdr _iph;
67ip:
68		iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
69		if (!iph || iph->ihl < 5)
70			return false;
71		nhoff += iph->ihl * 4;
72
73		ip_proto = iph->protocol;
74		if (ip_is_fragment(iph))
75			ip_proto = 0;
76
77		iph_to_flow_copy_addrs(flow, iph);
78		break;
79	}
80	case __constant_htons(ETH_P_IPV6): {
81		const struct ipv6hdr *iph;
82		struct ipv6hdr _iph;
83ipv6:
84		iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
85		if (!iph)
86			return false;
87
88		ip_proto = iph->nexthdr;
89		flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr);
90		flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr);
91		nhoff += sizeof(struct ipv6hdr);
92		break;
93	}
94	case __constant_htons(ETH_P_8021AD):
95	case __constant_htons(ETH_P_8021Q): {
96		const struct vlan_hdr *vlan;
97		struct vlan_hdr _vlan;
98
99		vlan = skb_header_pointer(skb, nhoff, sizeof(_vlan), &_vlan);
100		if (!vlan)
101			return false;
102
103		proto = vlan->h_vlan_encapsulated_proto;
104		nhoff += sizeof(*vlan);
105		goto again;
106	}
107	case __constant_htons(ETH_P_PPP_SES): {
108		struct {
109			struct pppoe_hdr hdr;
110			__be16 proto;
111		} *hdr, _hdr;
112		hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr);
113		if (!hdr)
114			return false;
115		proto = hdr->proto;
116		nhoff += PPPOE_SES_HLEN;
117		switch (proto) {
118		case __constant_htons(PPP_IP):
119			goto ip;
120		case __constant_htons(PPP_IPV6):
121			goto ipv6;
122		default:
123			return false;
124		}
125	}
126	default:
127		return false;
128	}
129
130	switch (ip_proto) {
131	case IPPROTO_GRE: {
132		struct gre_hdr {
133			__be16 flags;
134			__be16 proto;
135		} *hdr, _hdr;
136
137		hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr);
138		if (!hdr)
139			return false;
140		/*
141		 * Only look inside GRE if version zero and no
142		 * routing
143		 */
144		if (!(hdr->flags & (GRE_VERSION|GRE_ROUTING))) {
145			proto = hdr->proto;
146			nhoff += 4;
147			if (hdr->flags & GRE_CSUM)
148				nhoff += 4;
149			if (hdr->flags & GRE_KEY)
150				nhoff += 4;
151			if (hdr->flags & GRE_SEQ)
152				nhoff += 4;
153			if (proto == htons(ETH_P_TEB)) {
154				const struct ethhdr *eth;
155				struct ethhdr _eth;
156
157				eth = skb_header_pointer(skb, nhoff,
158							 sizeof(_eth), &_eth);
159				if (!eth)
160					return false;
161				proto = eth->h_proto;
162				nhoff += sizeof(*eth);
163			}
164			goto again;
165		}
166		break;
167	}
168	case IPPROTO_IPIP:
169		proto = htons(ETH_P_IP);
170		goto ip;
171	case IPPROTO_IPV6:
172		proto = htons(ETH_P_IPV6);
173		goto ipv6;
174	default:
175		break;
176	}
177
178	flow->ip_proto = ip_proto;
179	flow->ports = skb_flow_get_ports(skb, nhoff, ip_proto);
180	flow->thoff = (u16) nhoff;
181
182	return true;
183}
184EXPORT_SYMBOL(skb_flow_dissect);
185
186static u32 hashrnd __read_mostly;
187static __always_inline void __flow_hash_secret_init(void)
188{
189	net_get_random_once(&hashrnd, sizeof(hashrnd));
190}
191
192static __always_inline u32 __flow_hash_3words(u32 a, u32 b, u32 c)
193{
194	__flow_hash_secret_init();
195	return jhash_3words(a, b, c, hashrnd);
196}
197
198static __always_inline u32 __flow_hash_1word(u32 a)
199{
200	__flow_hash_secret_init();
201	return jhash_1word(a, hashrnd);
202}
203
204/*
205 * __skb_get_hash: calculate a flow hash based on src/dst addresses
206 * and src/dst port numbers.  Sets rxhash in skb to non-zero hash value
207 * on success, zero indicates no valid hash.  Also, sets l4_rxhash in skb
208 * if hash is a canonical 4-tuple hash over transport ports.
209 */
210void __skb_get_hash(struct sk_buff *skb)
211{
212	struct flow_keys keys;
213	u32 hash;
214
215	if (!skb_flow_dissect(skb, &keys))
216		return;
217
218	if (keys.ports)
219		skb->l4_rxhash = 1;
220
221	/* get a consistent hash (same value on both flow directions) */
222	if (((__force u32)keys.dst < (__force u32)keys.src) ||
223	    (((__force u32)keys.dst == (__force u32)keys.src) &&
224	     ((__force u16)keys.port16[1] < (__force u16)keys.port16[0]))) {
225		swap(keys.dst, keys.src);
226		swap(keys.port16[0], keys.port16[1]);
227	}
228
229	hash = __flow_hash_3words((__force u32)keys.dst,
230				  (__force u32)keys.src,
231				  (__force u32)keys.ports);
232	if (!hash)
233		hash = 1;
234
235	skb->rxhash = hash;
236}
237EXPORT_SYMBOL(__skb_get_hash);
238
239/*
240 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
241 * to be used as a distribution range.
242 */
243u16 __skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb,
244		  unsigned int num_tx_queues)
245{
246	u32 hash;
247	u16 qoffset = 0;
248	u16 qcount = num_tx_queues;
249
250	if (skb_rx_queue_recorded(skb)) {
251		hash = skb_get_rx_queue(skb);
252		while (unlikely(hash >= num_tx_queues))
253			hash -= num_tx_queues;
254		return hash;
255	}
256
257	if (dev->num_tc) {
258		u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
259		qoffset = dev->tc_to_txq[tc].offset;
260		qcount = dev->tc_to_txq[tc].count;
261	}
262
263	if (skb->sk && skb->sk->sk_hash)
264		hash = skb->sk->sk_hash;
265	else
266		hash = (__force u16) skb->protocol;
267	hash = __flow_hash_1word(hash);
268
269	return (u16) (((u64) hash * qcount) >> 32) + qoffset;
270}
271EXPORT_SYMBOL(__skb_tx_hash);
272
273/* __skb_get_poff() returns the offset to the payload as far as it could
274 * be dissected. The main user is currently BPF, so that we can dynamically
275 * truncate packets without needing to push actual payload to the user
276 * space and can analyze headers only, instead.
277 */
278u32 __skb_get_poff(const struct sk_buff *skb)
279{
280	struct flow_keys keys;
281	u32 poff = 0;
282
283	if (!skb_flow_dissect(skb, &keys))
284		return 0;
285
286	poff += keys.thoff;
287	switch (keys.ip_proto) {
288	case IPPROTO_TCP: {
289		const struct tcphdr *tcph;
290		struct tcphdr _tcph;
291
292		tcph = skb_header_pointer(skb, poff, sizeof(_tcph), &_tcph);
293		if (!tcph)
294			return poff;
295
296		poff += max_t(u32, sizeof(struct tcphdr), tcph->doff * 4);
297		break;
298	}
299	case IPPROTO_UDP:
300	case IPPROTO_UDPLITE:
301		poff += sizeof(struct udphdr);
302		break;
303	/* For the rest, we do not really care about header
304	 * extensions at this point for now.
305	 */
306	case IPPROTO_ICMP:
307		poff += sizeof(struct icmphdr);
308		break;
309	case IPPROTO_ICMPV6:
310		poff += sizeof(struct icmp6hdr);
311		break;
312	case IPPROTO_IGMP:
313		poff += sizeof(struct igmphdr);
314		break;
315	case IPPROTO_DCCP:
316		poff += sizeof(struct dccp_hdr);
317		break;
318	case IPPROTO_SCTP:
319		poff += sizeof(struct sctphdr);
320		break;
321	}
322
323	return poff;
324}
325
326static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
327{
328#ifdef CONFIG_XPS
329	struct xps_dev_maps *dev_maps;
330	struct xps_map *map;
331	int queue_index = -1;
332
333	rcu_read_lock();
334	dev_maps = rcu_dereference(dev->xps_maps);
335	if (dev_maps) {
336		map = rcu_dereference(
337		    dev_maps->cpu_map[raw_smp_processor_id()]);
338		if (map) {
339			if (map->len == 1)
340				queue_index = map->queues[0];
341			else {
342				u32 hash;
343				if (skb->sk && skb->sk->sk_hash)
344					hash = skb->sk->sk_hash;
345				else
346					hash = (__force u16) skb->protocol ^
347					    skb->rxhash;
348				hash = __flow_hash_1word(hash);
349				queue_index = map->queues[
350				    ((u64)hash * map->len) >> 32];
351			}
352			if (unlikely(queue_index >= dev->real_num_tx_queues))
353				queue_index = -1;
354		}
355	}
356	rcu_read_unlock();
357
358	return queue_index;
359#else
360	return -1;
361#endif
362}
363
364static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
365{
366	struct sock *sk = skb->sk;
367	int queue_index = sk_tx_queue_get(sk);
368
369	if (queue_index < 0 || skb->ooo_okay ||
370	    queue_index >= dev->real_num_tx_queues) {
371		int new_index = get_xps_queue(dev, skb);
372		if (new_index < 0)
373			new_index = skb_tx_hash(dev, skb);
374
375		if (queue_index != new_index && sk &&
376		    rcu_access_pointer(sk->sk_dst_cache))
377			sk_tx_queue_set(sk, new_index);
378
379		queue_index = new_index;
380	}
381
382	return queue_index;
383}
384
385struct netdev_queue *netdev_pick_tx(struct net_device *dev,
386				    struct sk_buff *skb,
387				    void *accel_priv)
388{
389	int queue_index = 0;
390
391	if (dev->real_num_tx_queues != 1) {
392		const struct net_device_ops *ops = dev->netdev_ops;
393		if (ops->ndo_select_queue)
394			queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
395							    __netdev_pick_tx);
396		else
397			queue_index = __netdev_pick_tx(dev, skb);
398
399		if (!accel_priv)
400			queue_index = netdev_cap_txqueue(dev, queue_index);
401	}
402
403	skb_set_queue_mapping(skb, queue_index);
404	return netdev_get_tx_queue(dev, queue_index);
405}
406