xref: /openbmc/linux/net/core/flow_dissector.c (revision 588b48ca)
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  */
21 static 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 }
52 EXPORT_SYMBOL(skb_flow_get_ports);
53 
54 bool 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 
62 again:
63 	switch (proto) {
64 	case htons(ETH_P_IP): {
65 		const struct iphdr *iph;
66 		struct iphdr _iph;
67 ip:
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 htons(ETH_P_IPV6): {
81 		const struct ipv6hdr *iph;
82 		struct ipv6hdr _iph;
83 		__be32 flow_label;
84 
85 ipv6:
86 		iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
87 		if (!iph)
88 			return false;
89 
90 		ip_proto = iph->nexthdr;
91 		flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr);
92 		flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr);
93 		nhoff += sizeof(struct ipv6hdr);
94 
95 		flow_label = ip6_flowlabel(iph);
96 		if (flow_label) {
97 			/* Awesome, IPv6 packet has a flow label so we can
98 			 * use that to represent the ports without any
99 			 * further dissection.
100 			 */
101 			flow->n_proto = proto;
102 			flow->ip_proto = ip_proto;
103 			flow->ports = flow_label;
104 			flow->thoff = (u16)nhoff;
105 
106 			return true;
107 		}
108 
109 		break;
110 	}
111 	case htons(ETH_P_8021AD):
112 	case htons(ETH_P_8021Q): {
113 		const struct vlan_hdr *vlan;
114 		struct vlan_hdr _vlan;
115 
116 		vlan = skb_header_pointer(skb, nhoff, sizeof(_vlan), &_vlan);
117 		if (!vlan)
118 			return false;
119 
120 		proto = vlan->h_vlan_encapsulated_proto;
121 		nhoff += sizeof(*vlan);
122 		goto again;
123 	}
124 	case htons(ETH_P_PPP_SES): {
125 		struct {
126 			struct pppoe_hdr hdr;
127 			__be16 proto;
128 		} *hdr, _hdr;
129 		hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr);
130 		if (!hdr)
131 			return false;
132 		proto = hdr->proto;
133 		nhoff += PPPOE_SES_HLEN;
134 		switch (proto) {
135 		case htons(PPP_IP):
136 			goto ip;
137 		case htons(PPP_IPV6):
138 			goto ipv6;
139 		default:
140 			return false;
141 		}
142 	}
143 	default:
144 		return false;
145 	}
146 
147 	switch (ip_proto) {
148 	case IPPROTO_GRE: {
149 		struct gre_hdr {
150 			__be16 flags;
151 			__be16 proto;
152 		} *hdr, _hdr;
153 
154 		hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr);
155 		if (!hdr)
156 			return false;
157 		/*
158 		 * Only look inside GRE if version zero and no
159 		 * routing
160 		 */
161 		if (!(hdr->flags & (GRE_VERSION|GRE_ROUTING))) {
162 			proto = hdr->proto;
163 			nhoff += 4;
164 			if (hdr->flags & GRE_CSUM)
165 				nhoff += 4;
166 			if (hdr->flags & GRE_KEY)
167 				nhoff += 4;
168 			if (hdr->flags & GRE_SEQ)
169 				nhoff += 4;
170 			if (proto == htons(ETH_P_TEB)) {
171 				const struct ethhdr *eth;
172 				struct ethhdr _eth;
173 
174 				eth = skb_header_pointer(skb, nhoff,
175 							 sizeof(_eth), &_eth);
176 				if (!eth)
177 					return false;
178 				proto = eth->h_proto;
179 				nhoff += sizeof(*eth);
180 			}
181 			goto again;
182 		}
183 		break;
184 	}
185 	case IPPROTO_IPIP:
186 		proto = htons(ETH_P_IP);
187 		goto ip;
188 	case IPPROTO_IPV6:
189 		proto = htons(ETH_P_IPV6);
190 		goto ipv6;
191 	default:
192 		break;
193 	}
194 
195 	flow->n_proto = proto;
196 	flow->ip_proto = ip_proto;
197 	flow->ports = skb_flow_get_ports(skb, nhoff, ip_proto);
198 	flow->thoff = (u16) nhoff;
199 
200 	return true;
201 }
202 EXPORT_SYMBOL(skb_flow_dissect);
203 
204 static u32 hashrnd __read_mostly;
205 static __always_inline void __flow_hash_secret_init(void)
206 {
207 	net_get_random_once(&hashrnd, sizeof(hashrnd));
208 }
209 
210 static __always_inline u32 __flow_hash_3words(u32 a, u32 b, u32 c)
211 {
212 	__flow_hash_secret_init();
213 	return jhash_3words(a, b, c, hashrnd);
214 }
215 
216 static inline u32 __flow_hash_from_keys(struct flow_keys *keys)
217 {
218 	u32 hash;
219 
220 	/* get a consistent hash (same value on both flow directions) */
221 	if (((__force u32)keys->dst < (__force u32)keys->src) ||
222 	    (((__force u32)keys->dst == (__force u32)keys->src) &&
223 	     ((__force u16)keys->port16[1] < (__force u16)keys->port16[0]))) {
224 		swap(keys->dst, keys->src);
225 		swap(keys->port16[0], keys->port16[1]);
226 	}
227 
228 	hash = __flow_hash_3words((__force u32)keys->dst,
229 				  (__force u32)keys->src,
230 				  (__force u32)keys->ports);
231 	if (!hash)
232 		hash = 1;
233 
234 	return hash;
235 }
236 
237 u32 flow_hash_from_keys(struct flow_keys *keys)
238 {
239 	return __flow_hash_from_keys(keys);
240 }
241 EXPORT_SYMBOL(flow_hash_from_keys);
242 
243 /*
244  * __skb_get_hash: calculate a flow hash based on src/dst addresses
245  * and src/dst port numbers.  Sets hash in skb to non-zero hash value
246  * on success, zero indicates no valid hash.  Also, sets l4_hash in skb
247  * if hash is a canonical 4-tuple hash over transport ports.
248  */
249 void __skb_get_hash(struct sk_buff *skb)
250 {
251 	struct flow_keys keys;
252 
253 	if (!skb_flow_dissect(skb, &keys))
254 		return;
255 
256 	if (keys.ports)
257 		skb->l4_hash = 1;
258 
259 	skb->sw_hash = 1;
260 
261 	skb->hash = __flow_hash_from_keys(&keys);
262 }
263 EXPORT_SYMBOL(__skb_get_hash);
264 
265 /*
266  * Returns a Tx hash based on the given packet descriptor a Tx queues' number
267  * to be used as a distribution range.
268  */
269 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
270 		  unsigned int num_tx_queues)
271 {
272 	u32 hash;
273 	u16 qoffset = 0;
274 	u16 qcount = num_tx_queues;
275 
276 	if (skb_rx_queue_recorded(skb)) {
277 		hash = skb_get_rx_queue(skb);
278 		while (unlikely(hash >= num_tx_queues))
279 			hash -= num_tx_queues;
280 		return hash;
281 	}
282 
283 	if (dev->num_tc) {
284 		u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
285 		qoffset = dev->tc_to_txq[tc].offset;
286 		qcount = dev->tc_to_txq[tc].count;
287 	}
288 
289 	return (u16) (((u64)skb_get_hash(skb) * qcount) >> 32) + qoffset;
290 }
291 EXPORT_SYMBOL(__skb_tx_hash);
292 
293 /* __skb_get_poff() returns the offset to the payload as far as it could
294  * be dissected. The main user is currently BPF, so that we can dynamically
295  * truncate packets without needing to push actual payload to the user
296  * space and can analyze headers only, instead.
297  */
298 u32 __skb_get_poff(const struct sk_buff *skb)
299 {
300 	struct flow_keys keys;
301 	u32 poff = 0;
302 
303 	if (!skb_flow_dissect(skb, &keys))
304 		return 0;
305 
306 	poff += keys.thoff;
307 	switch (keys.ip_proto) {
308 	case IPPROTO_TCP: {
309 		const struct tcphdr *tcph;
310 		struct tcphdr _tcph;
311 
312 		tcph = skb_header_pointer(skb, poff, sizeof(_tcph), &_tcph);
313 		if (!tcph)
314 			return poff;
315 
316 		poff += max_t(u32, sizeof(struct tcphdr), tcph->doff * 4);
317 		break;
318 	}
319 	case IPPROTO_UDP:
320 	case IPPROTO_UDPLITE:
321 		poff += sizeof(struct udphdr);
322 		break;
323 	/* For the rest, we do not really care about header
324 	 * extensions at this point for now.
325 	 */
326 	case IPPROTO_ICMP:
327 		poff += sizeof(struct icmphdr);
328 		break;
329 	case IPPROTO_ICMPV6:
330 		poff += sizeof(struct icmp6hdr);
331 		break;
332 	case IPPROTO_IGMP:
333 		poff += sizeof(struct igmphdr);
334 		break;
335 	case IPPROTO_DCCP:
336 		poff += sizeof(struct dccp_hdr);
337 		break;
338 	case IPPROTO_SCTP:
339 		poff += sizeof(struct sctphdr);
340 		break;
341 	}
342 
343 	return poff;
344 }
345 
346 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
347 {
348 #ifdef CONFIG_XPS
349 	struct xps_dev_maps *dev_maps;
350 	struct xps_map *map;
351 	int queue_index = -1;
352 
353 	rcu_read_lock();
354 	dev_maps = rcu_dereference(dev->xps_maps);
355 	if (dev_maps) {
356 		map = rcu_dereference(
357 		    dev_maps->cpu_map[raw_smp_processor_id()]);
358 		if (map) {
359 			if (map->len == 1)
360 				queue_index = map->queues[0];
361 			else
362 				queue_index = map->queues[
363 				    ((u64)skb_get_hash(skb) * map->len) >> 32];
364 
365 			if (unlikely(queue_index >= dev->real_num_tx_queues))
366 				queue_index = -1;
367 		}
368 	}
369 	rcu_read_unlock();
370 
371 	return queue_index;
372 #else
373 	return -1;
374 #endif
375 }
376 
377 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
378 {
379 	struct sock *sk = skb->sk;
380 	int queue_index = sk_tx_queue_get(sk);
381 
382 	if (queue_index < 0 || skb->ooo_okay ||
383 	    queue_index >= dev->real_num_tx_queues) {
384 		int new_index = get_xps_queue(dev, skb);
385 		if (new_index < 0)
386 			new_index = skb_tx_hash(dev, skb);
387 
388 		if (queue_index != new_index && sk &&
389 		    rcu_access_pointer(sk->sk_dst_cache))
390 			sk_tx_queue_set(sk, new_index);
391 
392 		queue_index = new_index;
393 	}
394 
395 	return queue_index;
396 }
397 
398 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
399 				    struct sk_buff *skb,
400 				    void *accel_priv)
401 {
402 	int queue_index = 0;
403 
404 	if (dev->real_num_tx_queues != 1) {
405 		const struct net_device_ops *ops = dev->netdev_ops;
406 		if (ops->ndo_select_queue)
407 			queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
408 							    __netdev_pick_tx);
409 		else
410 			queue_index = __netdev_pick_tx(dev, skb);
411 
412 		if (!accel_priv)
413 			queue_index = netdev_cap_txqueue(dev, queue_index);
414 	}
415 
416 	skb_set_queue_mapping(skb, queue_index);
417 	return netdev_get_tx_queue(dev, queue_index);
418 }
419