xref: /openbmc/linux/net/core/flow_dissector.c (revision 9d749629)
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/if_tunnel.h>
9 #include <linux/if_pppox.h>
10 #include <linux/ppp_defs.h>
11 #include <net/flow_keys.h>
12 
13 /* copy saddr & daddr, possibly using 64bit load/store
14  * Equivalent to :	flow->src = iph->saddr;
15  *			flow->dst = iph->daddr;
16  */
17 static void iph_to_flow_copy_addrs(struct flow_keys *flow, const struct iphdr *iph)
18 {
19 	BUILD_BUG_ON(offsetof(typeof(*flow), dst) !=
20 		     offsetof(typeof(*flow), src) + sizeof(flow->src));
21 	memcpy(&flow->src, &iph->saddr, sizeof(flow->src) + sizeof(flow->dst));
22 }
23 
24 bool skb_flow_dissect(const struct sk_buff *skb, struct flow_keys *flow)
25 {
26 	int poff, nhoff = skb_network_offset(skb);
27 	u8 ip_proto;
28 	__be16 proto = skb->protocol;
29 
30 	memset(flow, 0, sizeof(*flow));
31 
32 again:
33 	switch (proto) {
34 	case __constant_htons(ETH_P_IP): {
35 		const struct iphdr *iph;
36 		struct iphdr _iph;
37 ip:
38 		iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
39 		if (!iph)
40 			return false;
41 
42 		if (ip_is_fragment(iph))
43 			ip_proto = 0;
44 		else
45 			ip_proto = iph->protocol;
46 		iph_to_flow_copy_addrs(flow, iph);
47 		nhoff += iph->ihl * 4;
48 		break;
49 	}
50 	case __constant_htons(ETH_P_IPV6): {
51 		const struct ipv6hdr *iph;
52 		struct ipv6hdr _iph;
53 ipv6:
54 		iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
55 		if (!iph)
56 			return false;
57 
58 		ip_proto = iph->nexthdr;
59 		flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr);
60 		flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr);
61 		nhoff += sizeof(struct ipv6hdr);
62 		break;
63 	}
64 	case __constant_htons(ETH_P_8021Q): {
65 		const struct vlan_hdr *vlan;
66 		struct vlan_hdr _vlan;
67 
68 		vlan = skb_header_pointer(skb, nhoff, sizeof(_vlan), &_vlan);
69 		if (!vlan)
70 			return false;
71 
72 		proto = vlan->h_vlan_encapsulated_proto;
73 		nhoff += sizeof(*vlan);
74 		goto again;
75 	}
76 	case __constant_htons(ETH_P_PPP_SES): {
77 		struct {
78 			struct pppoe_hdr hdr;
79 			__be16 proto;
80 		} *hdr, _hdr;
81 		hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr);
82 		if (!hdr)
83 			return false;
84 		proto = hdr->proto;
85 		nhoff += PPPOE_SES_HLEN;
86 		switch (proto) {
87 		case __constant_htons(PPP_IP):
88 			goto ip;
89 		case __constant_htons(PPP_IPV6):
90 			goto ipv6;
91 		default:
92 			return false;
93 		}
94 	}
95 	default:
96 		return false;
97 	}
98 
99 	switch (ip_proto) {
100 	case IPPROTO_GRE: {
101 		struct gre_hdr {
102 			__be16 flags;
103 			__be16 proto;
104 		} *hdr, _hdr;
105 
106 		hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr);
107 		if (!hdr)
108 			return false;
109 		/*
110 		 * Only look inside GRE if version zero and no
111 		 * routing
112 		 */
113 		if (!(hdr->flags & (GRE_VERSION|GRE_ROUTING))) {
114 			proto = hdr->proto;
115 			nhoff += 4;
116 			if (hdr->flags & GRE_CSUM)
117 				nhoff += 4;
118 			if (hdr->flags & GRE_KEY)
119 				nhoff += 4;
120 			if (hdr->flags & GRE_SEQ)
121 				nhoff += 4;
122 			goto again;
123 		}
124 		break;
125 	}
126 	case IPPROTO_IPIP:
127 		goto again;
128 	default:
129 		break;
130 	}
131 
132 	flow->ip_proto = ip_proto;
133 	poff = proto_ports_offset(ip_proto);
134 	if (poff >= 0) {
135 		__be32 *ports, _ports;
136 
137 		nhoff += poff;
138 		ports = skb_header_pointer(skb, nhoff, sizeof(_ports), &_ports);
139 		if (ports)
140 			flow->ports = *ports;
141 	}
142 
143 	return true;
144 }
145 EXPORT_SYMBOL(skb_flow_dissect);
146 
147 static u32 hashrnd __read_mostly;
148 
149 /*
150  * __skb_get_rxhash: calculate a flow hash based on src/dst addresses
151  * and src/dst port numbers.  Sets rxhash in skb to non-zero hash value
152  * on success, zero indicates no valid hash.  Also, sets l4_rxhash in skb
153  * if hash is a canonical 4-tuple hash over transport ports.
154  */
155 void __skb_get_rxhash(struct sk_buff *skb)
156 {
157 	struct flow_keys keys;
158 	u32 hash;
159 
160 	if (!skb_flow_dissect(skb, &keys))
161 		return;
162 
163 	if (keys.ports)
164 		skb->l4_rxhash = 1;
165 
166 	/* get a consistent hash (same value on both flow directions) */
167 	if (((__force u32)keys.dst < (__force u32)keys.src) ||
168 	    (((__force u32)keys.dst == (__force u32)keys.src) &&
169 	     ((__force u16)keys.port16[1] < (__force u16)keys.port16[0]))) {
170 		swap(keys.dst, keys.src);
171 		swap(keys.port16[0], keys.port16[1]);
172 	}
173 
174 	hash = jhash_3words((__force u32)keys.dst,
175 			    (__force u32)keys.src,
176 			    (__force u32)keys.ports, hashrnd);
177 	if (!hash)
178 		hash = 1;
179 
180 	skb->rxhash = hash;
181 }
182 EXPORT_SYMBOL(__skb_get_rxhash);
183 
184 /*
185  * Returns a Tx hash based on the given packet descriptor a Tx queues' number
186  * to be used as a distribution range.
187  */
188 u16 __skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb,
189 		  unsigned int num_tx_queues)
190 {
191 	u32 hash;
192 	u16 qoffset = 0;
193 	u16 qcount = num_tx_queues;
194 
195 	if (skb_rx_queue_recorded(skb)) {
196 		hash = skb_get_rx_queue(skb);
197 		while (unlikely(hash >= num_tx_queues))
198 			hash -= num_tx_queues;
199 		return hash;
200 	}
201 
202 	if (dev->num_tc) {
203 		u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
204 		qoffset = dev->tc_to_txq[tc].offset;
205 		qcount = dev->tc_to_txq[tc].count;
206 	}
207 
208 	if (skb->sk && skb->sk->sk_hash)
209 		hash = skb->sk->sk_hash;
210 	else
211 		hash = (__force u16) skb->protocol;
212 	hash = jhash_1word(hash, hashrnd);
213 
214 	return (u16) (((u64) hash * qcount) >> 32) + qoffset;
215 }
216 EXPORT_SYMBOL(__skb_tx_hash);
217 
218 static inline u16 dev_cap_txqueue(struct net_device *dev, u16 queue_index)
219 {
220 	if (unlikely(queue_index >= dev->real_num_tx_queues)) {
221 		net_warn_ratelimited("%s selects TX queue %d, but real number of TX queues is %d\n",
222 				     dev->name, queue_index,
223 				     dev->real_num_tx_queues);
224 		return 0;
225 	}
226 	return queue_index;
227 }
228 
229 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
230 {
231 #ifdef CONFIG_XPS
232 	struct xps_dev_maps *dev_maps;
233 	struct xps_map *map;
234 	int queue_index = -1;
235 
236 	rcu_read_lock();
237 	dev_maps = rcu_dereference(dev->xps_maps);
238 	if (dev_maps) {
239 		map = rcu_dereference(
240 		    dev_maps->cpu_map[raw_smp_processor_id()]);
241 		if (map) {
242 			if (map->len == 1)
243 				queue_index = map->queues[0];
244 			else {
245 				u32 hash;
246 				if (skb->sk && skb->sk->sk_hash)
247 					hash = skb->sk->sk_hash;
248 				else
249 					hash = (__force u16) skb->protocol ^
250 					    skb->rxhash;
251 				hash = jhash_1word(hash, hashrnd);
252 				queue_index = map->queues[
253 				    ((u64)hash * map->len) >> 32];
254 			}
255 			if (unlikely(queue_index >= dev->real_num_tx_queues))
256 				queue_index = -1;
257 		}
258 	}
259 	rcu_read_unlock();
260 
261 	return queue_index;
262 #else
263 	return -1;
264 #endif
265 }
266 
267 u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
268 {
269 	struct sock *sk = skb->sk;
270 	int queue_index = sk_tx_queue_get(sk);
271 
272 	if (queue_index < 0 || skb->ooo_okay ||
273 	    queue_index >= dev->real_num_tx_queues) {
274 		int new_index = get_xps_queue(dev, skb);
275 		if (new_index < 0)
276 			new_index = skb_tx_hash(dev, skb);
277 
278 		if (queue_index != new_index && sk) {
279 			struct dst_entry *dst =
280 				    rcu_dereference_check(sk->sk_dst_cache, 1);
281 
282 			if (dst && skb_dst(skb) == dst)
283 				sk_tx_queue_set(sk, queue_index);
284 
285 		}
286 
287 		queue_index = new_index;
288 	}
289 
290 	return queue_index;
291 }
292 EXPORT_SYMBOL(__netdev_pick_tx);
293 
294 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
295 				    struct sk_buff *skb)
296 {
297 	int queue_index = 0;
298 
299 	if (dev->real_num_tx_queues != 1) {
300 		const struct net_device_ops *ops = dev->netdev_ops;
301 		if (ops->ndo_select_queue)
302 			queue_index = ops->ndo_select_queue(dev, skb);
303 		else
304 			queue_index = __netdev_pick_tx(dev, skb);
305 		queue_index = dev_cap_txqueue(dev, queue_index);
306 	}
307 
308 	skb_set_queue_mapping(skb, queue_index);
309 	return netdev_get_tx_queue(dev, queue_index);
310 }
311 
312 static int __init initialize_hashrnd(void)
313 {
314 	get_random_bytes(&hashrnd, sizeof(hashrnd));
315 	return 0;
316 }
317 
318 late_initcall_sync(initialize_hashrnd);
319