xref: /freebsd/sys/netinet6/in6_rss.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 /*-
2  * Copyright (c) 2010-2011 Juniper Networks, Inc.
3  * All rights reserved.
4  *
5  * This software was developed by Robert N. M. Watson under contract
6  * to Juniper Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_inet6.h"
35 #include "opt_pcbgroup.h"
36 
37 #ifndef PCBGROUP
38 #error "options RSS depends on options PCBGROUP"
39 #endif
40 
41 #include <sys/param.h>
42 #include <sys/mbuf.h>
43 #include <sys/socket.h>
44 #include <sys/priv.h>
45 #include <sys/kernel.h>
46 #include <sys/smp.h>
47 #include <sys/sysctl.h>
48 #include <sys/sbuf.h>
49 
50 #include <net/if.h>
51 #include <net/if_var.h>
52 #include <net/netisr.h>
53 #include <net/rss_config.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_pcb.h>
57 #include <netinet6/in6_rss.h>
58 #include <netinet/in_var.h>
59 
60 /* for software rss hash support */
61 #include <netinet/ip6.h>
62 #include <netinet6/ip6_var.h>
63 #include <netinet/tcp.h>
64 #include <netinet/udp.h>
65 
66 /*
67  * Hash an IPv6 2-tuple.
68  */
69 uint32_t
70 rss_hash_ip6_2tuple(const struct in6_addr *src, const struct in6_addr *dst)
71 {
72 	uint8_t data[sizeof(*src) + sizeof(*dst)];
73 	u_int datalen;
74 
75 	datalen = 0;
76 	bcopy(src, &data[datalen], sizeof(*src));
77 	datalen += sizeof(*src);
78 	bcopy(dst, &data[datalen], sizeof(*dst));
79 	datalen += sizeof(*dst);
80 	return (rss_hash(datalen, data));
81 }
82 
83 /*
84  * Hash an IPv6 4-tuple.
85  */
86 uint32_t
87 rss_hash_ip6_4tuple(const struct in6_addr *src, u_short srcport,
88     const struct in6_addr *dst, u_short dstport)
89 {
90 	uint8_t data[sizeof(*src) + sizeof(*dst) + sizeof(srcport) +
91 	    sizeof(dstport)];
92 	u_int datalen;
93 
94 	datalen = 0;
95 	bcopy(src, &data[datalen], sizeof(*src));
96 	datalen += sizeof(*src);
97 	bcopy(dst, &data[datalen], sizeof(*dst));
98 	datalen += sizeof(*dst);
99 	bcopy(&srcport, &data[datalen], sizeof(srcport));
100 	datalen += sizeof(srcport);
101 	bcopy(&dstport, &data[datalen], sizeof(dstport));
102 	datalen += sizeof(dstport);
103 	return (rss_hash(datalen, data));
104 }
105 
106 /*
107  * Calculate an appropriate ipv6 2-tuple or 4-tuple given the given
108  * IPv6 source/destination address, UDP or TCP source/destination ports
109  * and the protocol type.
110  *
111  * The protocol code may wish to do a software hash of the given
112  * tuple.  This depends upon the currently configured RSS hash types.
113  *
114  * This assumes that the packet in question isn't a fragment.
115  *
116  * It also assumes the packet source/destination address
117  * are in "incoming" packet order (ie, source is "far" address.)
118  */
119 int
120 rss_proto_software_hash_v6(const struct in6_addr *s, const struct in6_addr *d,
121     u_short sp, u_short dp, int proto,
122     uint32_t *hashval, uint32_t *hashtype)
123 {
124 	uint32_t hash;
125 
126 	/*
127 	 * Next, choose the hash type depending upon the protocol
128 	 * identifier.
129 	 */
130 	if ((proto == IPPROTO_TCP) &&
131 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6)) {
132 		hash = rss_hash_ip6_4tuple(s, sp, d, dp);
133 		*hashval = hash;
134 		*hashtype = M_HASHTYPE_RSS_TCP_IPV6;
135 		return (0);
136 	} else if ((proto == IPPROTO_UDP) &&
137 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6)) {
138 		hash = rss_hash_ip6_4tuple(s, sp, d, dp);
139 		*hashval = hash;
140 		*hashtype = M_HASHTYPE_RSS_UDP_IPV6;
141 		return (0);
142 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) {
143 		/* RSS doesn't hash on other protocols like SCTP; so 2-tuple */
144 		hash = rss_hash_ip6_2tuple(s, d);
145 		*hashval = hash;
146 		*hashtype = M_HASHTYPE_RSS_IPV6;
147 		return (0);
148 	}
149 
150 	/* No configured available hashtypes! */
151 	RSS_DEBUG("no available hashtypes!\n");
152 	return (-1);
153 }
154 
155 /*
156  * Do a software calculation of the RSS for the given mbuf.
157  *
158  * This is typically used by the input path to recalculate the RSS after
159  * some form of packet processing (eg de-capsulation, IP fragment reassembly.)
160  *
161  * dir is the packet direction - RSS_HASH_PKT_INGRESS for incoming and
162  * RSS_HASH_PKT_EGRESS for outgoing.
163  *
164  * Returns 0 if a hash was done, -1 if no hash was done, +1 if
165  * the mbuf already had a valid RSS flowid.
166  *
167  * This function doesn't modify the mbuf.  It's up to the caller to
168  * assign flowid/flowtype as appropriate.
169  */
170 int
171 rss_mbuf_software_hash_v6(const struct mbuf *m, int dir, uint32_t *hashval,
172     uint32_t *hashtype)
173 {
174 	const struct ip6_hdr *ip6;
175 	const struct ip6_frag *ip6f;
176 	const struct tcphdr *th;
177 	const struct udphdr *uh;
178 	uint32_t flowtype;
179 	uint8_t proto;
180 	int off, newoff;
181 	int nxt;
182 
183 	/*
184 	 * XXX For now this only handles hashing on incoming mbufs.
185 	 */
186 	if (dir != RSS_HASH_PKT_INGRESS) {
187 		RSS_DEBUG("called on EGRESS packet!\n");
188 		return (-1);
189 	}
190 
191 	off = sizeof(struct ip6_hdr);
192 
193 	/*
194 	 * First, validate that the mbuf we have is long enough
195 	 * to have an IPv6 header in it.
196 	 */
197 	if (m->m_pkthdr.len < off) {
198 		RSS_DEBUG("short mbuf pkthdr\n");
199 		return (-1);
200 	}
201 	if (m->m_len < off) {
202 		RSS_DEBUG("short mbuf len\n");
203 		return (-1);
204 	}
205 
206 	/* Ok, let's dereference that */
207 	ip6 = mtod(m, struct ip6_hdr *);
208 	proto = ip6->ip6_nxt;
209 
210 	/*
211 	 * Find the beginning of the TCP/UDP header.
212 	 *
213 	 * If this is a fragment then it shouldn't be four-tuple
214 	 * hashed just yet.  Once it's reassembled into a full
215 	 * frame it should be re-hashed.
216 	 */
217 	while (proto != IPPROTO_FRAGMENT) {
218 		newoff = ip6_nexthdr(m, off, proto, &nxt);
219 		if (newoff < 0)
220 			break;
221 		off = newoff;
222 		proto = nxt;
223 	}
224 
225 	/*
226 	 * Ignore the fragment header if this is an "atomic" fragment
227 	 * (offset and m bit set to 0)
228 	 */
229 	if (proto == IPPROTO_FRAGMENT) {
230 		if (m->m_len < off + sizeof(struct ip6_frag)) {
231 			RSS_DEBUG("short fragment frame?\n");
232 			return (-1);
233 		}
234 		ip6f = (const struct ip6_frag *)((c_caddr_t)ip6 + off);
235 		if ((ip6f->ip6f_offlg & ~IP6F_RESERVED_MASK) == 0) {
236 			off = ip6_lasthdr(m, off, proto, &nxt);
237 			if (off < 0) {
238 				RSS_DEBUG("invalid extension header\n");
239 				return (-1);
240 			}
241 			proto = nxt;
242 		}
243 	}
244 
245 	/*
246 	 * If the mbuf flowid/flowtype matches the packet type,
247 	 * and we don't support the 4-tuple version of the given protocol,
248 	 * then signal to the owner that it can trust the flowid/flowtype
249 	 * details.
250 	 *
251 	 * This is a little picky - eg, if TCPv6 / UDPv6 hashing
252 	 * is supported but we got a TCP/UDP frame only 2-tuple hashed,
253 	 * then we shouldn't just "trust" the 2-tuple hash.  We need
254 	 * a 4-tuple hash.
255 	 */
256 	flowtype = M_HASHTYPE_GET(m);
257 
258 	if (flowtype != M_HASHTYPE_NONE) {
259 		switch (proto) {
260 		case IPPROTO_UDP:
261 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) &&
262 			    (flowtype == M_HASHTYPE_RSS_UDP_IPV6)) {
263 				return (1);
264 			}
265 			/*
266 			 * Only allow 2-tuple for UDP frames if we don't also
267 			 * support 4-tuple for UDP.
268 			 */
269 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
270 			    ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) == 0) &&
271 			    flowtype == M_HASHTYPE_RSS_IPV6) {
272 				return (1);
273 			}
274 			break;
275 		case IPPROTO_TCP:
276 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) &&
277 			    (flowtype == M_HASHTYPE_RSS_TCP_IPV6)) {
278 				return (1);
279 			}
280 			/*
281 			 * Only allow 2-tuple for TCP frames if we don't also
282 			 * support 4-tuple for TCP.
283 			 */
284 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
285 			    ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) == 0) &&
286 			    flowtype == M_HASHTYPE_RSS_IPV6) {
287 				return (1);
288 			}
289 			break;
290 		default:
291 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
292 			    flowtype == M_HASHTYPE_RSS_IPV6) {
293 				return (1);
294 			}
295 			break;
296 		}
297 	}
298 
299 	/*
300 	 * Decode enough information to make a hash decision.
301 	 */
302 	if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) &&
303 	    (proto == IPPROTO_TCP)) {
304 		if (m->m_len < off + sizeof(struct tcphdr)) {
305 			RSS_DEBUG("short TCP frame?\n");
306 			return (-1);
307 		}
308 		th = (const struct tcphdr *)((c_caddr_t)ip6 + off);
309 		return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
310 		    th->th_sport,
311 		    th->th_dport,
312 		    proto,
313 		    hashval,
314 		    hashtype);
315 	} else if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) &&
316 	    (proto == IPPROTO_UDP)) {
317 		if (m->m_len < off + sizeof(struct udphdr)) {
318 			RSS_DEBUG("short UDP frame?\n");
319 			return (-1);
320 		}
321 		uh = (const struct udphdr *)((c_caddr_t)ip6 + off);
322 		return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
323 		    uh->uh_sport,
324 		    uh->uh_dport,
325 		    proto,
326 		    hashval,
327 		    hashtype);
328 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) {
329 		/* Default to 2-tuple hash */
330 		return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
331 		    0,	/* source port */
332 		    0,	/* destination port */
333 		    0,	/* IPPROTO_IP */
334 		    hashval,
335 		    hashtype);
336 	} else {
337 		RSS_DEBUG("no available hashtypes!\n");
338 		return (-1);
339 	}
340 }
341 
342 /*
343  * Similar to rss_m2cpuid, but designed to be used by the IPv6 NETISR
344  * on incoming frames.
345  *
346  * If an existing RSS hash exists and it matches what the configured
347  * hashing is, then use it.
348  *
349  * If there's an existing RSS hash but the desired hash is different,
350  * or if there's no useful RSS hash, then calculate it via
351  * the software path.
352  *
353  * XXX TODO: definitely want statistics here!
354  */
355 struct mbuf *
356 rss_soft_m2cpuid_v6(struct mbuf *m, uintptr_t source, u_int *cpuid)
357 {
358 	uint32_t hash_val, hash_type;
359 	int ret;
360 
361 	M_ASSERTPKTHDR(m);
362 
363 	ret = rss_mbuf_software_hash_v6(m, RSS_HASH_PKT_INGRESS,
364 	    &hash_val, &hash_type);
365 	if (ret > 0) {
366 		/* mbuf has a valid hash already; don't need to modify it */
367 		*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
368 	} else if (ret == 0) {
369 		/* hash was done; update */
370 		m->m_pkthdr.flowid = hash_val;
371 		M_HASHTYPE_SET(m, hash_type);
372 		*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
373 	} else { /* ret < 0 */
374 		/* no hash was done */
375 		*cpuid = NETISR_CPUID_NONE;
376 	}
377 	return (m);
378 }
379