xref: /freebsd/sys/netinet6/in6_rss.c (revision 9978553d0199e7ec0bdd1c44fc7f6c7b0c11e43b)
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 
31 #include "opt_inet6.h"
32 #include "opt_rss.h"
33 
34 #include <sys/param.h>
35 #include <sys/mbuf.h>
36 #include <sys/socket.h>
37 #include <sys/priv.h>
38 #include <sys/kernel.h>
39 #include <sys/smp.h>
40 #include <sys/sysctl.h>
41 #include <sys/sbuf.h>
42 
43 #include <net/if.h>
44 #include <net/if_var.h>
45 #include <net/netisr.h>
46 #include <net/rss_config.h>
47 
48 #include <netinet/in.h>
49 #include <netinet/in_pcb.h>
50 #include <netinet6/in6_rss.h>
51 #include <netinet/in_var.h>
52 
53 /* for software rss hash support */
54 #include <netinet/ip6.h>
55 #include <netinet6/ip6_var.h>
56 #include <netinet/tcp.h>
57 #include <netinet/udp.h>
58 
59 /*
60  * Hash an IPv6 2-tuple.
61  */
62 uint32_t
63 rss_hash_ip6_2tuple(const struct in6_addr *src, const struct in6_addr *dst)
64 {
65 	uint8_t data[sizeof(*src) + sizeof(*dst)];
66 	u_int datalen;
67 
68 	datalen = 0;
69 	bcopy(src, &data[datalen], sizeof(*src));
70 	datalen += sizeof(*src);
71 	bcopy(dst, &data[datalen], sizeof(*dst));
72 	datalen += sizeof(*dst);
73 	return (rss_hash(datalen, data));
74 }
75 
76 /*
77  * Hash an IPv6 4-tuple.
78  */
79 uint32_t
80 rss_hash_ip6_4tuple(const struct in6_addr *src, u_short srcport,
81     const struct in6_addr *dst, u_short dstport)
82 {
83 	uint8_t data[sizeof(*src) + sizeof(*dst) + sizeof(srcport) +
84 	    sizeof(dstport)];
85 	u_int datalen;
86 
87 	datalen = 0;
88 	bcopy(src, &data[datalen], sizeof(*src));
89 	datalen += sizeof(*src);
90 	bcopy(dst, &data[datalen], sizeof(*dst));
91 	datalen += sizeof(*dst);
92 	bcopy(&srcport, &data[datalen], sizeof(srcport));
93 	datalen += sizeof(srcport);
94 	bcopy(&dstport, &data[datalen], sizeof(dstport));
95 	datalen += sizeof(dstport);
96 	return (rss_hash(datalen, data));
97 }
98 
99 /*
100  * Calculate an appropriate ipv6 2-tuple or 4-tuple given the given
101  * IPv6 source/destination address, UDP or TCP source/destination ports
102  * and the protocol type.
103  *
104  * The protocol code may wish to do a software hash of the given
105  * tuple.  This depends upon the currently configured RSS hash types.
106  *
107  * This assumes that the packet in question isn't a fragment.
108  *
109  * It also assumes the packet source/destination address
110  * are in "incoming" packet order (ie, source is "far" address.)
111  */
112 int
113 rss_proto_software_hash_v6(const struct in6_addr *s, const struct in6_addr *d,
114     u_short sp, u_short dp, int proto,
115     uint32_t *hashval, uint32_t *hashtype)
116 {
117 	uint32_t hash;
118 
119 	/*
120 	 * Next, choose the hash type depending upon the protocol
121 	 * identifier.
122 	 */
123 	if ((proto == IPPROTO_TCP) &&
124 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6)) {
125 		hash = rss_hash_ip6_4tuple(s, sp, d, dp);
126 		*hashval = hash;
127 		*hashtype = M_HASHTYPE_RSS_TCP_IPV6;
128 		return (0);
129 	} else if ((proto == IPPROTO_UDP) &&
130 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6)) {
131 		hash = rss_hash_ip6_4tuple(s, sp, d, dp);
132 		*hashval = hash;
133 		*hashtype = M_HASHTYPE_RSS_UDP_IPV6;
134 		return (0);
135 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) {
136 		/* RSS doesn't hash on other protocols like SCTP; so 2-tuple */
137 		hash = rss_hash_ip6_2tuple(s, d);
138 		*hashval = hash;
139 		*hashtype = M_HASHTYPE_RSS_IPV6;
140 		return (0);
141 	}
142 
143 	/* No configured available hashtypes! */
144 	RSS_DEBUG("no available hashtypes!\n");
145 	return (-1);
146 }
147 
148 /*
149  * Calculate an appropriate ipv6 2-tuple or 4-tuple given the given
150  * IPv6 source/destination address, UDP or TCP source/destination ports
151  * and the protocol type.
152  *
153  * The protocol code may wish to do a software hash of the given
154  * tuple. This depends upon the currently configured RSS hash types.
155  *
156  * It assumes the packet source/destination address
157  * are in "outgoin" packet order (ie, destination is "far" address.)
158  */
159 uint32_t
160 xps_proto_software_hash_v6(const struct in6_addr *s, const struct in6_addr *d,
161     u_short sp, u_short dp, int proto, uint32_t *hashtype)
162 {
163 
164 	uint32_t hash;
165 
166 	/*
167 	 * Next, choose the hash type depending upon the protocol
168 	 * identifier.
169 	 */
170 	if ((proto == IPPROTO_TCP) &&
171 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6)) {
172 		hash = rss_hash_ip6_4tuple(d, dp, s, sp);
173 		*hashtype = M_HASHTYPE_RSS_TCP_IPV6;
174 		return (hash);
175 	} else if ((proto == IPPROTO_UDP) &&
176 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6)) {
177 		hash = rss_hash_ip6_4tuple(d, dp, s, sp);
178 		*hashtype = M_HASHTYPE_RSS_UDP_IPV6;
179 		return (hash);
180 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) {
181 		/* RSS doesn't hash on other protocols like SCTP; so 2-tuple */
182 		hash = rss_hash_ip6_2tuple(d, s);
183 		*hashtype = M_HASHTYPE_RSS_IPV6;
184 		return (hash);
185 	}
186 
187 	*hashtype = M_HASHTYPE_NONE;
188 	return (0);
189 }
190 
191 
192 /*
193  * Do a software calculation of the RSS for the given mbuf.
194  *
195  * This is typically used by the input path to recalculate the RSS after
196  * some form of packet processing (eg de-capsulation, IP fragment reassembly.)
197  *
198  * dir is the packet direction - RSS_HASH_PKT_INGRESS for incoming and
199  * RSS_HASH_PKT_EGRESS for outgoing.
200  *
201  * Returns 0 if a hash was done, -1 if no hash was done, +1 if
202  * the mbuf already had a valid RSS flowid.
203  *
204  * This function doesn't modify the mbuf.  It's up to the caller to
205  * assign flowid/flowtype as appropriate.
206  */
207 int
208 rss_mbuf_software_hash_v6(const struct mbuf *m, int dir, uint32_t *hashval,
209     uint32_t *hashtype)
210 {
211 	const struct ip6_hdr *ip6;
212 	const struct ip6_frag *ip6f;
213 	const struct tcphdr *th;
214 	const struct udphdr *uh;
215 	uint32_t flowtype;
216 	uint8_t proto;
217 	int off, newoff;
218 	int nxt;
219 
220 	/*
221 	 * XXX For now this only handles hashing on incoming mbufs.
222 	 */
223 	if (dir != RSS_HASH_PKT_INGRESS) {
224 		RSS_DEBUG("called on EGRESS packet!\n");
225 		return (-1);
226 	}
227 
228 	off = sizeof(struct ip6_hdr);
229 
230 	/*
231 	 * First, validate that the mbuf we have is long enough
232 	 * to have an IPv6 header in it.
233 	 */
234 	if (m->m_pkthdr.len < off) {
235 		RSS_DEBUG("short mbuf pkthdr\n");
236 		return (-1);
237 	}
238 	if (m->m_len < off) {
239 		RSS_DEBUG("short mbuf len\n");
240 		return (-1);
241 	}
242 
243 	/* Ok, let's dereference that */
244 	ip6 = mtod(m, struct ip6_hdr *);
245 	proto = ip6->ip6_nxt;
246 
247 	/*
248 	 * Find the beginning of the TCP/UDP header.
249 	 *
250 	 * If this is a fragment then it shouldn't be four-tuple
251 	 * hashed just yet.  Once it's reassembled into a full
252 	 * frame it should be re-hashed.
253 	 */
254 	while (proto != IPPROTO_FRAGMENT) {
255 		newoff = ip6_nexthdr(m, off, proto, &nxt);
256 		if (newoff < 0)
257 			break;
258 		off = newoff;
259 		proto = nxt;
260 	}
261 
262 	/*
263 	 * Ignore the fragment header if this is an "atomic" fragment
264 	 * (offset and m bit set to 0)
265 	 */
266 	if (proto == IPPROTO_FRAGMENT) {
267 		if (m->m_len < off + sizeof(struct ip6_frag)) {
268 			RSS_DEBUG("short fragment frame?\n");
269 			return (-1);
270 		}
271 		ip6f = (const struct ip6_frag *)((c_caddr_t)ip6 + off);
272 		if ((ip6f->ip6f_offlg & ~IP6F_RESERVED_MASK) == 0) {
273 			off = ip6_lasthdr(m, off, proto, &nxt);
274 			if (off < 0) {
275 				RSS_DEBUG("invalid extension header\n");
276 				return (-1);
277 			}
278 			proto = nxt;
279 		}
280 	}
281 
282 	/*
283 	 * If the mbuf flowid/flowtype matches the packet type,
284 	 * and we don't support the 4-tuple version of the given protocol,
285 	 * then signal to the owner that it can trust the flowid/flowtype
286 	 * details.
287 	 *
288 	 * This is a little picky - eg, if TCPv6 / UDPv6 hashing
289 	 * is supported but we got a TCP/UDP frame only 2-tuple hashed,
290 	 * then we shouldn't just "trust" the 2-tuple hash.  We need
291 	 * a 4-tuple hash.
292 	 */
293 	flowtype = M_HASHTYPE_GET(m);
294 
295 	if (flowtype != M_HASHTYPE_NONE) {
296 		switch (proto) {
297 		case IPPROTO_UDP:
298 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) &&
299 			    (flowtype == M_HASHTYPE_RSS_UDP_IPV6)) {
300 				return (1);
301 			}
302 			/*
303 			 * Only allow 2-tuple for UDP frames if we don't also
304 			 * support 4-tuple for UDP.
305 			 */
306 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
307 			    ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) == 0) &&
308 			    flowtype == M_HASHTYPE_RSS_IPV6) {
309 				return (1);
310 			}
311 			break;
312 		case IPPROTO_TCP:
313 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) &&
314 			    (flowtype == M_HASHTYPE_RSS_TCP_IPV6)) {
315 				return (1);
316 			}
317 			/*
318 			 * Only allow 2-tuple for TCP frames if we don't also
319 			 * support 4-tuple for TCP.
320 			 */
321 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
322 			    ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) == 0) &&
323 			    flowtype == M_HASHTYPE_RSS_IPV6) {
324 				return (1);
325 			}
326 			break;
327 		default:
328 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
329 			    flowtype == M_HASHTYPE_RSS_IPV6) {
330 				return (1);
331 			}
332 			break;
333 		}
334 	}
335 
336 	/*
337 	 * Decode enough information to make a hash decision.
338 	 */
339 	if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) &&
340 	    (proto == IPPROTO_TCP)) {
341 		if (m->m_len < off + sizeof(struct tcphdr)) {
342 			RSS_DEBUG("short TCP frame?\n");
343 			return (-1);
344 		}
345 		th = (const struct tcphdr *)((c_caddr_t)ip6 + off);
346 		return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
347 		    th->th_sport,
348 		    th->th_dport,
349 		    proto,
350 		    hashval,
351 		    hashtype);
352 	} else if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) &&
353 	    (proto == IPPROTO_UDP)) {
354 		if (m->m_len < off + sizeof(struct udphdr)) {
355 			RSS_DEBUG("short UDP frame?\n");
356 			return (-1);
357 		}
358 		uh = (const struct udphdr *)((c_caddr_t)ip6 + off);
359 		return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
360 		    uh->uh_sport,
361 		    uh->uh_dport,
362 		    proto,
363 		    hashval,
364 		    hashtype);
365 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) {
366 		/* Default to 2-tuple hash */
367 		return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
368 		    0,	/* source port */
369 		    0,	/* destination port */
370 		    0,	/* IPPROTO_IP */
371 		    hashval,
372 		    hashtype);
373 	} else {
374 		RSS_DEBUG("no available hashtypes!\n");
375 		return (-1);
376 	}
377 }
378 
379 #ifdef RSS
380 /*
381  * Similar to rss_m2cpuid, but designed to be used by the IPv6 NETISR
382  * on incoming frames.
383  *
384  * If an existing RSS hash exists and it matches what the configured
385  * hashing is, then use it.
386  *
387  * If there's an existing RSS hash but the desired hash is different,
388  * or if there's no useful RSS hash, then calculate it via
389  * the software path.
390  *
391  * XXX TODO: definitely want statistics here!
392  */
393 struct mbuf *
394 rss_soft_m2cpuid_v6(struct mbuf *m, uintptr_t source, u_int *cpuid)
395 {
396 	uint32_t hash_val, hash_type;
397 	int ret;
398 
399 	M_ASSERTPKTHDR(m);
400 
401 	ret = rss_mbuf_software_hash_v6(m, RSS_HASH_PKT_INGRESS,
402 	    &hash_val, &hash_type);
403 	if (ret > 0) {
404 		/* mbuf has a valid hash already; don't need to modify it */
405 		*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
406 	} else if (ret == 0) {
407 		/* hash was done; update */
408 		m->m_pkthdr.flowid = hash_val;
409 		M_HASHTYPE_SET(m, hash_type);
410 		*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
411 	} else { /* ret < 0 */
412 		/* no hash was done */
413 		*cpuid = NETISR_CPUID_NONE;
414 	}
415 	return (m);
416 }
417 #endif
418