xref: /freebsd/sys/netinet/in_rss.c (revision 681ce946f33e75c590e97c53076e86dff1fe8f4a)
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 
36 #include <sys/param.h>
37 #include <sys/mbuf.h>
38 #include <sys/socket.h>
39 #include <sys/priv.h>
40 #include <sys/kernel.h>
41 #include <sys/smp.h>
42 #include <sys/sysctl.h>
43 #include <sys/sbuf.h>
44 
45 #include <net/if.h>
46 #include <net/if_var.h>
47 #include <net/netisr.h>
48 #include <net/rss_config.h>
49 
50 #include <netinet/in.h>
51 #include <netinet/in_pcb.h>
52 #include <netinet/in_rss.h>
53 #include <netinet/in_var.h>
54 
55 /* for software rss hash support */
56 #include <netinet/ip.h>
57 #include <netinet/tcp.h>
58 #include <netinet/udp.h>
59 
60 /*
61  * Hash an IPv4 2-tuple.
62  */
63 uint32_t
64 rss_hash_ip4_2tuple(struct in_addr src, struct in_addr dst)
65 {
66 	uint8_t data[sizeof(src) + sizeof(dst)];
67 	u_int datalen;
68 
69 	datalen = 0;
70 	bcopy(&src, &data[datalen], sizeof(src));
71 	datalen += sizeof(src);
72 	bcopy(&dst, &data[datalen], sizeof(dst));
73 	datalen += sizeof(dst);
74 	return (rss_hash(datalen, data));
75 }
76 
77 /*
78  * Hash an IPv4 4-tuple.
79  */
80 uint32_t
81 rss_hash_ip4_4tuple(struct in_addr src, u_short srcport, struct in_addr dst,
82     u_short dstport)
83 {
84 	uint8_t data[sizeof(src) + sizeof(dst) + sizeof(srcport) +
85 	    sizeof(dstport)];
86 	u_int datalen;
87 
88 	datalen = 0;
89 	bcopy(&src, &data[datalen], sizeof(src));
90 	datalen += sizeof(src);
91 	bcopy(&dst, &data[datalen], sizeof(dst));
92 	datalen += sizeof(dst);
93 	bcopy(&srcport, &data[datalen], sizeof(srcport));
94 	datalen += sizeof(srcport);
95 	bcopy(&dstport, &data[datalen], sizeof(dstport));
96 	datalen += sizeof(dstport);
97 	return (rss_hash(datalen, data));
98 }
99 
100 /*
101  * Calculate an appropriate ipv4 2-tuple or 4-tuple given the given
102  * IPv4 source/destination address, UDP or TCP source/destination ports
103  * and the protocol type.
104  *
105  * The protocol code may wish to do a software hash of the given
106  * tuple.  This depends upon the currently configured RSS hash types.
107  *
108  * This assumes that the packet in question isn't a fragment.
109  *
110  * It also assumes the packet source/destination address
111  * are in "incoming" packet order (ie, source is "far" address.)
112  */
113 int
114 rss_proto_software_hash_v4(struct in_addr s, struct in_addr d,
115     u_short sp, u_short dp, int proto,
116     uint32_t *hashval, uint32_t *hashtype)
117 {
118 	uint32_t hash;
119 
120 	/*
121 	 * Next, choose the hash type depending upon the protocol
122 	 * identifier.
123 	 */
124 	if ((proto == IPPROTO_TCP) &&
125 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4)) {
126 		hash = rss_hash_ip4_4tuple(s, sp, d, dp);
127 		*hashval = hash;
128 		*hashtype = M_HASHTYPE_RSS_TCP_IPV4;
129 		return (0);
130 	} else if ((proto == IPPROTO_UDP) &&
131 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4)) {
132 		hash = rss_hash_ip4_4tuple(s, sp, d, dp);
133 		*hashval = hash;
134 		*hashtype = M_HASHTYPE_RSS_UDP_IPV4;
135 		return (0);
136 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) {
137 		/* RSS doesn't hash on other protocols like SCTP; so 2-tuple */
138 		hash = rss_hash_ip4_2tuple(s, d);
139 		*hashval = hash;
140 		*hashtype = M_HASHTYPE_RSS_IPV4;
141 		return (0);
142 	}
143 
144 	/* No configured available hashtypes! */
145 	RSS_DEBUG("no available hashtypes!\n");
146 	return (-1);
147 }
148 
149 /*
150  * Calculate an appropriate ipv4 2-tuple or 4-tuple given the given
151  * IPv4 source/destination address, UDP or TCP source/destination ports
152  * and the protocol type.
153  *
154  * The protocol code may wish to do a software hash of the given
155  * tuple.  This depends upon the currently configured RSS hash types.
156  *
157  * It assumes the packet source/destination address
158  * are in "outgoing" packet order (ie, destination is "far" address.)
159  */
160 uint32_t
161 xps_proto_software_hash_v4(struct in_addr s, struct in_addr d,
162     u_short sp, u_short dp, int proto, uint32_t *hashtype)
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_IPV4)) {
172 		hash = rss_hash_ip4_4tuple(d, dp, s, sp);
173 		*hashtype = M_HASHTYPE_RSS_TCP_IPV4;
174 		return (hash);
175 	} else if ((proto == IPPROTO_UDP) &&
176 	    (rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4)) {
177 		hash = rss_hash_ip4_4tuple(d, dp, s, sp);
178 		*hashtype = M_HASHTYPE_RSS_UDP_IPV4;
179 		return (hash);
180 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) {
181 		/* RSS doesn't hash on other protocols like SCTP; so 2-tuple */
182 		hash = rss_hash_ip4_2tuple(d, s);
183 		*hashtype = M_HASHTYPE_RSS_IPV4;
184 		return (hash);
185 	}
186 
187 	*hashtype = M_HASHTYPE_NONE;
188 	return (0);
189 }
190 
191 /*
192  * Do a software calculation of the RSS for the given mbuf.
193  *
194  * This is typically used by the input path to recalculate the RSS after
195  * some form of packet processing (eg de-capsulation, IP fragment reassembly.)
196  *
197  * dir is the packet direction - RSS_HASH_PKT_INGRESS for incoming and
198  * RSS_HASH_PKT_EGRESS for outgoing.
199  *
200  * Returns 0 if a hash was done, -1 if no hash was done, +1 if
201  * the mbuf already had a valid RSS flowid.
202  *
203  * This function doesn't modify the mbuf.  It's up to the caller to
204  * assign flowid/flowtype as appropriate.
205  */
206 int
207 rss_mbuf_software_hash_v4(const struct mbuf *m, int dir, uint32_t *hashval,
208     uint32_t *hashtype)
209 {
210 	const struct ip *ip;
211 	const struct tcphdr *th;
212 	const struct udphdr *uh;
213 	uint32_t flowid;
214 	uint32_t flowtype;
215 	uint8_t proto;
216 	int iphlen;
217 	int is_frag = 0;
218 
219 	/*
220 	 * XXX For now this only handles hashing on incoming mbufs.
221 	 */
222 	if (dir != RSS_HASH_PKT_INGRESS) {
223 		RSS_DEBUG("called on EGRESS packet!\n");
224 		return (-1);
225 	}
226 
227 	/*
228 	 * First, validate that the mbuf we have is long enough
229 	 * to have an IPv4 header in it.
230 	 */
231 	if (m->m_pkthdr.len < (sizeof(struct ip))) {
232 		RSS_DEBUG("short mbuf pkthdr\n");
233 		return (-1);
234 	}
235 	if (m->m_len < (sizeof(struct ip))) {
236 		RSS_DEBUG("short mbuf len\n");
237 		return (-1);
238 	}
239 
240 	/* Ok, let's dereference that */
241 	ip = mtod(m, struct ip *);
242 	proto = ip->ip_p;
243 	iphlen = ip->ip_hl << 2;
244 
245 	/*
246 	 * If this is a fragment then it shouldn't be four-tuple
247 	 * hashed just yet.  Once it's reassembled into a full
248 	 * frame it should be re-hashed.
249 	 */
250 	if (ip->ip_off & htons(IP_MF | IP_OFFMASK))
251 		is_frag = 1;
252 
253 	/*
254 	 * If the mbuf flowid/flowtype matches the packet type,
255 	 * and we don't support the 4-tuple version of the given protocol,
256 	 * then signal to the owner that it can trust the flowid/flowtype
257 	 * details.
258 	 *
259 	 * This is a little picky - eg, if TCPv4 / UDPv4 hashing
260 	 * is supported but we got a TCP/UDP frame only 2-tuple hashed,
261 	 * then we shouldn't just "trust" the 2-tuple hash.  We need
262 	 * a 4-tuple hash.
263 	 */
264 	flowid = m->m_pkthdr.flowid;
265 	flowtype = M_HASHTYPE_GET(m);
266 
267 	if (flowtype != M_HASHTYPE_NONE) {
268 		switch (proto) {
269 		case IPPROTO_UDP:
270 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) &&
271 			    (flowtype == M_HASHTYPE_RSS_UDP_IPV4) &&
272 			    (is_frag == 0)) {
273 				return (1);
274 			}
275 			/*
276 			 * Only allow 2-tuple for UDP frames if we don't also
277 			 * support 4-tuple for UDP.
278 			 */
279 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
280 			    ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) == 0) &&
281 			    flowtype == M_HASHTYPE_RSS_IPV4) {
282 				return (1);
283 			}
284 			break;
285 		case IPPROTO_TCP:
286 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) &&
287 			    (flowtype == M_HASHTYPE_RSS_TCP_IPV4) &&
288 			    (is_frag == 0)) {
289 				return (1);
290 			}
291 			/*
292 			 * Only allow 2-tuple for TCP frames if we don't also
293 			 * support 2-tuple for TCP.
294 			 */
295 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
296 			    ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) == 0) &&
297 			    flowtype == M_HASHTYPE_RSS_IPV4) {
298 				return (1);
299 			}
300 			break;
301 		default:
302 			if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
303 			    flowtype == M_HASHTYPE_RSS_IPV4) {
304 				return (1);
305 			}
306 			break;
307 		}
308 	}
309 
310 	/*
311 	 * Decode enough information to make a hash decision.
312 	 *
313 	 * XXX TODO: does the hardware hash on 4-tuple if IP
314 	 *    options are present?
315 	 */
316 	if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) &&
317 	    (proto == IPPROTO_TCP) &&
318 	    (is_frag == 0)) {
319 		if (m->m_len < iphlen + sizeof(struct tcphdr)) {
320 			RSS_DEBUG("short TCP frame?\n");
321 			return (-1);
322 		}
323 		th = (const struct tcphdr *)((c_caddr_t)ip + iphlen);
324 		return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
325 		    th->th_sport,
326 		    th->th_dport,
327 		    proto,
328 		    hashval,
329 		    hashtype);
330 	} else if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) &&
331 	    (proto == IPPROTO_UDP) &&
332 	    (is_frag == 0)) {
333 		uh = (const struct udphdr *)((c_caddr_t)ip + iphlen);
334 		if (m->m_len < iphlen + sizeof(struct udphdr)) {
335 			RSS_DEBUG("short UDP frame?\n");
336 			return (-1);
337 		}
338 		return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
339 		    uh->uh_sport,
340 		    uh->uh_dport,
341 		    proto,
342 		    hashval,
343 		    hashtype);
344 	} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) {
345 		/* Default to 2-tuple hash */
346 		return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
347 		    0,	/* source port */
348 		    0,	/* destination port */
349 		    0,	/* IPPROTO_IP */
350 		    hashval,
351 		    hashtype);
352 	} else {
353 		RSS_DEBUG("no available hashtypes!\n");
354 		return (-1);
355 	}
356 }
357 
358 /*
359  * Similar to rss_m2cpuid, but designed to be used by the IP NETISR
360  * on incoming frames.
361  *
362  * If an existing RSS hash exists and it matches what the configured
363  * hashing is, then use it.
364  *
365  * If there's an existing RSS hash but the desired hash is different,
366  * or if there's no useful RSS hash, then calculate it via
367  * the software path.
368  *
369  * XXX TODO: definitely want statistics here!
370  */
371 struct mbuf *
372 rss_soft_m2cpuid_v4(struct mbuf *m, uintptr_t source, u_int *cpuid)
373 {
374 	uint32_t hash_val, hash_type;
375 	int ret;
376 
377 	M_ASSERTPKTHDR(m);
378 
379 	ret = rss_mbuf_software_hash_v4(m, RSS_HASH_PKT_INGRESS,
380 	    &hash_val, &hash_type);
381 	if (ret > 0) {
382 		/* mbuf has a valid hash already; don't need to modify it */
383 		*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
384 	} else if (ret == 0) {
385 		/* hash was done; update */
386 		m->m_pkthdr.flowid = hash_val;
387 		M_HASHTYPE_SET(m, hash_type);
388 		*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
389 	} else { /* ret < 0 */
390 		/* no hash was done */
391 		*cpuid = NETISR_CPUID_NONE;
392 	}
393 	return (m);
394 }
395