2 * Copyright (c) 2010-2011 Juniper Networks, Inc.
5 * This software was developed by Robert N. M. Watson under contract
6 * to Juniper Networks, Inc.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
30 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
34 #include "opt_inet6.h"
35 #include "opt_pcbgroup.h"
38 #error "options RSS depends on options PCBGROUP"
41 #include <sys/param.h>
43 #include <sys/socket.h>
45 #include <sys/kernel.h>
47 #include <sys/sysctl.h>
51 #include <net/if_var.h>
52 #include <net/netisr.h>
53 #include <net/rss_config.h>
54 #include <net/toeplitz.h>
57 #include <netinet/in.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/in_rss.h>
60 #include <netinet/in_var.h>
62 /* for software rss hash support */
63 #include <netinet/ip.h>
64 #include <netinet/tcp.h>
65 #include <netinet/udp.h>
69 * Operating system parts of receiver-side scaling (RSS), which allows
70 * network cards to direct flows to particular receive queues based on hashes
71 * of header tuples. This implementation aligns RSS buckets with connection
72 * groups at the TCP/IP layer, so each bucket is associated with exactly one
73 * group. As a result, the group lookup structures (and lock) should have an
74 * effective affinity with exactly one CPU.
76 * Network device drivers needing to configure RSS will query this framework
77 * for parameters, such as the current RSS key, hashing policies, number of
78 * bits, and indirection table mapping hashes to buckets and CPUs. They may
79 * provide their own supplementary information, such as queue<->CPU bindings.
80 * It is the responsibility of the network device driver to inject packets
81 * into the stack on as close to the right CPU as possible, if playing by RSS
86 * - Synchronization for rss_key and other future-configurable parameters.
87 * - Event handler drivers can register to pick up RSS configuration changes.
88 * - Should we allow rss_basecpu to be configured?
89 * - Randomize key on boot.
91 * - Statistics on how often there's a misalignment between hardware
92 * placement and pcbgroup expectations.
95 SYSCTL_DECL(_net_inet);
96 SYSCTL_NODE(_net_inet, OID_AUTO, rss, CTLFLAG_RW, 0, "Receive-side steering");
99 * Toeplitz is the only required hash function in the RSS spec, so use it by
102 static u_int rss_hashalgo = RSS_HASH_TOEPLITZ;
103 SYSCTL_INT(_net_inet_rss, OID_AUTO, hashalgo, CTLFLAG_RDTUN, &rss_hashalgo, 0,
104 "RSS hash algorithm");
107 * Size of the indirection table; at most 128 entries per the RSS spec. We
108 * size it to at least 2 times the number of CPUs by default to allow useful
109 * rebalancing. If not set explicitly with a loader tunable, we tune based
110 * on the number of CPUs present.
112 * XXXRW: buckets might be better to use for the tunable than bits.
114 static u_int rss_bits;
115 SYSCTL_INT(_net_inet_rss, OID_AUTO, bits, CTLFLAG_RDTUN, &rss_bits, 0,
118 static u_int rss_mask;
119 SYSCTL_INT(_net_inet_rss, OID_AUTO, mask, CTLFLAG_RD, &rss_mask, 0,
122 static const u_int rss_maxbits = RSS_MAXBITS;
123 SYSCTL_INT(_net_inet_rss, OID_AUTO, maxbits, CTLFLAG_RD,
124 __DECONST(int *, &rss_maxbits), 0, "RSS maximum bits");
127 * RSS's own count of the number of CPUs it could be using for processing.
128 * Bounded to 64 by RSS constants.
130 static u_int rss_ncpus;
131 SYSCTL_INT(_net_inet_rss, OID_AUTO, ncpus, CTLFLAG_RD, &rss_ncpus, 0,
132 "Number of CPUs available to RSS");
134 #define RSS_MAXCPUS (1 << (RSS_MAXBITS - 1))
135 static const u_int rss_maxcpus = RSS_MAXCPUS;
136 SYSCTL_INT(_net_inet_rss, OID_AUTO, maxcpus, CTLFLAG_RD,
137 __DECONST(int *, &rss_maxcpus), 0, "RSS maximum CPUs that can be used");
140 * Variable exists just for reporting rss_bits in a user-friendly way.
142 static u_int rss_buckets;
143 SYSCTL_INT(_net_inet_rss, OID_AUTO, buckets, CTLFLAG_RD, &rss_buckets, 0,
147 * Base CPU number; devices will add this to all CPU numbers returned by the
148 * RSS indirection table. Currently unmodifable in FreeBSD.
150 static const u_int rss_basecpu;
151 SYSCTL_INT(_net_inet_rss, OID_AUTO, basecpu, CTLFLAG_RD,
152 __DECONST(int *, &rss_basecpu), 0, "RSS base CPU");
155 * RSS secret key, intended to prevent attacks on load-balancing. Its
156 * effectiveness may be limited by algorithm choice and available entropy
159 * XXXRW: And that we don't randomize it yet!
161 * This is the default Microsoft RSS specification key which is also
162 * the Chelsio T5 firmware default key.
164 static uint8_t rss_key[RSS_KEYSIZE] = {
165 0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
166 0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
167 0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
168 0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
169 0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
173 * RSS hash->CPU table, which maps hashed packet headers to particular CPUs.
174 * Drivers may supplement this table with a seperate CPU<->queue table when
175 * programming devices.
177 struct rss_table_entry {
178 uint8_t rte_cpu; /* CPU affinity of bucket. */
180 static struct rss_table_entry rss_table[RSS_TABLE_MAXLEN];
183 rss_init(__unused void *arg)
189 * Validate tunables, coerce to sensible values.
191 switch (rss_hashalgo) {
192 case RSS_HASH_TOEPLITZ:
197 printf("%s: invalid RSS hashalgo %u, coercing to %u",
198 __func__, rss_hashalgo, RSS_HASH_TOEPLITZ);
199 rss_hashalgo = RSS_HASH_TOEPLITZ;
203 * Count available CPUs.
205 * XXXRW: Note incorrect assumptions regarding contiguity of this set
209 for (i = 0; i <= mp_maxid; i++) {
214 if (rss_ncpus > RSS_MAXCPUS)
215 rss_ncpus = RSS_MAXCPUS;
218 * Tune RSS table entries to be no less than 2x the number of CPUs
219 * -- unless we're running uniprocessor, in which case there's not
220 * much point in having buckets to rearrange for load-balancing!
224 rss_bits = fls(rss_ncpus - 1) + 1;
227 * Microsoft limits RSS table entries to 128, so apply that
228 * limit to both auto-detected CPU counts and user-configured
231 if (rss_bits == 0 || rss_bits > RSS_MAXBITS) {
232 printf("%s: RSS bits %u not valid, coercing to %u",
233 __func__, rss_bits, RSS_MAXBITS);
234 rss_bits = RSS_MAXBITS;
238 * Figure out how many buckets to use; warn if less than the
239 * number of configured CPUs, although this is not a fatal
242 rss_buckets = (1 << rss_bits);
243 if (rss_buckets < rss_ncpus)
244 printf("%s: WARNING: rss_buckets (%u) less than "
245 "rss_ncpus (%u)\n", __func__, rss_buckets,
247 rss_mask = rss_buckets - 1;
255 * Set up initial CPU assignments: round-robin by default.
258 for (i = 0; i < rss_buckets; i++) {
259 rss_table[i].rte_cpu = cpuid;
260 cpuid = CPU_NEXT(cpuid);
266 * XXXRW: Not yet. If nothing else, will require an rss_isbadkey()
267 * loop to check for "bad" RSS keys.
270 SYSINIT(rss_init, SI_SUB_SOFTINTR, SI_ORDER_SECOND, rss_init, NULL);
273 rss_naive_hash(u_int keylen, const uint8_t *key, u_int datalen,
280 for (i = 0; i < keylen; i++)
282 for (i = 0; i < datalen; i++)
288 rss_hash(u_int datalen, const uint8_t *data)
291 switch (rss_hashalgo) {
292 case RSS_HASH_TOEPLITZ:
293 return (toeplitz_hash(sizeof(rss_key), rss_key, datalen,
297 return (rss_naive_hash(sizeof(rss_key), rss_key, datalen,
301 panic("%s: unsupported/unknown hashalgo %d", __func__,
307 * Query the number of RSS bits in use.
317 * Query the RSS bucket associated with an RSS hash.
320 rss_getbucket(u_int hash)
323 return (hash & rss_mask);
327 * Query the RSS layer bucket associated with the given
328 * entry in the RSS hash space.
330 * The RSS indirection table is 0 .. rss_buckets-1,
331 * covering the low 'rss_bits' of the total 128 slot
332 * RSS indirection table. So just mask off rss_bits and
335 * NIC drivers can then iterate over the 128 slot RSS
336 * indirection table and fetch which RSS bucket to
337 * map it to. This will typically be a CPU queue
340 rss_get_indirection_to_bucket(u_int index)
343 return (index & rss_mask);
347 * Query the RSS CPU associated with an RSS bucket.
350 rss_getcpu(u_int bucket)
353 return (rss_table[bucket].rte_cpu);
357 * netisr CPU affinity lookup given just the hash and hashtype.
360 rss_hash2cpuid(uint32_t hash_val, uint32_t hash_type)
364 case M_HASHTYPE_RSS_IPV4:
365 case M_HASHTYPE_RSS_TCP_IPV4:
366 case M_HASHTYPE_RSS_UDP_IPV4:
367 case M_HASHTYPE_RSS_IPV6:
368 case M_HASHTYPE_RSS_TCP_IPV6:
369 case M_HASHTYPE_RSS_UDP_IPV6:
370 return (rss_getcpu(rss_getbucket(hash_val)));
372 return (NETISR_CPUID_NONE);
377 * Query the RSS bucket associated with the given hash value and
381 rss_hash2bucket(uint32_t hash_val, uint32_t hash_type, uint32_t *bucket_id)
385 case M_HASHTYPE_RSS_IPV4:
386 case M_HASHTYPE_RSS_TCP_IPV4:
387 case M_HASHTYPE_RSS_UDP_IPV4:
388 case M_HASHTYPE_RSS_IPV6:
389 case M_HASHTYPE_RSS_TCP_IPV6:
390 case M_HASHTYPE_RSS_UDP_IPV6:
391 *bucket_id = rss_getbucket(hash_val);
399 * netisr CPU affinity lookup routine for use by protocols.
402 rss_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
406 *cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
411 rss_m2bucket(struct mbuf *m, uint32_t *bucket_id)
416 return(rss_hash2bucket(m->m_pkthdr.flowid, M_HASHTYPE_GET(m),
421 * Query the RSS hash algorithm.
424 rss_gethashalgo(void)
427 return (rss_hashalgo);
431 * Query the current RSS key; likely to be used by device drivers when
432 * configuring hardware RSS. Caller must pass an array of size RSS_KEYSIZE.
434 * XXXRW: Perhaps we should do the accept-a-length-and-truncate thing?
437 rss_getkey(uint8_t *key)
440 bcopy(rss_key, key, sizeof(rss_key));
444 * Query the number of buckets; this may be used by both network device
445 * drivers, which will need to populate hardware shadows of the software
446 * indirection table, and the network stack itself (such as when deciding how
447 * many connection groups to allocate).
450 rss_getnumbuckets(void)
453 return (rss_buckets);
457 * Query the number of CPUs in use by RSS; may be useful to device drivers
458 * trying to figure out how to map a larger number of CPUs into a smaller
459 * number of receive queues.
469 * Return the supported RSS hash configuration.
471 * NICs should query this to determine what to configure in their redirection
475 rss_gethashconfig(void)
478 /* Return 4-tuple for TCP; 2-tuple for others */
480 * UDP may fragment more often than TCP and thus we'll end up with
481 * NICs returning 2-tuple fragments.
482 * udp_init() and udplite_init() both currently initialise things
484 * So for now disable UDP 4-tuple hashing until all of the other
485 * pieces are in place.
488 RSS_HASHTYPE_RSS_IPV4
489 | RSS_HASHTYPE_RSS_TCP_IPV4
490 | RSS_HASHTYPE_RSS_IPV6
491 | RSS_HASHTYPE_RSS_TCP_IPV6
492 | RSS_HASHTYPE_RSS_IPV6_EX
493 | RSS_HASHTYPE_RSS_TCP_IPV6_EX
495 | RSS_HASHTYPE_RSS_UDP_IPV4
496 | RSS_HASHTYPE_RSS_UDP_IPV4_EX
497 | RSS_HASHTYPE_RSS_UDP_IPV6
498 | RSS_HASHTYPE_RSS_UDP_IPV6_EX
504 * XXXRW: Confirm that sysctl -a won't dump this keying material, don't want
505 * it appearing in debugging output unnecessarily.
508 sysctl_rss_key(SYSCTL_HANDLER_ARGS)
510 uint8_t temp_rss_key[RSS_KEYSIZE];
513 error = priv_check(req->td, PRIV_NETINET_HASHKEY);
517 bcopy(rss_key, temp_rss_key, sizeof(temp_rss_key));
518 error = sysctl_handle_opaque(oidp, temp_rss_key,
519 sizeof(temp_rss_key), req);
522 if (req->newptr != NULL) {
523 /* XXXRW: Not yet. */
528 SYSCTL_PROC(_net_inet_rss, OID_AUTO, key,
529 CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_rss_key,
530 "", "RSS keying material");
533 sysctl_rss_bucket_mapping(SYSCTL_HANDLER_ARGS)
540 error = sysctl_wire_old_buffer(req, 0);
543 sb = sbuf_new_for_sysctl(NULL, NULL, 512, req);
546 for (i = 0; i < rss_buckets; i++) {
547 sbuf_printf(sb, "%s%d:%d", i == 0 ? "" : " ",
551 error = sbuf_finish(sb);
556 SYSCTL_PROC(_net_inet_rss, OID_AUTO, bucket_mapping,
557 CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
558 sysctl_rss_bucket_mapping, "", "RSS bucket -> CPU mapping");