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#include <net/toeplitz.h>
55
56/*-
57 * Operating system parts of receiver-side scaling (RSS), which allows
58 * network cards to direct flows to particular receive queues based on hashes
59 * of header tuples.  This implementation aligns RSS buckets with connection
60 * groups at the TCP/IP layer, so each bucket is associated with exactly one
61 * group.  As a result, the group lookup structures (and lock) should have an
62 * effective affinity with exactly one CPU.
63 *
64 * Network device drivers needing to configure RSS will query this framework
65 * for parameters, such as the current RSS key, hashing policies, number of
66 * bits, and indirection table mapping hashes to buckets and CPUs.  They may
67 * provide their own supplementary information, such as queue<->CPU bindings.
68 * It is the responsibility of the network device driver to inject packets
69 * into the stack on as close to the right CPU as possible, if playing by RSS
70 * rules.
71 *
72 * TODO:
73 *
74 * - Synchronization for rss_key and other future-configurable parameters.
75 * - Event handler drivers can register to pick up RSS configuration changes.
76 * - Should we allow rss_basecpu to be configured?
77 * - Randomize key on boot.
78 * - IPv6 support.
79 * - Statistics on how often there's a misalignment between hardware
80 *   placement and pcbgroup expectations.
81 */
82
83SYSCTL_DECL(_net_inet);
84SYSCTL_NODE(_net_inet, OID_AUTO, rss, CTLFLAG_RW, 0, "Receive-side steering");
85
86/*
87 * Toeplitz is the only required hash function in the RSS spec, so use it by
88 * default.
89 */
90static u_int	rss_hashalgo = RSS_HASH_TOEPLITZ;
91SYSCTL_INT(_net_inet_rss, OID_AUTO, hashalgo, CTLFLAG_RDTUN, &rss_hashalgo, 0,
92    "RSS hash algorithm");
93
94/*
95 * Size of the indirection table; at most 128 entries per the RSS spec.  We
96 * size it to at least 2 times the number of CPUs by default to allow useful
97 * rebalancing.  If not set explicitly with a loader tunable, we tune based
98 * on the number of CPUs present.
99 *
100 * XXXRW: buckets might be better to use for the tunable than bits.
101 */
102static u_int	rss_bits;
103SYSCTL_INT(_net_inet_rss, OID_AUTO, bits, CTLFLAG_RDTUN, &rss_bits, 0,
104    "RSS bits");
105
106static u_int	rss_mask;
107SYSCTL_INT(_net_inet_rss, OID_AUTO, mask, CTLFLAG_RD, &rss_mask, 0,
108    "RSS mask");
109
110static const u_int	rss_maxbits = RSS_MAXBITS;
111SYSCTL_INT(_net_inet_rss, OID_AUTO, maxbits, CTLFLAG_RD,
112    __DECONST(int *, &rss_maxbits), 0, "RSS maximum bits");
113
114/*
115 * RSS's own count of the number of CPUs it could be using for processing.
116 * Bounded to 64 by RSS constants.
117 */
118static u_int	rss_ncpus;
119SYSCTL_INT(_net_inet_rss, OID_AUTO, ncpus, CTLFLAG_RD, &rss_ncpus, 0,
120    "Number of CPUs available to RSS");
121
122#define	RSS_MAXCPUS	(1 << (RSS_MAXBITS - 1))
123static const u_int	rss_maxcpus = RSS_MAXCPUS;
124SYSCTL_INT(_net_inet_rss, OID_AUTO, maxcpus, CTLFLAG_RD,
125    __DECONST(int *, &rss_maxcpus), 0, "RSS maximum CPUs that can be used");
126
127/*
128 * Variable exists just for reporting rss_bits in a user-friendly way.
129 */
130static u_int	rss_buckets;
131SYSCTL_INT(_net_inet_rss, OID_AUTO, buckets, CTLFLAG_RD, &rss_buckets, 0,
132    "RSS buckets");
133
134/*
135 * Base CPU number; devices will add this to all CPU numbers returned by the
136 * RSS indirection table.  Currently unmodifable in FreeBSD.
137 */
138static const u_int	rss_basecpu;
139SYSCTL_INT(_net_inet_rss, OID_AUTO, basecpu, CTLFLAG_RD,
140    __DECONST(int *, &rss_basecpu), 0, "RSS base CPU");
141
142/*
143 * Print verbose debugging messages.
144 * 0 - disable
145 * non-zero - enable
146 */
147int	rss_debug = 0;
148SYSCTL_INT(_net_inet_rss, OID_AUTO, debug, CTLFLAG_RWTUN, &rss_debug, 0,
149    "RSS debug level");
150
151/*
152 * RSS secret key, intended to prevent attacks on load-balancing.  Its
153 * effectiveness may be limited by algorithm choice and available entropy
154 * during the boot.
155 *
156 * XXXRW: And that we don't randomize it yet!
157 *
158 * This is the default Microsoft RSS specification key which is also
159 * the Chelsio T5 firmware default key.
160 */
161static uint8_t rss_key[RSS_KEYSIZE] = {
162	0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
163	0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
164	0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
165	0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
166	0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa,
167};
168
169/*
170 * RSS hash->CPU table, which maps hashed packet headers to particular CPUs.
171 * Drivers may supplement this table with a separate CPU<->queue table when
172 * programming devices.
173 */
174struct rss_table_entry {
175	uint8_t		rte_cpu;	/* CPU affinity of bucket. */
176};
177static struct rss_table_entry	rss_table[RSS_TABLE_MAXLEN];
178
179static void
180rss_init(__unused void *arg)
181{
182	u_int i;
183	u_int cpuid;
184
185	/*
186	 * Validate tunables, coerce to sensible values.
187	 */
188	switch (rss_hashalgo) {
189	case RSS_HASH_TOEPLITZ:
190	case RSS_HASH_NAIVE:
191		break;
192
193	default:
194		RSS_DEBUG("invalid RSS hashalgo %u, coercing to %u\n",
195		    rss_hashalgo, RSS_HASH_TOEPLITZ);
196		rss_hashalgo = RSS_HASH_TOEPLITZ;
197	}
198
199	/*
200	 * Count available CPUs.
201	 *
202	 * XXXRW: Note incorrect assumptions regarding contiguity of this set
203	 * elsewhere.
204	 */
205	rss_ncpus = 0;
206	for (i = 0; i <= mp_maxid; i++) {
207		if (CPU_ABSENT(i))
208			continue;
209		rss_ncpus++;
210	}
211	if (rss_ncpus > RSS_MAXCPUS)
212		rss_ncpus = RSS_MAXCPUS;
213
214	/*
215	 * Tune RSS table entries to be no less than 2x the number of CPUs
216	 * -- unless we're running uniprocessor, in which case there's not
217	 * much point in having buckets to rearrange for load-balancing!
218	 */
219	if (rss_ncpus > 1) {
220		if (rss_bits == 0)
221			rss_bits = fls(rss_ncpus - 1) + 1;
222
223		/*
224		 * Microsoft limits RSS table entries to 128, so apply that
225		 * limit to both auto-detected CPU counts and user-configured
226		 * ones.
227		 */
228		if (rss_bits == 0 || rss_bits > RSS_MAXBITS) {
229			RSS_DEBUG("RSS bits %u not valid, coercing to %u\n",
230			    rss_bits, RSS_MAXBITS);
231			rss_bits = RSS_MAXBITS;
232		}
233
234		/*
235		 * Figure out how many buckets to use; warn if less than the
236		 * number of configured CPUs, although this is not a fatal
237		 * problem.
238		 */
239		rss_buckets = (1 << rss_bits);
240		if (rss_buckets < rss_ncpus)
241			RSS_DEBUG("WARNING: rss_buckets (%u) less than "
242			    "rss_ncpus (%u)\n", rss_buckets, rss_ncpus);
243		rss_mask = rss_buckets - 1;
244	} else {
245		rss_bits = 0;
246		rss_buckets = 1;
247		rss_mask = 0;
248	}
249
250	/*
251	 * Set up initial CPU assignments: round-robin by default.
252	 */
253	cpuid = CPU_FIRST();
254	for (i = 0; i < rss_buckets; i++) {
255		rss_table[i].rte_cpu = cpuid;
256		cpuid = CPU_NEXT(cpuid);
257	}
258
259	/*
260	 * Randomize rrs_key.
261	 *
262	 * XXXRW: Not yet.  If nothing else, will require an rss_isbadkey()
263	 * loop to check for "bad" RSS keys.
264	 */
265}
266SYSINIT(rss_init, SI_SUB_SOFTINTR, SI_ORDER_SECOND, rss_init, NULL);
267
268static uint32_t
269rss_naive_hash(u_int keylen, const uint8_t *key, u_int datalen,
270    const uint8_t *data)
271{
272	uint32_t v;
273	u_int i;
274
275	v = 0;
276	for (i = 0; i < keylen; i++)
277		v += key[i];
278	for (i = 0; i < datalen; i++)
279		v += data[i];
280	return (v);
281}
282
283uint32_t
284rss_hash(u_int datalen, const uint8_t *data)
285{
286
287	switch (rss_hashalgo) {
288	case RSS_HASH_TOEPLITZ:
289		return (toeplitz_hash(sizeof(rss_key), rss_key, datalen,
290		    data));
291
292	case RSS_HASH_NAIVE:
293		return (rss_naive_hash(sizeof(rss_key), rss_key, datalen,
294		    data));
295
296	default:
297		panic("%s: unsupported/unknown hashalgo %d", __func__,
298		    rss_hashalgo);
299	}
300}
301
302/*
303 * Query the number of RSS bits in use.
304 */
305u_int
306rss_getbits(void)
307{
308
309	return (rss_bits);
310}
311
312/*
313 * Query the RSS bucket associated with an RSS hash.
314 */
315u_int
316rss_getbucket(u_int hash)
317{
318
319	return (hash & rss_mask);
320}
321
322/*
323 * Query the RSS layer bucket associated with the given
324 * entry in the RSS hash space.
325 *
326 * The RSS indirection table is 0 .. rss_buckets-1,
327 * covering the low 'rss_bits' of the total 128 slot
328 * RSS indirection table.  So just mask off rss_bits and
329 * return that.
330 *
331 * NIC drivers can then iterate over the 128 slot RSS
332 * indirection table and fetch which RSS bucket to
333 * map it to.  This will typically be a CPU queue
334 */
335u_int
336rss_get_indirection_to_bucket(u_int index)
337{
338
339	return (index & rss_mask);
340}
341
342/*
343 * Query the RSS CPU associated with an RSS bucket.
344 */
345u_int
346rss_getcpu(u_int bucket)
347{
348
349	return (rss_table[bucket].rte_cpu);
350}
351
352/*
353 * netisr CPU affinity lookup given just the hash and hashtype.
354 */
355u_int
356rss_hash2cpuid(uint32_t hash_val, uint32_t hash_type)
357{
358
359	switch (hash_type) {
360	case M_HASHTYPE_RSS_IPV4:
361	case M_HASHTYPE_RSS_TCP_IPV4:
362	case M_HASHTYPE_RSS_UDP_IPV4:
363	case M_HASHTYPE_RSS_IPV6:
364	case M_HASHTYPE_RSS_TCP_IPV6:
365	case M_HASHTYPE_RSS_UDP_IPV6:
366		return (rss_getcpu(rss_getbucket(hash_val)));
367	default:
368		return (NETISR_CPUID_NONE);
369	}
370}
371
372/*
373 * Query the RSS bucket associated with the given hash value and
374 * type.
375 */
376int
377rss_hash2bucket(uint32_t hash_val, uint32_t hash_type, uint32_t *bucket_id)
378{
379
380	switch (hash_type) {
381	case M_HASHTYPE_RSS_IPV4:
382	case M_HASHTYPE_RSS_TCP_IPV4:
383	case M_HASHTYPE_RSS_UDP_IPV4:
384	case M_HASHTYPE_RSS_IPV6:
385	case M_HASHTYPE_RSS_TCP_IPV6:
386	case M_HASHTYPE_RSS_UDP_IPV6:
387		*bucket_id = rss_getbucket(hash_val);
388		return (0);
389	default:
390		return (-1);
391	}
392}
393
394/*
395 * netisr CPU affinity lookup routine for use by protocols.
396 */
397struct mbuf *
398rss_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
399{
400
401	M_ASSERTPKTHDR(m);
402	*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
403	return (m);
404}
405
406int
407rss_m2bucket(struct mbuf *m, uint32_t *bucket_id)
408{
409
410	M_ASSERTPKTHDR(m);
411
412	return(rss_hash2bucket(m->m_pkthdr.flowid, M_HASHTYPE_GET(m),
413	    bucket_id));
414}
415
416/*
417 * Query the RSS hash algorithm.
418 */
419u_int
420rss_gethashalgo(void)
421{
422
423	return (rss_hashalgo);
424}
425
426/*
427 * Query the current RSS key; likely to be used by device drivers when
428 * configuring hardware RSS.  Caller must pass an array of size RSS_KEYSIZE.
429 *
430 * XXXRW: Perhaps we should do the accept-a-length-and-truncate thing?
431 */
432void
433rss_getkey(uint8_t *key)
434{
435
436	bcopy(rss_key, key, sizeof(rss_key));
437}
438
439/*
440 * Query the number of buckets; this may be used by both network device
441 * drivers, which will need to populate hardware shadows of the software
442 * indirection table, and the network stack itself (such as when deciding how
443 * many connection groups to allocate).
444 */
445u_int
446rss_getnumbuckets(void)
447{
448
449	return (rss_buckets);
450}
451
452/*
453 * Query the number of CPUs in use by RSS; may be useful to device drivers
454 * trying to figure out how to map a larger number of CPUs into a smaller
455 * number of receive queues.
456 */
457u_int
458rss_getnumcpus(void)
459{
460
461	return (rss_ncpus);
462}
463
464/*
465 * Return the supported RSS hash configuration.
466 *
467 * NICs should query this to determine what to configure in their redirection
468 * matching table.
469 */
470inline u_int
471rss_gethashconfig(void)
472{
473
474	/* Return 4-tuple for TCP; 2-tuple for others */
475	/*
476	 * UDP may fragment more often than TCP and thus we'll end up with
477	 * NICs returning 2-tuple fragments.
478	 * udp_init() and udplite_init() both currently initialise things
479	 * as 2-tuple.
480	 * So for now disable UDP 4-tuple hashing until all of the other
481	 * pieces are in place.
482	 */
483	return (
484	    RSS_HASHTYPE_RSS_IPV4
485	|    RSS_HASHTYPE_RSS_TCP_IPV4
486	|    RSS_HASHTYPE_RSS_IPV6
487	|    RSS_HASHTYPE_RSS_TCP_IPV6
488	|    RSS_HASHTYPE_RSS_IPV6_EX
489	|    RSS_HASHTYPE_RSS_TCP_IPV6_EX
490#if 0
491	|    RSS_HASHTYPE_RSS_UDP_IPV4
492	|    RSS_HASHTYPE_RSS_UDP_IPV4_EX
493	|    RSS_HASHTYPE_RSS_UDP_IPV6
494	|    RSS_HASHTYPE_RSS_UDP_IPV6_EX
495#endif
496	);
497}
498
499/*
500 * XXXRW: Confirm that sysctl -a won't dump this keying material, don't want
501 * it appearing in debugging output unnecessarily.
502 */
503static int
504sysctl_rss_key(SYSCTL_HANDLER_ARGS)
505{
506	uint8_t temp_rss_key[RSS_KEYSIZE];
507	int error;
508
509	error = priv_check(req->td, PRIV_NETINET_HASHKEY);
510	if (error)
511		return (error);
512
513	bcopy(rss_key, temp_rss_key, sizeof(temp_rss_key));
514	error = sysctl_handle_opaque(oidp, temp_rss_key,
515	    sizeof(temp_rss_key), req);
516	if (error)
517		return (error);
518	if (req->newptr != NULL) {
519		/* XXXRW: Not yet. */
520		return (EINVAL);
521	}
522	return (0);
523}
524SYSCTL_PROC(_net_inet_rss, OID_AUTO, key,
525    CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_rss_key,
526    "", "RSS keying material");
527
528static int
529sysctl_rss_bucket_mapping(SYSCTL_HANDLER_ARGS)
530{
531	struct sbuf *sb;
532	int error;
533	int i;
534
535	error = 0;
536	error = sysctl_wire_old_buffer(req, 0);
537	if (error != 0)
538		return (error);
539	sb = sbuf_new_for_sysctl(NULL, NULL, 512, req);
540	if (sb == NULL)
541		return (ENOMEM);
542	for (i = 0; i < rss_buckets; i++) {
543		sbuf_printf(sb, "%s%d:%d", i == 0 ? "" : " ",
544		    i,
545		    rss_getcpu(i));
546	}
547	error = sbuf_finish(sb);
548	sbuf_delete(sb);
549
550	return (error);
551}
552SYSCTL_PROC(_net_inet_rss, OID_AUTO, bucket_mapping,
553    CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
554    sysctl_rss_bucket_mapping, "", "RSS bucket -> CPU mapping");
555