xref: /freebsd/sys/net/if_epair.c (revision e4ea162509e400340a2bc3e755071a92f3465e2d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2008 The FreeBSD Foundation
5  * Copyright (c) 2009-2021 Bjoern A. Zeeb <bz@FreeBSD.org>
6  *
7  * This software was developed by CK Software GmbH under sponsorship
8  * from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  * notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  * notice, this list of conditions and the following disclaimer in the
17  * documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * A pair of virtual back-to-back connected ethernet like interfaces
34  * (``two interfaces with a virtual cross-over cable'').
35  *
36  * This is mostly intended to be used to provide connectivity between
37  * different virtual network stack instances.
38  */
39 
40 #include <sys/cdefs.h>
41 #include "opt_rss.h"
42 #include "opt_inet.h"
43 #include "opt_inet6.h"
44 
45 #include <sys/param.h>
46 #include <sys/bus.h>
47 #include <sys/hash.h>
48 #include <sys/interrupt.h>
49 #include <sys/jail.h>
50 #include <sys/kernel.h>
51 #include <sys/libkern.h>
52 #include <sys/malloc.h>
53 #include <sys/mbuf.h>
54 #include <sys/module.h>
55 #include <sys/proc.h>
56 #include <sys/queue.h>
57 #include <sys/sched.h>
58 #include <sys/smp.h>
59 #include <sys/socket.h>
60 #include <sys/sockio.h>
61 #include <sys/sysctl.h>
62 #include <sys/taskqueue.h>
63 
64 #include <net/bpf.h>
65 #include <net/ethernet.h>
66 #include <net/if.h>
67 #include <net/if_var.h>
68 #include <net/if_clone.h>
69 #include <net/if_media.h>
70 #include <net/if_private.h>
71 #include <net/if_types.h>
72 #include <net/if_vlan_var.h>
73 #include <net/netisr.h>
74 #ifdef RSS
75 #include <net/rss_config.h>
76 #ifdef INET
77 #include <netinet/in_rss.h>
78 #endif
79 #ifdef INET6
80 #include <netinet6/in6_rss.h>
81 #endif
82 #endif
83 #include <net/vnet.h>
84 
85 static const char epairname[] = "epair";
86 #define	RXRSIZE	4096	/* Probably overkill by 4-8x. */
87 
88 static MALLOC_DEFINE(M_EPAIR, epairname,
89     "Pair of virtual cross-over connected Ethernet-like interfaces");
90 
91 VNET_DEFINE_STATIC(struct if_clone *, epair_cloner);
92 #define	V_epair_cloner	VNET(epair_cloner)
93 
94 static unsigned int next_index = 0;
95 #define	EPAIR_LOCK_INIT()		mtx_init(&epair_n_index_mtx, "epairidx", \
96 					    NULL, MTX_DEF)
97 #define	EPAIR_LOCK_DESTROY()		mtx_destroy(&epair_n_index_mtx)
98 #define	EPAIR_LOCK()			mtx_lock(&epair_n_index_mtx)
99 #define	EPAIR_UNLOCK()			mtx_unlock(&epair_n_index_mtx)
100 
101 SYSCTL_DECL(_net_link);
102 static SYSCTL_NODE(_net_link, OID_AUTO, epair, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
103 	"Pair of virtual cross-over connected Ethernet-like interfaces");
104 
105 static bool use_ether_gen_addr = true;
106 SYSCTL_BOOL(_net_link_epair, OID_AUTO, ether_gen_addr, CTLFLAG_RWTUN,
107 	&use_ether_gen_addr, false,
108 	"Generate MAC with FreeBSD OUI using ether_gen_addr(9)");
109 
110 struct epair_softc;
111 struct epair_queue {
112 	struct mtx		 mtx;
113 	struct mbufq		 q;
114 	int			 id;
115 	enum {
116 		EPAIR_QUEUE_IDLE,
117 		EPAIR_QUEUE_WAKING,
118 		EPAIR_QUEUE_RUNNING,
119 	}			 state;
120 	struct task		 tx_task;
121 	struct epair_softc	*sc;
122 };
123 
124 static struct mtx epair_n_index_mtx;
125 struct epair_softc {
126 	struct ifnet		*ifp;		/* This ifp. */
127 	struct ifnet		*oifp;		/* other ifp of pair. */
128 	int			 num_queues;
129 	struct epair_queue	*queues;
130 	struct ifmedia		 media;		/* Media config (fake). */
131 	STAILQ_ENTRY(epair_softc) entry;
132 };
133 
134 struct epair_tasks_t {
135 	int			 tasks;
136 	struct taskqueue	 *tq[MAXCPU];
137 };
138 
139 static struct epair_tasks_t epair_tasks;
140 
141 static void
142 epair_clear_mbuf(struct mbuf *m)
143 {
144 	M_ASSERTPKTHDR(m);
145 
146 	/* Remove any CSUM_SND_TAG as ether_input will barf. */
147 	if (m->m_pkthdr.csum_flags & CSUM_SND_TAG) {
148 		m_snd_tag_rele(m->m_pkthdr.snd_tag);
149 		m->m_pkthdr.snd_tag = NULL;
150 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
151 	}
152 
153 	/* Clear vlan information. */
154 	m->m_flags &= ~M_VLANTAG;
155 	m->m_pkthdr.ether_vtag = 0;
156 
157 	m_tag_delete_nonpersistent(m);
158 }
159 
160 static void
161 epair_tx_start_deferred(void *arg, int pending)
162 {
163 	struct epair_queue *q = (struct epair_queue *)arg;
164 	if_t ifp;
165 	struct mbuf *m, *n;
166 	bool resched;
167 
168 	ifp = q->sc->ifp;
169 
170 	if_ref(ifp);
171 	CURVNET_SET(ifp->if_vnet);
172 
173 	mtx_lock(&q->mtx);
174 	m = mbufq_flush(&q->q);
175 	q->state = EPAIR_QUEUE_RUNNING;
176 	mtx_unlock(&q->mtx);
177 
178 	while (m != NULL) {
179 		n = STAILQ_NEXT(m, m_stailqpkt);
180 		m->m_nextpkt = NULL;
181 		if_input(ifp, m);
182 		m = n;
183 	}
184 
185 	/*
186 	 * Avoid flushing the queue more than once per task.  We can otherwise
187 	 * end up starving ourselves in a multi-epair routing configuration.
188 	 */
189 	mtx_lock(&q->mtx);
190 	if (!mbufq_empty(&q->q)) {
191 		resched = true;
192 		q->state = EPAIR_QUEUE_WAKING;
193 	} else {
194 		resched = false;
195 		q->state = EPAIR_QUEUE_IDLE;
196 	}
197 	mtx_unlock(&q->mtx);
198 
199 	if (resched)
200 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
201 
202 	CURVNET_RESTORE();
203 	if_rele(ifp);
204 }
205 
206 static struct epair_queue *
207 epair_select_queue(struct epair_softc *sc, struct mbuf *m)
208 {
209 	uint32_t bucket;
210 #ifdef RSS
211 	struct ether_header *eh;
212 	int ret;
213 
214 	ret = rss_m2bucket(m, &bucket);
215 	if (ret) {
216 		/* Actually hash the packet. */
217 		eh = mtod(m, struct ether_header *);
218 
219 		switch (ntohs(eh->ether_type)) {
220 #ifdef INET
221 		case ETHERTYPE_IP:
222 			rss_soft_m2cpuid_v4(m, 0, &bucket);
223 			break;
224 #endif
225 #ifdef INET6
226 		case ETHERTYPE_IPV6:
227 			rss_soft_m2cpuid_v6(m, 0, &bucket);
228 			break;
229 #endif
230 		default:
231 			bucket = 0;
232 			break;
233 		}
234 	}
235 	bucket %= sc->num_queues;
236 #else
237 	bucket = 0;
238 #endif
239 	return (&sc->queues[bucket]);
240 }
241 
242 static void
243 epair_prepare_mbuf(struct mbuf *m, struct ifnet *src_ifp)
244 {
245 	M_ASSERTPKTHDR(m);
246 	epair_clear_mbuf(m);
247 	if_setrcvif(m, src_ifp);
248 	M_SETFIB(m, src_ifp->if_fib);
249 
250 	MPASS(m->m_nextpkt == NULL);
251 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
252 }
253 
254 static void
255 epair_menq(struct mbuf *m, struct epair_softc *osc)
256 {
257 	struct epair_queue *q;
258 	struct ifnet *ifp, *oifp;
259 	int error, len;
260 	bool mcast;
261 
262 	/*
263 	 * I know this looks weird. We pass the "other sc" as we need that one
264 	 * and can get both ifps from it as well.
265 	 */
266 	oifp = osc->ifp;
267 	ifp = osc->oifp;
268 
269 	epair_prepare_mbuf(m, oifp);
270 
271 	/* Save values as once the mbuf is queued, it's not ours anymore. */
272 	len = m->m_pkthdr.len;
273 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
274 
275 	q = epair_select_queue(osc, m);
276 
277 	mtx_lock(&q->mtx);
278 	if (q->state == EPAIR_QUEUE_IDLE) {
279 		q->state = EPAIR_QUEUE_WAKING;
280 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
281 	}
282 	error = mbufq_enqueue(&q->q, m);
283 	mtx_unlock(&q->mtx);
284 
285 	if (error != 0) {
286 		m_freem(m);
287 		if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
288 	} else {
289 		if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
290 		if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
291 		if (mcast)
292 			if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
293 		if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
294 	}
295 }
296 
297 static void
298 epair_start(struct ifnet *ifp)
299 {
300 	struct mbuf *m;
301 	struct epair_softc *sc;
302 	struct ifnet *oifp;
303 
304 	/*
305 	 * We get packets here from ether_output via if_handoff()
306 	 * and need to put them into the input queue of the oifp
307 	 * and will put the packet into the receive-queue (rxq) of the
308 	 * other interface (oifp) of our pair.
309 	 */
310 	sc = ifp->if_softc;
311 	oifp = sc->oifp;
312 	sc = oifp->if_softc;
313 	for (;;) {
314 		IFQ_DEQUEUE(&ifp->if_snd, m);
315 		if (m == NULL)
316 			break;
317 		M_ASSERTPKTHDR(m);
318 		BPF_MTAP(ifp, m);
319 
320 		/* In case either interface is not usable drop the packet. */
321 		if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
322 		    (ifp->if_flags & IFF_UP) == 0 ||
323 		    (oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
324 		    (oifp->if_flags & IFF_UP) == 0) {
325 			m_freem(m);
326 			continue;
327 		}
328 
329 		epair_menq(m, sc);
330 	}
331 }
332 
333 static int
334 epair_transmit(struct ifnet *ifp, struct mbuf *m)
335 {
336 	struct epair_softc *sc;
337 	struct ifnet *oifp;
338 #ifdef ALTQ
339 	int len;
340 	bool mcast;
341 #endif
342 
343 	if (m == NULL)
344 		return (0);
345 	M_ASSERTPKTHDR(m);
346 
347 	/*
348 	 * We could just transmit this, but it makes testing easier if we're a
349 	 * little bit more like real hardware.
350 	 * Allow just that little bit extra for ethernet (and vlan) headers.
351 	 */
352 	if (m->m_pkthdr.len > (ifp->if_mtu + sizeof(struct ether_vlan_header))) {
353 		m_freem(m);
354 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
355 		return (E2BIG);
356 	}
357 
358 	/*
359 	 * We are not going to use the interface en/dequeue mechanism
360 	 * on the TX side. We are called from ether_output_frame()
361 	 * and will put the packet into the receive-queue (rxq) of the
362 	 * other interface (oifp) of our pair.
363 	 */
364 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
365 		m_freem(m);
366 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
367 		return (ENXIO);
368 	}
369 	if ((ifp->if_flags & IFF_UP) == 0) {
370 		m_freem(m);
371 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
372 		return (ENETDOWN);
373 	}
374 
375 	BPF_MTAP(ifp, m);
376 
377 	/*
378 	 * In case the outgoing interface is not usable,
379 	 * drop the packet.
380 	 */
381 	sc = ifp->if_softc;
382 	oifp = sc->oifp;
383 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
384 	    (oifp->if_flags & IFF_UP) == 0) {
385 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
386 		m_freem(m);
387 		return (0);
388 	}
389 
390 #ifdef ALTQ
391 	len = m->m_pkthdr.len;
392 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
393 	int error = 0;
394 
395 	/* Support ALTQ via the classic if_start() path. */
396 	IF_LOCK(&ifp->if_snd);
397 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
398 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
399 		if (error)
400 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
401 		IF_UNLOCK(&ifp->if_snd);
402 		if (!error) {
403 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
404 			if (mcast)
405 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
406 			epair_start(ifp);
407 		}
408 		return (error);
409 	}
410 	IF_UNLOCK(&ifp->if_snd);
411 #endif
412 
413 	epair_menq(m, oifp->if_softc);
414 	return (0);
415 }
416 
417 static void
418 epair_qflush(struct ifnet *ifp __unused)
419 {
420 }
421 
422 static int
423 epair_media_change(struct ifnet *ifp __unused)
424 {
425 
426 	/* Do nothing. */
427 	return (0);
428 }
429 
430 static void
431 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
432 {
433 
434 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
435 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
436 }
437 
438 /*
439  * Update ifp->if_hwassist according to the current value of ifp->if_capenable.
440  */
441 static void
442 epair_caps_changed(struct ifnet *ifp)
443 {
444 	uint64_t hwassist = 0;
445 
446 	if (ifp->if_capenable & IFCAP_TXCSUM)
447 		hwassist |= CSUM_IP_TCP | CSUM_IP_UDP;
448 	if (ifp->if_capenable & IFCAP_TXCSUM_IPV6)
449 		hwassist |= CSUM_IP6_TCP | CSUM_IP6_UDP;
450 	ifp->if_hwassist = hwassist;
451 }
452 
453 static int
454 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
455 {
456 	struct epair_softc *sc;
457 	struct ifreq *ifr;
458 	int error;
459 
460 	ifr = (struct ifreq *)data;
461 	switch (cmd) {
462 	case SIOCSIFFLAGS:
463 	case SIOCADDMULTI:
464 	case SIOCDELMULTI:
465 		error = 0;
466 		break;
467 
468 	case SIOCSIFMEDIA:
469 	case SIOCGIFMEDIA:
470 		sc = ifp->if_softc;
471 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
472 		break;
473 
474 	case SIOCSIFMTU:
475 		/* We basically allow all kinds of MTUs. */
476 		ifp->if_mtu = ifr->ifr_mtu;
477 		error = 0;
478 		break;
479 
480 	case SIOCGIFCAP:
481 		ifr->ifr_reqcap = ifp->if_capabilities;
482 		ifr->ifr_curcap = ifp->if_capenable;
483 		error = 0;
484 		break;
485 	case SIOCSIFCAP:
486 		/*
487 		 * Enable/disable capabilities as requested, besides
488 		 * IFCAP_RXCSUM(_IPV6), which always remain enabled.
489 		 * Incoming packets may have the mbuf flag CSUM_DATA_VALID set.
490 		 * Without IFCAP_RXCSUM(_IPV6), this flag would have to be
491 		 * removed, which does not seem helpful.
492 		 */
493 		ifp->if_capenable = ifr->ifr_reqcap | IFCAP_RXCSUM |
494 		    IFCAP_RXCSUM_IPV6;
495 		epair_caps_changed(ifp);
496 		/*
497 		 * If IFCAP_TXCSUM(_IPV6) has been changed, change it on the
498 		 * other epair interface as well.
499 		 * A bridge disables IFCAP_TXCSUM(_IPV6) when adding one epair
500 		 * interface if another interface in the bridge has it disabled.
501 		 * In that case this capability needs to be disabled on the
502 		 * other epair interface to avoid sending packets in the bridge
503 		 * that rely on this capability.
504 		 */
505 		sc = ifp->if_softc;
506 		if ((ifp->if_capenable ^ sc->oifp->if_capenable) &
507 		    (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6)) {
508 			sc->oifp->if_capenable &=
509 			    ~(IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
510 			sc->oifp->if_capenable |= ifp->if_capenable &
511 			    (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
512 			epair_caps_changed(sc->oifp);
513 		}
514 		VLAN_CAPABILITIES(ifp);
515 		error = 0;
516 		break;
517 
518 	default:
519 		/* Let the common ethernet handler process this. */
520 		error = ether_ioctl(ifp, cmd, data);
521 		break;
522 	}
523 
524 	return (error);
525 }
526 
527 static void
528 epair_init(void *dummy __unused)
529 {
530 }
531 
532 /*
533  * Interface cloning functions.
534  * We use our private ones so that we can create/destroy our secondary
535  * device along with the primary one.
536  */
537 static int
538 epair_clone_match(struct if_clone *ifc, const char *name)
539 {
540 	const char *cp;
541 
542 	/*
543 	 * Our base name is epair.
544 	 * Our interfaces will be named epair<n>[ab].
545 	 * So accept anything of the following list:
546 	 * - epair
547 	 * - epair<n>
548 	 * but not the epair<n>[ab] versions.
549 	 */
550 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
551 		return (0);
552 
553 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
554 		if (*cp < '0' || *cp > '9')
555 			return (0);
556 	}
557 
558 	return (1);
559 }
560 
561 static void
562 epair_generate_mac_byname(struct epair_softc *sc, uint8_t eaddr[])
563 {
564 	struct ether_addr gen_eaddr;
565 	int i;
566 
567 	ether_gen_addr_byname(if_name(sc->ifp), &gen_eaddr);
568 	for (i = 0; i < ETHER_ADDR_LEN; i++)
569 		eaddr[i] = gen_eaddr.octet[i];
570 }
571 
572 static void
573 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
574 {
575 	struct ifnet *ifp;
576 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
577 
578 	ifp = scb->ifp;
579 	if (!use_ether_gen_addr) {
580 		/* Copy epairNa etheraddr and change the last byte. */
581 		memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
582 		eaddr[5] = 0x0b;
583 	} else
584 		epair_generate_mac_byname(scb, eaddr);
585 	ether_ifattach(ifp, eaddr);
586 
587 	if_clone_addif(ifc, ifp);
588 }
589 
590 static struct epair_softc *
591 epair_alloc_sc(struct if_clone *ifc)
592 {
593 	struct epair_softc *sc;
594 
595 	struct ifnet *ifp = if_alloc(IFT_ETHER);
596 	sc = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
597 	sc->ifp = ifp;
598 	sc->num_queues = epair_tasks.tasks;
599 	sc->queues = mallocarray(sc->num_queues, sizeof(struct epair_queue),
600 	    M_EPAIR, M_WAITOK);
601 	for (int i = 0; i < sc->num_queues; i++) {
602 		struct epair_queue *q = &sc->queues[i];
603 		q->id = i;
604 		q->state = EPAIR_QUEUE_IDLE;
605 		mtx_init(&q->mtx, "epairq", NULL, MTX_DEF | MTX_NEW);
606 		mbufq_init(&q->q, RXRSIZE);
607 		q->sc = sc;
608 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
609 	}
610 
611 	/* Initialise pseudo media types. */
612 	ifmedia_init(&sc->media, 0, epair_media_change, epair_media_status);
613 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T, 0, NULL);
614 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T);
615 
616 	return (sc);
617 }
618 
619 static void
620 epair_setup_ifp(struct epair_softc *sc, char *name, int unit)
621 {
622 	struct ifnet *ifp = sc->ifp;
623 
624 	ifp->if_softc = sc;
625 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
626 	ifp->if_dname = epairname;
627 	ifp->if_dunit = unit;
628 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
629 	ifp->if_capabilities = IFCAP_VLAN_MTU | IFCAP_TXCSUM |
630 	    IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6;
631 	ifp->if_capenable = IFCAP_VLAN_MTU | IFCAP_TXCSUM |
632 	    IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6;
633 	epair_caps_changed(ifp);
634 	ifp->if_transmit = epair_transmit;
635 	ifp->if_qflush = epair_qflush;
636 	ifp->if_start = epair_start;
637 	ifp->if_ioctl = epair_ioctl;
638 	ifp->if_init  = epair_init;
639 	if_setsendqlen(ifp, ifqmaxlen);
640 	if_setsendqready(ifp);
641 
642 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
643 }
644 
645 static void
646 epair_generate_mac(struct epair_softc *sc, uint8_t *eaddr)
647 {
648 	uint32_t key[3];
649 	uint32_t hash;
650 	uint64_t hostid;
651 
652 	EPAIR_LOCK();
653 #ifdef SMP
654 	/* Get an approximate distribution. */
655 	hash = next_index % mp_ncpus;
656 #else
657 	hash = 0;
658 #endif
659 	EPAIR_UNLOCK();
660 
661 	/*
662 	 * Calculate the etheraddr hashing the hostid and the
663 	 * interface index. The result would be hopefully unique.
664 	 * Note that the "a" component of an epair instance may get moved
665 	 * to a different VNET after creation. In that case its index
666 	 * will be freed and the index can get reused by new epair instance.
667 	 * Make sure we do not create same etheraddr again.
668 	 */
669 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
670 	if (hostid == 0)
671 		arc4rand(&hostid, sizeof(hostid), 0);
672 
673 	struct ifnet *ifp = sc->ifp;
674 	EPAIR_LOCK();
675 	if (ifp->if_index > next_index)
676 		next_index = ifp->if_index;
677 	else
678 		next_index++;
679 
680 	key[0] = (uint32_t)next_index;
681 	EPAIR_UNLOCK();
682 	key[1] = (uint32_t)(hostid & 0xffffffff);
683 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
684 	hash = jenkins_hash32(key, 3, 0);
685 
686 	eaddr[0] = 0x02;
687 	memcpy(&eaddr[1], &hash, 4);
688 	eaddr[5] = 0x0a;
689 }
690 
691 static void
692 epair_free_sc(struct epair_softc *sc)
693 {
694 
695 	if_free(sc->ifp);
696 	ifmedia_removeall(&sc->media);
697 	for (int i = 0; i < sc->num_queues; i++) {
698 		struct epair_queue *q = &sc->queues[i];
699 		mtx_destroy(&q->mtx);
700 	}
701 	free(sc->queues, M_EPAIR);
702 	free(sc, M_EPAIR);
703 }
704 
705 static void
706 epair_set_state(struct ifnet *ifp, bool running)
707 {
708 	if (running) {
709 		ifp->if_drv_flags |= IFF_DRV_RUNNING;
710 		if_link_state_change(ifp, LINK_STATE_UP);
711 	} else {
712 		if_link_state_change(ifp, LINK_STATE_DOWN);
713 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
714 	}
715 }
716 
717 static int
718 epair_handle_unit(struct if_clone *ifc, char *name, size_t len, int *punit)
719 {
720 	int error = 0, unit, wildcard;
721 	char *dp;
722 
723 	/* Try to see if a special unit was requested. */
724 	error = ifc_name2unit(name, &unit);
725 	if (error != 0)
726 		return (error);
727 	wildcard = (unit < 0);
728 
729 	error = ifc_alloc_unit(ifc, &unit);
730 	if (error != 0)
731 		return (error);
732 
733 	/*
734 	 * If no unit had been given, we need to adjust the ifName.
735 	 * Also make sure there is space for our extra [ab] suffix.
736 	 */
737 	for (dp = name; *dp != '\0'; dp++);
738 	if (wildcard) {
739 		int slen = snprintf(dp, len - (dp - name), "%d", unit);
740 		if (slen > len - (dp - name) - 1) {
741 			/* ifName too long. */
742 			error = ENOSPC;
743 			goto done;
744 		}
745 		dp += slen;
746 	}
747 	if (len - (dp - name) - 1 < 1) {
748 		/* No space left for our [ab] suffix. */
749 		error = ENOSPC;
750 		goto done;
751 	}
752 	*dp = 'b';
753 	/* Must not change dp so we can replace 'a' by 'b' later. */
754 	*(dp+1) = '\0';
755 
756 	/* Check if 'a' and 'b' interfaces already exist. */
757 	if (ifunit(name) != NULL) {
758 		error = EEXIST;
759 		goto done;
760 	}
761 
762 	*dp = 'a';
763 	if (ifunit(name) != NULL) {
764 		error = EEXIST;
765 		goto done;
766 	}
767 	*punit = unit;
768 done:
769 	if (error != 0)
770 		ifc_free_unit(ifc, unit);
771 
772 	return (error);
773 }
774 
775 static int
776 epair_clone_create(struct if_clone *ifc, char *name, size_t len,
777     struct ifc_data *ifd, struct ifnet **ifpp)
778 {
779 	struct epair_softc *sca, *scb;
780 	struct ifnet *ifp;
781 	char *dp;
782 	int error, unit;
783 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
784 
785 	error = epair_handle_unit(ifc, name, len, &unit);
786 	if (error != 0)
787 		return (error);
788 
789 	/* Allocate memory for both [ab] interfaces */
790 	sca = epair_alloc_sc(ifc);
791 	scb = epair_alloc_sc(ifc);
792 
793 	/*
794 	 * Cross-reference the interfaces so we will be able to free both.
795 	 */
796 	sca->oifp = scb->ifp;
797 	scb->oifp = sca->ifp;
798 
799 	/* Finish initialization of interface <n>a. */
800 	ifp = sca->ifp;
801 	epair_setup_ifp(sca, name, unit);
802 	if (!use_ether_gen_addr)
803 		epair_generate_mac(sca, eaddr);
804 	else
805 		epair_generate_mac_byname(sca, eaddr);
806 
807 	ether_ifattach(ifp, eaddr);
808 
809 	/* Swap the name and finish initialization of interface <n>b. */
810 	dp = name + strlen(name) - 1;
811 	*dp = 'b';
812 
813 	epair_setup_ifp(scb, name, unit);
814 
815 	ifp = scb->ifp;
816 	/* We need to play some tricks here for the second interface. */
817 	strlcpy(name, epairname, len);
818 	/* Correctly set the name for the cloner list. */
819 	strlcpy(name, scb->ifp->if_xname, len);
820 
821 	epair_clone_add(ifc, scb);
822 
823 	/*
824 	 * Restore name to <n>a as the ifp for this will go into the
825 	 * cloner list for the initial call.
826 	 */
827 	strlcpy(name, sca->ifp->if_xname, len);
828 
829 	/* Tell the world, that we are ready to rock. */
830 	epair_set_state(sca->ifp, true);
831 	epair_set_state(scb->ifp, true);
832 
833 	*ifpp = sca->ifp;
834 
835 	return (0);
836 }
837 
838 static void
839 epair_drain_rings(struct epair_softc *sc)
840 {
841 	for (int i = 0; i < sc->num_queues; i++) {
842 		struct epair_queue *q;
843 		struct mbuf *m, *n;
844 
845 		q = &sc->queues[i];
846 		mtx_lock(&q->mtx);
847 		m = mbufq_flush(&q->q);
848 		mtx_unlock(&q->mtx);
849 
850 		for (; m != NULL; m = n) {
851 			n = m->m_nextpkt;
852 			m_freem(m);
853 		}
854 	}
855 }
856 
857 static int
858 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp, uint32_t flags)
859 {
860 	struct ifnet *oifp;
861 	struct epair_softc *sca, *scb;
862 	int unit, error;
863 
864 	/*
865 	 * In case we called into if_clone_destroyif() ourselves
866 	 * again to remove the second interface, the softc will be
867 	 * NULL. In that case so not do anything but return success.
868 	 */
869 	if (ifp->if_softc == NULL)
870 		return (0);
871 
872 	unit = ifp->if_dunit;
873 	sca = ifp->if_softc;
874 	oifp = sca->oifp;
875 	scb = oifp->if_softc;
876 
877 	/* Frist get the interfaces down and detached. */
878 	epair_set_state(ifp, false);
879 	epair_set_state(oifp, false);
880 
881 	ether_ifdetach(ifp);
882 	ether_ifdetach(oifp);
883 
884 	/* Third free any queued packets and all the resources. */
885 	CURVNET_SET_QUIET(oifp->if_vnet);
886 	epair_drain_rings(scb);
887 	oifp->if_softc = NULL;
888 	error = if_clone_destroyif(ifc, oifp);
889 	if (error)
890 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
891 		    __func__, error);
892 	epair_free_sc(scb);
893 	CURVNET_RESTORE();
894 
895 	epair_drain_rings(sca);
896 	epair_free_sc(sca);
897 
898 	/* Last free the cloner unit. */
899 	ifc_free_unit(ifc, unit);
900 
901 	return (0);
902 }
903 
904 static void
905 vnet_epair_init(const void *unused __unused)
906 {
907 	struct if_clone_addreq req = {
908 		.match_f = epair_clone_match,
909 		.create_f = epair_clone_create,
910 		.destroy_f = epair_clone_destroy,
911 	};
912 	V_epair_cloner = ifc_attach_cloner(epairname, &req);
913 }
914 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
915     vnet_epair_init, NULL);
916 
917 static void
918 vnet_epair_uninit(const void *unused __unused)
919 {
920 
921 	ifc_detach_cloner(V_epair_cloner);
922 }
923 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
924     vnet_epair_uninit, NULL);
925 
926 static int
927 epair_mod_init(void)
928 {
929 	char name[32];
930 	epair_tasks.tasks = 0;
931 
932 #ifdef RSS
933 	int cpu;
934 
935 	CPU_FOREACH(cpu) {
936 		cpuset_t cpu_mask;
937 
938 		/* Pin to this CPU so we get appropriate NUMA allocations. */
939 		thread_lock(curthread);
940 		sched_bind(curthread, cpu);
941 		thread_unlock(curthread);
942 
943 		snprintf(name, sizeof(name), "epair_task_%d", cpu);
944 
945 		epair_tasks.tq[cpu] = taskqueue_create(name, M_WAITOK,
946 		    taskqueue_thread_enqueue,
947 		    &epair_tasks.tq[cpu]);
948 		CPU_SETOF(cpu, &cpu_mask);
949 		taskqueue_start_threads_cpuset(&epair_tasks.tq[cpu], 1, PI_NET,
950 		    &cpu_mask, "%s", name);
951 
952 		epair_tasks.tasks++;
953 	}
954 	thread_lock(curthread);
955 	sched_unbind(curthread);
956 	thread_unlock(curthread);
957 #else
958 	snprintf(name, sizeof(name), "epair_task");
959 
960 	epair_tasks.tq[0] = taskqueue_create(name, M_WAITOK,
961 	    taskqueue_thread_enqueue,
962 	    &epair_tasks.tq[0]);
963 	taskqueue_start_threads(&epair_tasks.tq[0], 1, PI_NET, "%s", name);
964 
965 	epair_tasks.tasks = 1;
966 #endif
967 
968 	return (0);
969 }
970 
971 static void
972 epair_mod_cleanup(void)
973 {
974 
975 	for (int i = 0; i < epair_tasks.tasks; i++) {
976 		taskqueue_drain_all(epair_tasks.tq[i]);
977 		taskqueue_free(epair_tasks.tq[i]);
978 	}
979 }
980 
981 static int
982 epair_modevent(module_t mod, int type, void *data)
983 {
984 	int ret;
985 
986 	switch (type) {
987 	case MOD_LOAD:
988 		EPAIR_LOCK_INIT();
989 		ret = epair_mod_init();
990 		if (ret != 0)
991 			return (ret);
992 		if (bootverbose)
993 			printf("%s: %s initialized.\n", __func__, epairname);
994 		break;
995 	case MOD_UNLOAD:
996 		epair_mod_cleanup();
997 		EPAIR_LOCK_DESTROY();
998 		if (bootverbose)
999 			printf("%s: %s unloaded.\n", __func__, epairname);
1000 		break;
1001 	default:
1002 		return (EOPNOTSUPP);
1003 	}
1004 	return (0);
1005 }
1006 
1007 static moduledata_t epair_mod = {
1008 	"if_epair",
1009 	epair_modevent,
1010 	0
1011 };
1012 
1013 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
1014 MODULE_VERSION(if_epair, 3);
1015