xref: /freebsd/sys/net/if_epair.c (revision fc76ddee9be0d7f98c9f9a162627950f8102964e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
42 
43 #include "opt_rss.h"
44 #include "opt_inet.h"
45 #include "opt_inet6.h"
46 
47 #include <sys/param.h>
48 #include <sys/bus.h>
49 #include <sys/hash.h>
50 #include <sys/interrupt.h>
51 #include <sys/jail.h>
52 #include <sys/kernel.h>
53 #include <sys/libkern.h>
54 #include <sys/malloc.h>
55 #include <sys/mbuf.h>
56 #include <sys/module.h>
57 #include <sys/proc.h>
58 #include <sys/queue.h>
59 #include <sys/sched.h>
60 #include <sys/smp.h>
61 #include <sys/socket.h>
62 #include <sys/sockio.h>
63 #include <sys/taskqueue.h>
64 
65 #include <net/bpf.h>
66 #include <net/ethernet.h>
67 #include <net/if.h>
68 #include <net/if_var.h>
69 #include <net/if_clone.h>
70 #include <net/if_media.h>
71 #include <net/if_var.h>
72 #include <net/if_private.h>
73 #include <net/if_types.h>
74 #include <net/netisr.h>
75 #ifdef RSS
76 #include <net/rss_config.h>
77 #ifdef INET
78 #include <netinet/in_rss.h>
79 #endif
80 #ifdef INET6
81 #include <netinet6/in6_rss.h>
82 #endif
83 #endif
84 #include <net/vnet.h>
85 
86 static const char epairname[] = "epair";
87 #define	RXRSIZE	4096	/* Probably overkill by 4-8x. */
88 
89 static MALLOC_DEFINE(M_EPAIR, epairname,
90     "Pair of virtual cross-over connected Ethernet-like interfaces");
91 
92 VNET_DEFINE_STATIC(struct if_clone *, epair_cloner);
93 #define	V_epair_cloner	VNET(epair_cloner)
94 
95 static unsigned int next_index = 0;
96 #define	EPAIR_LOCK_INIT()		mtx_init(&epair_n_index_mtx, "epairidx", \
97 					    NULL, MTX_DEF)
98 #define	EPAIR_LOCK_DESTROY()		mtx_destroy(&epair_n_index_mtx)
99 #define	EPAIR_LOCK()			mtx_lock(&epair_n_index_mtx)
100 #define	EPAIR_UNLOCK()			mtx_unlock(&epair_n_index_mtx)
101 
102 struct epair_softc;
103 struct epair_queue {
104 	struct mtx		 mtx;
105 	struct mbufq		 q;
106 	int			 id;
107 	enum {
108 		EPAIR_QUEUE_IDLE,
109 		EPAIR_QUEUE_WAKING,
110 		EPAIR_QUEUE_RUNNING,
111 	}			 state;
112 	struct task		 tx_task;
113 	struct epair_softc	*sc;
114 };
115 
116 static struct mtx epair_n_index_mtx;
117 struct epair_softc {
118 	struct ifnet		*ifp;		/* This ifp. */
119 	struct ifnet		*oifp;		/* other ifp of pair. */
120 	int			 num_queues;
121 	struct epair_queue	*queues;
122 	struct ifmedia		 media;		/* Media config (fake). */
123 	STAILQ_ENTRY(epair_softc) entry;
124 };
125 
126 struct epair_tasks_t {
127 	int			 tasks;
128 	struct taskqueue	 *tq[MAXCPU];
129 };
130 
131 static struct epair_tasks_t epair_tasks;
132 
133 static void
134 epair_clear_mbuf(struct mbuf *m)
135 {
136 	/* Remove any CSUM_SND_TAG as ether_input will barf. */
137 	if (m->m_pkthdr.csum_flags & CSUM_SND_TAG) {
138 		m_snd_tag_rele(m->m_pkthdr.snd_tag);
139 		m->m_pkthdr.snd_tag = NULL;
140 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
141 	}
142 
143 	m_tag_delete_nonpersistent(m);
144 }
145 
146 static void
147 epair_tx_start_deferred(void *arg, int pending)
148 {
149 	struct epair_queue *q = (struct epair_queue *)arg;
150 	if_t ifp;
151 	struct mbuf *m, *n;
152 	bool resched;
153 
154 	ifp = q->sc->ifp;
155 
156 	if_ref(ifp);
157 	CURVNET_SET(ifp->if_vnet);
158 
159 	mtx_lock(&q->mtx);
160 	m = mbufq_flush(&q->q);
161 	q->state = EPAIR_QUEUE_RUNNING;
162 	mtx_unlock(&q->mtx);
163 
164 	while (m != NULL) {
165 		n = STAILQ_NEXT(m, m_stailqpkt);
166 		m->m_nextpkt = NULL;
167 		if_input(ifp, m);
168 		m = n;
169 	}
170 
171 	/*
172 	 * Avoid flushing the queue more than once per task.  We can otherwise
173 	 * end up starving ourselves in a multi-epair routing configuration.
174 	 */
175 	mtx_lock(&q->mtx);
176 	if (mbufq_len(&q->q) > 0) {
177 		resched = true;
178 		q->state = EPAIR_QUEUE_WAKING;
179 	} else {
180 		resched = false;
181 		q->state = EPAIR_QUEUE_IDLE;
182 	}
183 	mtx_unlock(&q->mtx);
184 
185 	if (resched)
186 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
187 
188 	CURVNET_RESTORE();
189 	if_rele(ifp);
190 }
191 
192 static struct epair_queue *
193 epair_select_queue(struct epair_softc *sc, struct mbuf *m)
194 {
195 	uint32_t bucket;
196 #ifdef RSS
197 	struct ether_header *eh;
198 	int ret;
199 
200 	ret = rss_m2bucket(m, &bucket);
201 	if (ret) {
202 		/* Actually hash the packet. */
203 		eh = mtod(m, struct ether_header *);
204 
205 		switch (ntohs(eh->ether_type)) {
206 #ifdef INET
207 		case ETHERTYPE_IP:
208 			rss_soft_m2cpuid_v4(m, 0, &bucket);
209 			break;
210 #endif
211 #ifdef INET6
212 		case ETHERTYPE_IPV6:
213 			rss_soft_m2cpuid_v6(m, 0, &bucket);
214 			break;
215 #endif
216 		default:
217 			bucket = 0;
218 			break;
219 		}
220 	}
221 	bucket %= sc->num_queues;
222 #else
223 	bucket = 0;
224 #endif
225 	return (&sc->queues[bucket]);
226 }
227 
228 static void
229 epair_prepare_mbuf(struct mbuf *m, struct ifnet *src_ifp)
230 {
231 	M_ASSERTPKTHDR(m);
232 	epair_clear_mbuf(m);
233 	if_setrcvif(m, src_ifp);
234 	M_SETFIB(m, src_ifp->if_fib);
235 
236 	MPASS(m->m_nextpkt == NULL);
237 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
238 }
239 
240 static void
241 epair_menq(struct mbuf *m, struct epair_softc *osc)
242 {
243 	struct epair_queue *q;
244 	struct ifnet *ifp, *oifp;
245 	int error, len;
246 	bool mcast;
247 
248 	/*
249 	 * I know this looks weird. We pass the "other sc" as we need that one
250 	 * and can get both ifps from it as well.
251 	 */
252 	oifp = osc->ifp;
253 	ifp = osc->oifp;
254 
255 	epair_prepare_mbuf(m, oifp);
256 
257 	/* Save values as once the mbuf is queued, it's not ours anymore. */
258 	len = m->m_pkthdr.len;
259 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
260 
261 	q = epair_select_queue(osc, m);
262 
263 	mtx_lock(&q->mtx);
264 	if (q->state == EPAIR_QUEUE_IDLE) {
265 		q->state = EPAIR_QUEUE_WAKING;
266 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
267 	}
268 	error = mbufq_enqueue(&q->q, m);
269 	mtx_unlock(&q->mtx);
270 
271 	if (error != 0) {
272 		m_freem(m);
273 		if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
274 	} else {
275 		if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
276 		if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
277 		if (mcast)
278 			if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
279 		if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
280 	}
281 }
282 
283 static void
284 epair_start(struct ifnet *ifp)
285 {
286 	struct mbuf *m;
287 	struct epair_softc *sc;
288 	struct ifnet *oifp;
289 
290 	/*
291 	 * We get packets here from ether_output via if_handoff()
292 	 * and need to put them into the input queue of the oifp
293 	 * and will put the packet into the receive-queue (rxq) of the
294 	 * other interface (oifp) of our pair.
295 	 */
296 	sc = ifp->if_softc;
297 	oifp = sc->oifp;
298 	sc = oifp->if_softc;
299 	for (;;) {
300 		IFQ_DEQUEUE(&ifp->if_snd, m);
301 		if (m == NULL)
302 			break;
303 		M_ASSERTPKTHDR(m);
304 		BPF_MTAP(ifp, m);
305 
306 		/* In case either interface is not usable drop the packet. */
307 		if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
308 		    (ifp->if_flags & IFF_UP) == 0 ||
309 		    (oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
310 		    (oifp->if_flags & IFF_UP) == 0) {
311 			m_freem(m);
312 			continue;
313 		}
314 
315 		epair_menq(m, sc);
316 	}
317 }
318 
319 static int
320 epair_transmit(struct ifnet *ifp, struct mbuf *m)
321 {
322 	struct epair_softc *sc;
323 	struct ifnet *oifp;
324 #ifdef ALTQ
325 	int len;
326 	bool mcast;
327 #endif
328 
329 	if (m == NULL)
330 		return (0);
331 	M_ASSERTPKTHDR(m);
332 
333 	/*
334 	 * We are not going to use the interface en/dequeue mechanism
335 	 * on the TX side. We are called from ether_output_frame()
336 	 * and will put the packet into the receive-queue (rxq) of the
337 	 * other interface (oifp) of our pair.
338 	 */
339 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
340 		m_freem(m);
341 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
342 		return (ENXIO);
343 	}
344 	if ((ifp->if_flags & IFF_UP) == 0) {
345 		m_freem(m);
346 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
347 		return (ENETDOWN);
348 	}
349 
350 	BPF_MTAP(ifp, m);
351 
352 	/*
353 	 * In case the outgoing interface is not usable,
354 	 * drop the packet.
355 	 */
356 	sc = ifp->if_softc;
357 	oifp = sc->oifp;
358 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
359 	    (oifp->if_flags & IFF_UP) == 0) {
360 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
361 		m_freem(m);
362 		return (0);
363 	}
364 
365 #ifdef ALTQ
366 	len = m->m_pkthdr.len;
367 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
368 	int error = 0;
369 
370 	/* Support ALTQ via the classic if_start() path. */
371 	IF_LOCK(&ifp->if_snd);
372 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
373 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
374 		if (error)
375 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
376 		IF_UNLOCK(&ifp->if_snd);
377 		if (!error) {
378 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
379 			if (mcast)
380 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
381 			epair_start(ifp);
382 		}
383 		return (error);
384 	}
385 	IF_UNLOCK(&ifp->if_snd);
386 #endif
387 
388 	epair_menq(m, oifp->if_softc);
389 	return (0);
390 }
391 
392 static void
393 epair_qflush(struct ifnet *ifp __unused)
394 {
395 }
396 
397 static int
398 epair_media_change(struct ifnet *ifp __unused)
399 {
400 
401 	/* Do nothing. */
402 	return (0);
403 }
404 
405 static void
406 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
407 {
408 
409 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
410 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
411 }
412 
413 static int
414 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
415 {
416 	struct epair_softc *sc;
417 	struct ifreq *ifr;
418 	int error;
419 
420 	ifr = (struct ifreq *)data;
421 	switch (cmd) {
422 	case SIOCSIFFLAGS:
423 	case SIOCADDMULTI:
424 	case SIOCDELMULTI:
425 		error = 0;
426 		break;
427 
428 	case SIOCSIFMEDIA:
429 	case SIOCGIFMEDIA:
430 		sc = ifp->if_softc;
431 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
432 		break;
433 
434 	case SIOCSIFMTU:
435 		/* We basically allow all kinds of MTUs. */
436 		ifp->if_mtu = ifr->ifr_mtu;
437 		error = 0;
438 		break;
439 
440 	default:
441 		/* Let the common ethernet handler process this. */
442 		error = ether_ioctl(ifp, cmd, data);
443 		break;
444 	}
445 
446 	return (error);
447 }
448 
449 static void
450 epair_init(void *dummy __unused)
451 {
452 }
453 
454 /*
455  * Interface cloning functions.
456  * We use our private ones so that we can create/destroy our secondary
457  * device along with the primary one.
458  */
459 static int
460 epair_clone_match(struct if_clone *ifc, const char *name)
461 {
462 	const char *cp;
463 
464 	/*
465 	 * Our base name is epair.
466 	 * Our interfaces will be named epair<n>[ab].
467 	 * So accept anything of the following list:
468 	 * - epair
469 	 * - epair<n>
470 	 * but not the epair<n>[ab] versions.
471 	 */
472 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
473 		return (0);
474 
475 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
476 		if (*cp < '0' || *cp > '9')
477 			return (0);
478 	}
479 
480 	return (1);
481 }
482 
483 static void
484 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
485 {
486 	struct ifnet *ifp;
487 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
488 
489 	ifp = scb->ifp;
490 	/* Copy epairNa etheraddr and change the last byte. */
491 	memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
492 	eaddr[5] = 0x0b;
493 	ether_ifattach(ifp, eaddr);
494 
495 	if_clone_addif(ifc, ifp);
496 }
497 
498 static struct epair_softc *
499 epair_alloc_sc(struct if_clone *ifc)
500 {
501 	struct epair_softc *sc;
502 
503 	struct ifnet *ifp = if_alloc(IFT_ETHER);
504 	if (ifp == NULL)
505 		return (NULL);
506 
507 	sc = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
508 	sc->ifp = ifp;
509 	sc->num_queues = epair_tasks.tasks;
510 	sc->queues = mallocarray(sc->num_queues, sizeof(struct epair_queue),
511 	    M_EPAIR, M_WAITOK);
512 	for (int i = 0; i < sc->num_queues; i++) {
513 		struct epair_queue *q = &sc->queues[i];
514 		q->id = i;
515 		q->state = EPAIR_QUEUE_IDLE;
516 		mtx_init(&q->mtx, "epairq", NULL, MTX_DEF | MTX_NEW);
517 		mbufq_init(&q->q, RXRSIZE);
518 		q->sc = sc;
519 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
520 	}
521 
522 	/* Initialise pseudo media types. */
523 	ifmedia_init(&sc->media, 0, epair_media_change, epair_media_status);
524 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T, 0, NULL);
525 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T);
526 
527 	return (sc);
528 }
529 
530 static void
531 epair_setup_ifp(struct epair_softc *sc, char *name, int unit)
532 {
533 	struct ifnet *ifp = sc->ifp;
534 
535 	ifp->if_softc = sc;
536 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
537 	ifp->if_dname = epairname;
538 	ifp->if_dunit = unit;
539 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
540 	ifp->if_flags |= IFF_KNOWSEPOCH;
541 	ifp->if_capabilities = IFCAP_VLAN_MTU;
542 	ifp->if_capenable = IFCAP_VLAN_MTU;
543 	ifp->if_transmit = epair_transmit;
544 	ifp->if_qflush = epair_qflush;
545 	ifp->if_start = epair_start;
546 	ifp->if_ioctl = epair_ioctl;
547 	ifp->if_init  = epair_init;
548 	if_setsendqlen(ifp, ifqmaxlen);
549 	if_setsendqready(ifp);
550 
551 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
552 }
553 
554 static void
555 epair_generate_mac(struct epair_softc *sc, uint8_t *eaddr)
556 {
557 	uint32_t key[3];
558 	uint32_t hash;
559 	uint64_t hostid;
560 
561 	EPAIR_LOCK();
562 #ifdef SMP
563 	/* Get an approximate distribution. */
564 	hash = next_index % mp_ncpus;
565 #else
566 	hash = 0;
567 #endif
568 	EPAIR_UNLOCK();
569 
570 	/*
571 	 * Calculate the etheraddr hashing the hostid and the
572 	 * interface index. The result would be hopefully unique.
573 	 * Note that the "a" component of an epair instance may get moved
574 	 * to a different VNET after creation. In that case its index
575 	 * will be freed and the index can get reused by new epair instance.
576 	 * Make sure we do not create same etheraddr again.
577 	 */
578 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
579 	if (hostid == 0)
580 		arc4rand(&hostid, sizeof(hostid), 0);
581 
582 	struct ifnet *ifp = sc->ifp;
583 	EPAIR_LOCK();
584 	if (ifp->if_index > next_index)
585 		next_index = ifp->if_index;
586 	else
587 		next_index++;
588 
589 	key[0] = (uint32_t)next_index;
590 	EPAIR_UNLOCK();
591 	key[1] = (uint32_t)(hostid & 0xffffffff);
592 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
593 	hash = jenkins_hash32(key, 3, 0);
594 
595 	eaddr[0] = 0x02;
596 	memcpy(&eaddr[1], &hash, 4);
597 	eaddr[5] = 0x0a;
598 }
599 
600 static void
601 epair_free_sc(struct epair_softc *sc)
602 {
603 	if (sc == NULL)
604 		return;
605 
606 	if_free(sc->ifp);
607 	ifmedia_removeall(&sc->media);
608 	for (int i = 0; i < sc->num_queues; i++) {
609 		struct epair_queue *q = &sc->queues[i];
610 		mtx_destroy(&q->mtx);
611 	}
612 	free(sc->queues, M_EPAIR);
613 	free(sc, M_EPAIR);
614 }
615 
616 static void
617 epair_set_state(struct ifnet *ifp, bool running)
618 {
619 	if (running) {
620 		ifp->if_drv_flags |= IFF_DRV_RUNNING;
621 		if_link_state_change(ifp, LINK_STATE_UP);
622 	} else {
623 		if_link_state_change(ifp, LINK_STATE_DOWN);
624 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
625 	}
626 }
627 
628 static int
629 epair_handle_unit(struct if_clone *ifc, char *name, size_t len, int *punit)
630 {
631 	int error = 0, unit, wildcard;
632 	char *dp;
633 
634 	/* Try to see if a special unit was requested. */
635 	error = ifc_name2unit(name, &unit);
636 	if (error != 0)
637 		return (error);
638 	wildcard = (unit < 0);
639 
640 	error = ifc_alloc_unit(ifc, &unit);
641 	if (error != 0)
642 		return (error);
643 
644 	/*
645 	 * If no unit had been given, we need to adjust the ifName.
646 	 * Also make sure there is space for our extra [ab] suffix.
647 	 */
648 	for (dp = name; *dp != '\0'; dp++);
649 	if (wildcard) {
650 		int slen = snprintf(dp, len - (dp - name), "%d", unit);
651 		if (slen > len - (dp - name) - 1) {
652 			/* ifName too long. */
653 			error = ENOSPC;
654 			goto done;
655 		}
656 		dp += slen;
657 	}
658 	if (len - (dp - name) - 1 < 1) {
659 		/* No space left for our [ab] suffix. */
660 		error = ENOSPC;
661 		goto done;
662 	}
663 	*dp = 'b';
664 	/* Must not change dp so we can replace 'a' by 'b' later. */
665 	*(dp+1) = '\0';
666 
667 	/* Check if 'a' and 'b' interfaces already exist. */
668 	if (ifunit(name) != NULL) {
669 		error = EEXIST;
670 		goto done;
671 	}
672 
673 	*dp = 'a';
674 	if (ifunit(name) != NULL) {
675 		error = EEXIST;
676 		goto done;
677 	}
678 	*punit = unit;
679 done:
680 	if (error != 0)
681 		ifc_free_unit(ifc, unit);
682 
683 	return (error);
684 }
685 
686 static int
687 epair_clone_create(struct if_clone *ifc, char *name, size_t len,
688     struct ifc_data *ifd, struct ifnet **ifpp)
689 {
690 	struct epair_softc *sca, *scb;
691 	struct ifnet *ifp;
692 	char *dp;
693 	int error, unit;
694 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
695 
696 	error = epair_handle_unit(ifc, name, len, &unit);
697 	if (error != 0)
698 		return (error);
699 
700 	/* Allocate memory for both [ab] interfaces */
701 	sca = epair_alloc_sc(ifc);
702 	scb = epair_alloc_sc(ifc);
703 	if (sca == NULL || scb == NULL) {
704 		epair_free_sc(sca);
705 		epair_free_sc(scb);
706 		ifc_free_unit(ifc, unit);
707 		return (ENOSPC);
708 	}
709 
710 	/*
711 	 * Cross-reference the interfaces so we will be able to free both.
712 	 */
713 	sca->oifp = scb->ifp;
714 	scb->oifp = sca->ifp;
715 
716 	/* Finish initialization of interface <n>a. */
717 	ifp = sca->ifp;
718 	epair_setup_ifp(sca, name, unit);
719 	epair_generate_mac(sca, eaddr);
720 
721 	ether_ifattach(ifp, eaddr);
722 
723 	/* Swap the name and finish initialization of interface <n>b. */
724 	dp = name + strlen(name) - 1;
725 	*dp = 'b';
726 
727 	epair_setup_ifp(scb, name, unit);
728 
729 	ifp = scb->ifp;
730 	/* We need to play some tricks here for the second interface. */
731 	strlcpy(name, epairname, len);
732 	/* Correctly set the name for the cloner list. */
733 	strlcpy(name, scb->ifp->if_xname, len);
734 
735 	epair_clone_add(ifc, scb);
736 
737 	/*
738 	 * Restore name to <n>a as the ifp for this will go into the
739 	 * cloner list for the initial call.
740 	 */
741 	strlcpy(name, sca->ifp->if_xname, len);
742 
743 	/* Tell the world, that we are ready to rock. */
744 	epair_set_state(sca->ifp, true);
745 	epair_set_state(scb->ifp, true);
746 
747 	*ifpp = sca->ifp;
748 
749 	return (0);
750 }
751 
752 static void
753 epair_drain_rings(struct epair_softc *sc)
754 {
755 	for (int i = 0; i < sc->num_queues; i++) {
756 		struct epair_queue *q;
757 		struct mbuf *m, *n;
758 
759 		q = &sc->queues[i];
760 		mtx_lock(&q->mtx);
761 		m = mbufq_flush(&q->q);
762 		mtx_unlock(&q->mtx);
763 
764 		for (; m != NULL; m = n) {
765 			n = m->m_nextpkt;
766 			m_freem(m);
767 		}
768 	}
769 }
770 
771 static int
772 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp, uint32_t flags)
773 {
774 	struct ifnet *oifp;
775 	struct epair_softc *sca, *scb;
776 	int unit, error;
777 
778 	/*
779 	 * In case we called into if_clone_destroyif() ourselves
780 	 * again to remove the second interface, the softc will be
781 	 * NULL. In that case so not do anything but return success.
782 	 */
783 	if (ifp->if_softc == NULL)
784 		return (0);
785 
786 	unit = ifp->if_dunit;
787 	sca = ifp->if_softc;
788 	oifp = sca->oifp;
789 	scb = oifp->if_softc;
790 
791 	/* Frist get the interfaces down and detached. */
792 	epair_set_state(ifp, false);
793 	epair_set_state(oifp, false);
794 
795 	ether_ifdetach(ifp);
796 	ether_ifdetach(oifp);
797 
798 	/* Third free any queued packets and all the resources. */
799 	CURVNET_SET_QUIET(oifp->if_vnet);
800 	epair_drain_rings(scb);
801 	oifp->if_softc = NULL;
802 	error = if_clone_destroyif(ifc, oifp);
803 	if (error)
804 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
805 		    __func__, error);
806 	epair_free_sc(scb);
807 	CURVNET_RESTORE();
808 
809 	epair_drain_rings(sca);
810 	epair_free_sc(sca);
811 
812 	/* Last free the cloner unit. */
813 	ifc_free_unit(ifc, unit);
814 
815 	return (0);
816 }
817 
818 static void
819 vnet_epair_init(const void *unused __unused)
820 {
821 	struct if_clone_addreq req = {
822 		.match_f = epair_clone_match,
823 		.create_f = epair_clone_create,
824 		.destroy_f = epair_clone_destroy,
825 	};
826 	V_epair_cloner = ifc_attach_cloner(epairname, &req);
827 }
828 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
829     vnet_epair_init, NULL);
830 
831 static void
832 vnet_epair_uninit(const void *unused __unused)
833 {
834 
835 	ifc_detach_cloner(V_epair_cloner);
836 }
837 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
838     vnet_epair_uninit, NULL);
839 
840 static int
841 epair_mod_init(void)
842 {
843 	char name[32];
844 	epair_tasks.tasks = 0;
845 
846 #ifdef RSS
847 	int cpu;
848 
849 	CPU_FOREACH(cpu) {
850 		cpuset_t cpu_mask;
851 
852 		/* Pin to this CPU so we get appropriate NUMA allocations. */
853 		thread_lock(curthread);
854 		sched_bind(curthread, cpu);
855 		thread_unlock(curthread);
856 
857 		snprintf(name, sizeof(name), "epair_task_%d", cpu);
858 
859 		epair_tasks.tq[cpu] = taskqueue_create(name, M_WAITOK,
860 		    taskqueue_thread_enqueue,
861 		    &epair_tasks.tq[cpu]);
862 		CPU_SETOF(cpu, &cpu_mask);
863 		taskqueue_start_threads_cpuset(&epair_tasks.tq[cpu], 1, PI_NET,
864 		    &cpu_mask, "%s", name);
865 
866 		epair_tasks.tasks++;
867 	}
868 	thread_lock(curthread);
869 	sched_unbind(curthread);
870 	thread_unlock(curthread);
871 #else
872 	snprintf(name, sizeof(name), "epair_task");
873 
874 	epair_tasks.tq[0] = taskqueue_create(name, M_WAITOK,
875 	    taskqueue_thread_enqueue,
876 	    &epair_tasks.tq[0]);
877 	taskqueue_start_threads(&epair_tasks.tq[0], 1, PI_NET, "%s", name);
878 
879 	epair_tasks.tasks = 1;
880 #endif
881 
882 	return (0);
883 }
884 
885 static void
886 epair_mod_cleanup(void)
887 {
888 
889 	for (int i = 0; i < epair_tasks.tasks; i++) {
890 		taskqueue_drain_all(epair_tasks.tq[i]);
891 		taskqueue_free(epair_tasks.tq[i]);
892 	}
893 }
894 
895 static int
896 epair_modevent(module_t mod, int type, void *data)
897 {
898 	int ret;
899 
900 	switch (type) {
901 	case MOD_LOAD:
902 		EPAIR_LOCK_INIT();
903 		ret = epair_mod_init();
904 		if (ret != 0)
905 			return (ret);
906 		if (bootverbose)
907 			printf("%s: %s initialized.\n", __func__, epairname);
908 		break;
909 	case MOD_UNLOAD:
910 		epair_mod_cleanup();
911 		EPAIR_LOCK_DESTROY();
912 		if (bootverbose)
913 			printf("%s: %s unloaded.\n", __func__, epairname);
914 		break;
915 	default:
916 		return (EOPNOTSUPP);
917 	}
918 	return (0);
919 }
920 
921 static moduledata_t epair_mod = {
922 	"if_epair",
923 	epair_modevent,
924 	0
925 };
926 
927 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
928 MODULE_VERSION(if_epair, 3);
929