xref: /freebsd/sys/net/if_epair.c (revision f1c4c3daccbaf3820f0e2224de53df12fc952fcc)
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_var.h>
71 #include <net/if_private.h>
72 #include <net/if_types.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 static int
439 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
440 {
441 	struct epair_softc *sc;
442 	struct ifreq *ifr;
443 	int error;
444 
445 	ifr = (struct ifreq *)data;
446 	switch (cmd) {
447 	case SIOCSIFFLAGS:
448 	case SIOCADDMULTI:
449 	case SIOCDELMULTI:
450 		error = 0;
451 		break;
452 
453 	case SIOCSIFMEDIA:
454 	case SIOCGIFMEDIA:
455 		sc = ifp->if_softc;
456 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
457 		break;
458 
459 	case SIOCSIFMTU:
460 		/* We basically allow all kinds of MTUs. */
461 		ifp->if_mtu = ifr->ifr_mtu;
462 		error = 0;
463 		break;
464 
465 	default:
466 		/* Let the common ethernet handler process this. */
467 		error = ether_ioctl(ifp, cmd, data);
468 		break;
469 	}
470 
471 	return (error);
472 }
473 
474 static void
475 epair_init(void *dummy __unused)
476 {
477 }
478 
479 /*
480  * Interface cloning functions.
481  * We use our private ones so that we can create/destroy our secondary
482  * device along with the primary one.
483  */
484 static int
485 epair_clone_match(struct if_clone *ifc, const char *name)
486 {
487 	const char *cp;
488 
489 	/*
490 	 * Our base name is epair.
491 	 * Our interfaces will be named epair<n>[ab].
492 	 * So accept anything of the following list:
493 	 * - epair
494 	 * - epair<n>
495 	 * but not the epair<n>[ab] versions.
496 	 */
497 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
498 		return (0);
499 
500 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
501 		if (*cp < '0' || *cp > '9')
502 			return (0);
503 	}
504 
505 	return (1);
506 }
507 
508 static void
509 epair_generate_mac_byname(struct epair_softc *sc, uint8_t eaddr[])
510 {
511 	struct ether_addr gen_eaddr;
512 	int i;
513 
514 	ether_gen_addr_byname(if_name(sc->ifp), &gen_eaddr);
515 	for (i = 0; i < ETHER_ADDR_LEN; i++)
516 		eaddr[i] = gen_eaddr.octet[i];
517 }
518 
519 static void
520 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
521 {
522 	struct ifnet *ifp;
523 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
524 
525 	ifp = scb->ifp;
526 	if (!use_ether_gen_addr) {
527 		/* Copy epairNa etheraddr and change the last byte. */
528 		memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
529 		eaddr[5] = 0x0b;
530 	} else
531 		epair_generate_mac_byname(scb, eaddr);
532 	ether_ifattach(ifp, eaddr);
533 
534 	if_clone_addif(ifc, ifp);
535 }
536 
537 static struct epair_softc *
538 epair_alloc_sc(struct if_clone *ifc)
539 {
540 	struct epair_softc *sc;
541 
542 	struct ifnet *ifp = if_alloc(IFT_ETHER);
543 	sc = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
544 	sc->ifp = ifp;
545 	sc->num_queues = epair_tasks.tasks;
546 	sc->queues = mallocarray(sc->num_queues, sizeof(struct epair_queue),
547 	    M_EPAIR, M_WAITOK);
548 	for (int i = 0; i < sc->num_queues; i++) {
549 		struct epair_queue *q = &sc->queues[i];
550 		q->id = i;
551 		q->state = EPAIR_QUEUE_IDLE;
552 		mtx_init(&q->mtx, "epairq", NULL, MTX_DEF | MTX_NEW);
553 		mbufq_init(&q->q, RXRSIZE);
554 		q->sc = sc;
555 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
556 	}
557 
558 	/* Initialise pseudo media types. */
559 	ifmedia_init(&sc->media, 0, epair_media_change, epair_media_status);
560 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T, 0, NULL);
561 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T);
562 
563 	return (sc);
564 }
565 
566 static void
567 epair_setup_ifp(struct epair_softc *sc, char *name, int unit)
568 {
569 	struct ifnet *ifp = sc->ifp;
570 
571 	ifp->if_softc = sc;
572 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
573 	ifp->if_dname = epairname;
574 	ifp->if_dunit = unit;
575 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
576 	ifp->if_capabilities = IFCAP_VLAN_MTU;
577 	ifp->if_capenable = IFCAP_VLAN_MTU;
578 	ifp->if_transmit = epair_transmit;
579 	ifp->if_qflush = epair_qflush;
580 	ifp->if_start = epair_start;
581 	ifp->if_ioctl = epair_ioctl;
582 	ifp->if_init  = epair_init;
583 	if_setsendqlen(ifp, ifqmaxlen);
584 	if_setsendqready(ifp);
585 
586 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
587 }
588 
589 static void
590 epair_generate_mac(struct epair_softc *sc, uint8_t *eaddr)
591 {
592 	uint32_t key[3];
593 	uint32_t hash;
594 	uint64_t hostid;
595 
596 	EPAIR_LOCK();
597 #ifdef SMP
598 	/* Get an approximate distribution. */
599 	hash = next_index % mp_ncpus;
600 #else
601 	hash = 0;
602 #endif
603 	EPAIR_UNLOCK();
604 
605 	/*
606 	 * Calculate the etheraddr hashing the hostid and the
607 	 * interface index. The result would be hopefully unique.
608 	 * Note that the "a" component of an epair instance may get moved
609 	 * to a different VNET after creation. In that case its index
610 	 * will be freed and the index can get reused by new epair instance.
611 	 * Make sure we do not create same etheraddr again.
612 	 */
613 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
614 	if (hostid == 0)
615 		arc4rand(&hostid, sizeof(hostid), 0);
616 
617 	struct ifnet *ifp = sc->ifp;
618 	EPAIR_LOCK();
619 	if (ifp->if_index > next_index)
620 		next_index = ifp->if_index;
621 	else
622 		next_index++;
623 
624 	key[0] = (uint32_t)next_index;
625 	EPAIR_UNLOCK();
626 	key[1] = (uint32_t)(hostid & 0xffffffff);
627 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
628 	hash = jenkins_hash32(key, 3, 0);
629 
630 	eaddr[0] = 0x02;
631 	memcpy(&eaddr[1], &hash, 4);
632 	eaddr[5] = 0x0a;
633 }
634 
635 static void
636 epair_free_sc(struct epair_softc *sc)
637 {
638 
639 	if_free(sc->ifp);
640 	ifmedia_removeall(&sc->media);
641 	for (int i = 0; i < sc->num_queues; i++) {
642 		struct epair_queue *q = &sc->queues[i];
643 		mtx_destroy(&q->mtx);
644 	}
645 	free(sc->queues, M_EPAIR);
646 	free(sc, M_EPAIR);
647 }
648 
649 static void
650 epair_set_state(struct ifnet *ifp, bool running)
651 {
652 	if (running) {
653 		ifp->if_drv_flags |= IFF_DRV_RUNNING;
654 		if_link_state_change(ifp, LINK_STATE_UP);
655 	} else {
656 		if_link_state_change(ifp, LINK_STATE_DOWN);
657 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
658 	}
659 }
660 
661 static int
662 epair_handle_unit(struct if_clone *ifc, char *name, size_t len, int *punit)
663 {
664 	int error = 0, unit, wildcard;
665 	char *dp;
666 
667 	/* Try to see if a special unit was requested. */
668 	error = ifc_name2unit(name, &unit);
669 	if (error != 0)
670 		return (error);
671 	wildcard = (unit < 0);
672 
673 	error = ifc_alloc_unit(ifc, &unit);
674 	if (error != 0)
675 		return (error);
676 
677 	/*
678 	 * If no unit had been given, we need to adjust the ifName.
679 	 * Also make sure there is space for our extra [ab] suffix.
680 	 */
681 	for (dp = name; *dp != '\0'; dp++);
682 	if (wildcard) {
683 		int slen = snprintf(dp, len - (dp - name), "%d", unit);
684 		if (slen > len - (dp - name) - 1) {
685 			/* ifName too long. */
686 			error = ENOSPC;
687 			goto done;
688 		}
689 		dp += slen;
690 	}
691 	if (len - (dp - name) - 1 < 1) {
692 		/* No space left for our [ab] suffix. */
693 		error = ENOSPC;
694 		goto done;
695 	}
696 	*dp = 'b';
697 	/* Must not change dp so we can replace 'a' by 'b' later. */
698 	*(dp+1) = '\0';
699 
700 	/* Check if 'a' and 'b' interfaces already exist. */
701 	if (ifunit(name) != NULL) {
702 		error = EEXIST;
703 		goto done;
704 	}
705 
706 	*dp = 'a';
707 	if (ifunit(name) != NULL) {
708 		error = EEXIST;
709 		goto done;
710 	}
711 	*punit = unit;
712 done:
713 	if (error != 0)
714 		ifc_free_unit(ifc, unit);
715 
716 	return (error);
717 }
718 
719 static int
720 epair_clone_create(struct if_clone *ifc, char *name, size_t len,
721     struct ifc_data *ifd, struct ifnet **ifpp)
722 {
723 	struct epair_softc *sca, *scb;
724 	struct ifnet *ifp;
725 	char *dp;
726 	int error, unit;
727 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
728 
729 	error = epair_handle_unit(ifc, name, len, &unit);
730 	if (error != 0)
731 		return (error);
732 
733 	/* Allocate memory for both [ab] interfaces */
734 	sca = epair_alloc_sc(ifc);
735 	scb = epair_alloc_sc(ifc);
736 
737 	/*
738 	 * Cross-reference the interfaces so we will be able to free both.
739 	 */
740 	sca->oifp = scb->ifp;
741 	scb->oifp = sca->ifp;
742 
743 	/* Finish initialization of interface <n>a. */
744 	ifp = sca->ifp;
745 	epair_setup_ifp(sca, name, unit);
746 	if (!use_ether_gen_addr)
747 		epair_generate_mac(sca, eaddr);
748 	else
749 		epair_generate_mac_byname(sca, eaddr);
750 
751 	ether_ifattach(ifp, eaddr);
752 
753 	/* Swap the name and finish initialization of interface <n>b. */
754 	dp = name + strlen(name) - 1;
755 	*dp = 'b';
756 
757 	epair_setup_ifp(scb, name, unit);
758 
759 	ifp = scb->ifp;
760 	/* We need to play some tricks here for the second interface. */
761 	strlcpy(name, epairname, len);
762 	/* Correctly set the name for the cloner list. */
763 	strlcpy(name, scb->ifp->if_xname, len);
764 
765 	epair_clone_add(ifc, scb);
766 
767 	/*
768 	 * Restore name to <n>a as the ifp for this will go into the
769 	 * cloner list for the initial call.
770 	 */
771 	strlcpy(name, sca->ifp->if_xname, len);
772 
773 	/* Tell the world, that we are ready to rock. */
774 	epair_set_state(sca->ifp, true);
775 	epair_set_state(scb->ifp, true);
776 
777 	*ifpp = sca->ifp;
778 
779 	return (0);
780 }
781 
782 static void
783 epair_drain_rings(struct epair_softc *sc)
784 {
785 	for (int i = 0; i < sc->num_queues; i++) {
786 		struct epair_queue *q;
787 		struct mbuf *m, *n;
788 
789 		q = &sc->queues[i];
790 		mtx_lock(&q->mtx);
791 		m = mbufq_flush(&q->q);
792 		mtx_unlock(&q->mtx);
793 
794 		for (; m != NULL; m = n) {
795 			n = m->m_nextpkt;
796 			m_freem(m);
797 		}
798 	}
799 }
800 
801 static int
802 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp, uint32_t flags)
803 {
804 	struct ifnet *oifp;
805 	struct epair_softc *sca, *scb;
806 	int unit, error;
807 
808 	/*
809 	 * In case we called into if_clone_destroyif() ourselves
810 	 * again to remove the second interface, the softc will be
811 	 * NULL. In that case so not do anything but return success.
812 	 */
813 	if (ifp->if_softc == NULL)
814 		return (0);
815 
816 	unit = ifp->if_dunit;
817 	sca = ifp->if_softc;
818 	oifp = sca->oifp;
819 	scb = oifp->if_softc;
820 
821 	/* Frist get the interfaces down and detached. */
822 	epair_set_state(ifp, false);
823 	epair_set_state(oifp, false);
824 
825 	ether_ifdetach(ifp);
826 	ether_ifdetach(oifp);
827 
828 	/* Third free any queued packets and all the resources. */
829 	CURVNET_SET_QUIET(oifp->if_vnet);
830 	epair_drain_rings(scb);
831 	oifp->if_softc = NULL;
832 	error = if_clone_destroyif(ifc, oifp);
833 	if (error)
834 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
835 		    __func__, error);
836 	epair_free_sc(scb);
837 	CURVNET_RESTORE();
838 
839 	epair_drain_rings(sca);
840 	epair_free_sc(sca);
841 
842 	/* Last free the cloner unit. */
843 	ifc_free_unit(ifc, unit);
844 
845 	return (0);
846 }
847 
848 static void
849 vnet_epair_init(const void *unused __unused)
850 {
851 	struct if_clone_addreq req = {
852 		.match_f = epair_clone_match,
853 		.create_f = epair_clone_create,
854 		.destroy_f = epair_clone_destroy,
855 	};
856 	V_epair_cloner = ifc_attach_cloner(epairname, &req);
857 }
858 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
859     vnet_epair_init, NULL);
860 
861 static void
862 vnet_epair_uninit(const void *unused __unused)
863 {
864 
865 	ifc_detach_cloner(V_epair_cloner);
866 }
867 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
868     vnet_epair_uninit, NULL);
869 
870 static int
871 epair_mod_init(void)
872 {
873 	char name[32];
874 	epair_tasks.tasks = 0;
875 
876 #ifdef RSS
877 	int cpu;
878 
879 	CPU_FOREACH(cpu) {
880 		cpuset_t cpu_mask;
881 
882 		/* Pin to this CPU so we get appropriate NUMA allocations. */
883 		thread_lock(curthread);
884 		sched_bind(curthread, cpu);
885 		thread_unlock(curthread);
886 
887 		snprintf(name, sizeof(name), "epair_task_%d", cpu);
888 
889 		epair_tasks.tq[cpu] = taskqueue_create(name, M_WAITOK,
890 		    taskqueue_thread_enqueue,
891 		    &epair_tasks.tq[cpu]);
892 		CPU_SETOF(cpu, &cpu_mask);
893 		taskqueue_start_threads_cpuset(&epair_tasks.tq[cpu], 1, PI_NET,
894 		    &cpu_mask, "%s", name);
895 
896 		epair_tasks.tasks++;
897 	}
898 	thread_lock(curthread);
899 	sched_unbind(curthread);
900 	thread_unlock(curthread);
901 #else
902 	snprintf(name, sizeof(name), "epair_task");
903 
904 	epair_tasks.tq[0] = taskqueue_create(name, M_WAITOK,
905 	    taskqueue_thread_enqueue,
906 	    &epair_tasks.tq[0]);
907 	taskqueue_start_threads(&epair_tasks.tq[0], 1, PI_NET, "%s", name);
908 
909 	epair_tasks.tasks = 1;
910 #endif
911 
912 	return (0);
913 }
914 
915 static void
916 epair_mod_cleanup(void)
917 {
918 
919 	for (int i = 0; i < epair_tasks.tasks; i++) {
920 		taskqueue_drain_all(epair_tasks.tq[i]);
921 		taskqueue_free(epair_tasks.tq[i]);
922 	}
923 }
924 
925 static int
926 epair_modevent(module_t mod, int type, void *data)
927 {
928 	int ret;
929 
930 	switch (type) {
931 	case MOD_LOAD:
932 		EPAIR_LOCK_INIT();
933 		ret = epair_mod_init();
934 		if (ret != 0)
935 			return (ret);
936 		if (bootverbose)
937 			printf("%s: %s initialized.\n", __func__, epairname);
938 		break;
939 	case MOD_UNLOAD:
940 		epair_mod_cleanup();
941 		EPAIR_LOCK_DESTROY();
942 		if (bootverbose)
943 			printf("%s: %s unloaded.\n", __func__, epairname);
944 		break;
945 	default:
946 		return (EOPNOTSUPP);
947 	}
948 	return (0);
949 }
950 
951 static moduledata_t epair_mod = {
952 	"if_epair",
953 	epair_modevent,
954 	0
955 };
956 
957 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
958 MODULE_VERSION(if_epair, 3);
959