xref: /freebsd/sys/net/if_epair.c (revision 3c4ba5f55438f7afd4f4b0b56f88f2bb505fd6a6)
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 	M_ASSERTPKTHDR(m);
137 
138 	/* Remove any CSUM_SND_TAG as ether_input will barf. */
139 	if (m->m_pkthdr.csum_flags & CSUM_SND_TAG) {
140 		m_snd_tag_rele(m->m_pkthdr.snd_tag);
141 		m->m_pkthdr.snd_tag = NULL;
142 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
143 	}
144 
145 	/* Clear vlan information. */
146 	m->m_flags &= ~M_VLANTAG;
147 	m->m_pkthdr.ether_vtag = 0;
148 
149 	m_tag_delete_nonpersistent(m);
150 }
151 
152 static void
153 epair_tx_start_deferred(void *arg, int pending)
154 {
155 	struct epair_queue *q = (struct epair_queue *)arg;
156 	if_t ifp;
157 	struct mbuf *m, *n;
158 	bool resched;
159 
160 	ifp = q->sc->ifp;
161 
162 	if_ref(ifp);
163 	CURVNET_SET(ifp->if_vnet);
164 
165 	mtx_lock(&q->mtx);
166 	m = mbufq_flush(&q->q);
167 	q->state = EPAIR_QUEUE_RUNNING;
168 	mtx_unlock(&q->mtx);
169 
170 	while (m != NULL) {
171 		n = STAILQ_NEXT(m, m_stailqpkt);
172 		m->m_nextpkt = NULL;
173 		if_input(ifp, m);
174 		m = n;
175 	}
176 
177 	/*
178 	 * Avoid flushing the queue more than once per task.  We can otherwise
179 	 * end up starving ourselves in a multi-epair routing configuration.
180 	 */
181 	mtx_lock(&q->mtx);
182 	if (mbufq_len(&q->q) > 0) {
183 		resched = true;
184 		q->state = EPAIR_QUEUE_WAKING;
185 	} else {
186 		resched = false;
187 		q->state = EPAIR_QUEUE_IDLE;
188 	}
189 	mtx_unlock(&q->mtx);
190 
191 	if (resched)
192 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
193 
194 	CURVNET_RESTORE();
195 	if_rele(ifp);
196 }
197 
198 static struct epair_queue *
199 epair_select_queue(struct epair_softc *sc, struct mbuf *m)
200 {
201 	uint32_t bucket;
202 #ifdef RSS
203 	struct ether_header *eh;
204 	int ret;
205 
206 	ret = rss_m2bucket(m, &bucket);
207 	if (ret) {
208 		/* Actually hash the packet. */
209 		eh = mtod(m, struct ether_header *);
210 
211 		switch (ntohs(eh->ether_type)) {
212 #ifdef INET
213 		case ETHERTYPE_IP:
214 			rss_soft_m2cpuid_v4(m, 0, &bucket);
215 			break;
216 #endif
217 #ifdef INET6
218 		case ETHERTYPE_IPV6:
219 			rss_soft_m2cpuid_v6(m, 0, &bucket);
220 			break;
221 #endif
222 		default:
223 			bucket = 0;
224 			break;
225 		}
226 	}
227 	bucket %= sc->num_queues;
228 #else
229 	bucket = 0;
230 #endif
231 	return (&sc->queues[bucket]);
232 }
233 
234 static void
235 epair_prepare_mbuf(struct mbuf *m, struct ifnet *src_ifp)
236 {
237 	M_ASSERTPKTHDR(m);
238 	epair_clear_mbuf(m);
239 	if_setrcvif(m, src_ifp);
240 	M_SETFIB(m, src_ifp->if_fib);
241 
242 	MPASS(m->m_nextpkt == NULL);
243 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
244 }
245 
246 static void
247 epair_menq(struct mbuf *m, struct epair_softc *osc)
248 {
249 	struct epair_queue *q;
250 	struct ifnet *ifp, *oifp;
251 	int error, len;
252 	bool mcast;
253 
254 	/*
255 	 * I know this looks weird. We pass the "other sc" as we need that one
256 	 * and can get both ifps from it as well.
257 	 */
258 	oifp = osc->ifp;
259 	ifp = osc->oifp;
260 
261 	epair_prepare_mbuf(m, oifp);
262 
263 	/* Save values as once the mbuf is queued, it's not ours anymore. */
264 	len = m->m_pkthdr.len;
265 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
266 
267 	q = epair_select_queue(osc, m);
268 
269 	mtx_lock(&q->mtx);
270 	if (q->state == EPAIR_QUEUE_IDLE) {
271 		q->state = EPAIR_QUEUE_WAKING;
272 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
273 	}
274 	error = mbufq_enqueue(&q->q, m);
275 	mtx_unlock(&q->mtx);
276 
277 	if (error != 0) {
278 		m_freem(m);
279 		if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
280 	} else {
281 		if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
282 		if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
283 		if (mcast)
284 			if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
285 		if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
286 	}
287 }
288 
289 static void
290 epair_start(struct ifnet *ifp)
291 {
292 	struct mbuf *m;
293 	struct epair_softc *sc;
294 	struct ifnet *oifp;
295 
296 	/*
297 	 * We get packets here from ether_output via if_handoff()
298 	 * and need to put them into the input queue of the oifp
299 	 * and will put the packet into the receive-queue (rxq) of the
300 	 * other interface (oifp) of our pair.
301 	 */
302 	sc = ifp->if_softc;
303 	oifp = sc->oifp;
304 	sc = oifp->if_softc;
305 	for (;;) {
306 		IFQ_DEQUEUE(&ifp->if_snd, m);
307 		if (m == NULL)
308 			break;
309 		M_ASSERTPKTHDR(m);
310 		BPF_MTAP(ifp, m);
311 
312 		/* In case either interface is not usable drop the packet. */
313 		if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
314 		    (ifp->if_flags & IFF_UP) == 0 ||
315 		    (oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
316 		    (oifp->if_flags & IFF_UP) == 0) {
317 			m_freem(m);
318 			continue;
319 		}
320 
321 		epair_menq(m, sc);
322 	}
323 }
324 
325 static int
326 epair_transmit(struct ifnet *ifp, struct mbuf *m)
327 {
328 	struct epair_softc *sc;
329 	struct ifnet *oifp;
330 #ifdef ALTQ
331 	int len;
332 	bool mcast;
333 #endif
334 
335 	if (m == NULL)
336 		return (0);
337 	M_ASSERTPKTHDR(m);
338 
339 	/*
340 	 * We are not going to use the interface en/dequeue mechanism
341 	 * on the TX side. We are called from ether_output_frame()
342 	 * and will put the packet into the receive-queue (rxq) of the
343 	 * other interface (oifp) of our pair.
344 	 */
345 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
346 		m_freem(m);
347 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
348 		return (ENXIO);
349 	}
350 	if ((ifp->if_flags & IFF_UP) == 0) {
351 		m_freem(m);
352 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
353 		return (ENETDOWN);
354 	}
355 
356 	BPF_MTAP(ifp, m);
357 
358 	/*
359 	 * In case the outgoing interface is not usable,
360 	 * drop the packet.
361 	 */
362 	sc = ifp->if_softc;
363 	oifp = sc->oifp;
364 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
365 	    (oifp->if_flags & IFF_UP) == 0) {
366 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
367 		m_freem(m);
368 		return (0);
369 	}
370 
371 #ifdef ALTQ
372 	len = m->m_pkthdr.len;
373 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
374 	int error = 0;
375 
376 	/* Support ALTQ via the classic if_start() path. */
377 	IF_LOCK(&ifp->if_snd);
378 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
379 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
380 		if (error)
381 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
382 		IF_UNLOCK(&ifp->if_snd);
383 		if (!error) {
384 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
385 			if (mcast)
386 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
387 			epair_start(ifp);
388 		}
389 		return (error);
390 	}
391 	IF_UNLOCK(&ifp->if_snd);
392 #endif
393 
394 	epair_menq(m, oifp->if_softc);
395 	return (0);
396 }
397 
398 static void
399 epair_qflush(struct ifnet *ifp __unused)
400 {
401 }
402 
403 static int
404 epair_media_change(struct ifnet *ifp __unused)
405 {
406 
407 	/* Do nothing. */
408 	return (0);
409 }
410 
411 static void
412 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
413 {
414 
415 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
416 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
417 }
418 
419 static int
420 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
421 {
422 	struct epair_softc *sc;
423 	struct ifreq *ifr;
424 	int error;
425 
426 	ifr = (struct ifreq *)data;
427 	switch (cmd) {
428 	case SIOCSIFFLAGS:
429 	case SIOCADDMULTI:
430 	case SIOCDELMULTI:
431 		error = 0;
432 		break;
433 
434 	case SIOCSIFMEDIA:
435 	case SIOCGIFMEDIA:
436 		sc = ifp->if_softc;
437 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
438 		break;
439 
440 	case SIOCSIFMTU:
441 		/* We basically allow all kinds of MTUs. */
442 		ifp->if_mtu = ifr->ifr_mtu;
443 		error = 0;
444 		break;
445 
446 	default:
447 		/* Let the common ethernet handler process this. */
448 		error = ether_ioctl(ifp, cmd, data);
449 		break;
450 	}
451 
452 	return (error);
453 }
454 
455 static void
456 epair_init(void *dummy __unused)
457 {
458 }
459 
460 /*
461  * Interface cloning functions.
462  * We use our private ones so that we can create/destroy our secondary
463  * device along with the primary one.
464  */
465 static int
466 epair_clone_match(struct if_clone *ifc, const char *name)
467 {
468 	const char *cp;
469 
470 	/*
471 	 * Our base name is epair.
472 	 * Our interfaces will be named epair<n>[ab].
473 	 * So accept anything of the following list:
474 	 * - epair
475 	 * - epair<n>
476 	 * but not the epair<n>[ab] versions.
477 	 */
478 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
479 		return (0);
480 
481 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
482 		if (*cp < '0' || *cp > '9')
483 			return (0);
484 	}
485 
486 	return (1);
487 }
488 
489 static void
490 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
491 {
492 	struct ifnet *ifp;
493 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
494 
495 	ifp = scb->ifp;
496 	/* Copy epairNa etheraddr and change the last byte. */
497 	memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
498 	eaddr[5] = 0x0b;
499 	ether_ifattach(ifp, eaddr);
500 
501 	if_clone_addif(ifc, ifp);
502 }
503 
504 static struct epair_softc *
505 epair_alloc_sc(struct if_clone *ifc)
506 {
507 	struct epair_softc *sc;
508 
509 	struct ifnet *ifp = if_alloc(IFT_ETHER);
510 	if (ifp == NULL)
511 		return (NULL);
512 
513 	sc = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
514 	sc->ifp = ifp;
515 	sc->num_queues = epair_tasks.tasks;
516 	sc->queues = mallocarray(sc->num_queues, sizeof(struct epair_queue),
517 	    M_EPAIR, M_WAITOK);
518 	for (int i = 0; i < sc->num_queues; i++) {
519 		struct epair_queue *q = &sc->queues[i];
520 		q->id = i;
521 		q->state = EPAIR_QUEUE_IDLE;
522 		mtx_init(&q->mtx, "epairq", NULL, MTX_DEF | MTX_NEW);
523 		mbufq_init(&q->q, RXRSIZE);
524 		q->sc = sc;
525 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
526 	}
527 
528 	/* Initialise pseudo media types. */
529 	ifmedia_init(&sc->media, 0, epair_media_change, epair_media_status);
530 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T, 0, NULL);
531 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T);
532 
533 	return (sc);
534 }
535 
536 static void
537 epair_setup_ifp(struct epair_softc *sc, char *name, int unit)
538 {
539 	struct ifnet *ifp = sc->ifp;
540 
541 	ifp->if_softc = sc;
542 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
543 	ifp->if_dname = epairname;
544 	ifp->if_dunit = unit;
545 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
546 	ifp->if_capabilities = IFCAP_VLAN_MTU;
547 	ifp->if_capenable = IFCAP_VLAN_MTU;
548 	ifp->if_transmit = epair_transmit;
549 	ifp->if_qflush = epair_qflush;
550 	ifp->if_start = epair_start;
551 	ifp->if_ioctl = epair_ioctl;
552 	ifp->if_init  = epair_init;
553 	if_setsendqlen(ifp, ifqmaxlen);
554 	if_setsendqready(ifp);
555 
556 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
557 }
558 
559 static void
560 epair_generate_mac(struct epair_softc *sc, uint8_t *eaddr)
561 {
562 	uint32_t key[3];
563 	uint32_t hash;
564 	uint64_t hostid;
565 
566 	EPAIR_LOCK();
567 #ifdef SMP
568 	/* Get an approximate distribution. */
569 	hash = next_index % mp_ncpus;
570 #else
571 	hash = 0;
572 #endif
573 	EPAIR_UNLOCK();
574 
575 	/*
576 	 * Calculate the etheraddr hashing the hostid and the
577 	 * interface index. The result would be hopefully unique.
578 	 * Note that the "a" component of an epair instance may get moved
579 	 * to a different VNET after creation. In that case its index
580 	 * will be freed and the index can get reused by new epair instance.
581 	 * Make sure we do not create same etheraddr again.
582 	 */
583 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
584 	if (hostid == 0)
585 		arc4rand(&hostid, sizeof(hostid), 0);
586 
587 	struct ifnet *ifp = sc->ifp;
588 	EPAIR_LOCK();
589 	if (ifp->if_index > next_index)
590 		next_index = ifp->if_index;
591 	else
592 		next_index++;
593 
594 	key[0] = (uint32_t)next_index;
595 	EPAIR_UNLOCK();
596 	key[1] = (uint32_t)(hostid & 0xffffffff);
597 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
598 	hash = jenkins_hash32(key, 3, 0);
599 
600 	eaddr[0] = 0x02;
601 	memcpy(&eaddr[1], &hash, 4);
602 	eaddr[5] = 0x0a;
603 }
604 
605 static void
606 epair_free_sc(struct epair_softc *sc)
607 {
608 	if (sc == NULL)
609 		return;
610 
611 	if_free(sc->ifp);
612 	ifmedia_removeall(&sc->media);
613 	for (int i = 0; i < sc->num_queues; i++) {
614 		struct epair_queue *q = &sc->queues[i];
615 		mtx_destroy(&q->mtx);
616 	}
617 	free(sc->queues, M_EPAIR);
618 	free(sc, M_EPAIR);
619 }
620 
621 static void
622 epair_set_state(struct ifnet *ifp, bool running)
623 {
624 	if (running) {
625 		ifp->if_drv_flags |= IFF_DRV_RUNNING;
626 		if_link_state_change(ifp, LINK_STATE_UP);
627 	} else {
628 		if_link_state_change(ifp, LINK_STATE_DOWN);
629 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
630 	}
631 }
632 
633 static int
634 epair_handle_unit(struct if_clone *ifc, char *name, size_t len, int *punit)
635 {
636 	int error = 0, unit, wildcard;
637 	char *dp;
638 
639 	/* Try to see if a special unit was requested. */
640 	error = ifc_name2unit(name, &unit);
641 	if (error != 0)
642 		return (error);
643 	wildcard = (unit < 0);
644 
645 	error = ifc_alloc_unit(ifc, &unit);
646 	if (error != 0)
647 		return (error);
648 
649 	/*
650 	 * If no unit had been given, we need to adjust the ifName.
651 	 * Also make sure there is space for our extra [ab] suffix.
652 	 */
653 	for (dp = name; *dp != '\0'; dp++);
654 	if (wildcard) {
655 		int slen = snprintf(dp, len - (dp - name), "%d", unit);
656 		if (slen > len - (dp - name) - 1) {
657 			/* ifName too long. */
658 			error = ENOSPC;
659 			goto done;
660 		}
661 		dp += slen;
662 	}
663 	if (len - (dp - name) - 1 < 1) {
664 		/* No space left for our [ab] suffix. */
665 		error = ENOSPC;
666 		goto done;
667 	}
668 	*dp = 'b';
669 	/* Must not change dp so we can replace 'a' by 'b' later. */
670 	*(dp+1) = '\0';
671 
672 	/* Check if 'a' and 'b' interfaces already exist. */
673 	if (ifunit(name) != NULL) {
674 		error = EEXIST;
675 		goto done;
676 	}
677 
678 	*dp = 'a';
679 	if (ifunit(name) != NULL) {
680 		error = EEXIST;
681 		goto done;
682 	}
683 	*punit = unit;
684 done:
685 	if (error != 0)
686 		ifc_free_unit(ifc, unit);
687 
688 	return (error);
689 }
690 
691 static int
692 epair_clone_create(struct if_clone *ifc, char *name, size_t len,
693     struct ifc_data *ifd, struct ifnet **ifpp)
694 {
695 	struct epair_softc *sca, *scb;
696 	struct ifnet *ifp;
697 	char *dp;
698 	int error, unit;
699 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
700 
701 	error = epair_handle_unit(ifc, name, len, &unit);
702 	if (error != 0)
703 		return (error);
704 
705 	/* Allocate memory for both [ab] interfaces */
706 	sca = epair_alloc_sc(ifc);
707 	scb = epair_alloc_sc(ifc);
708 	if (sca == NULL || scb == NULL) {
709 		epair_free_sc(sca);
710 		epair_free_sc(scb);
711 		ifc_free_unit(ifc, unit);
712 		return (ENOSPC);
713 	}
714 
715 	/*
716 	 * Cross-reference the interfaces so we will be able to free both.
717 	 */
718 	sca->oifp = scb->ifp;
719 	scb->oifp = sca->ifp;
720 
721 	/* Finish initialization of interface <n>a. */
722 	ifp = sca->ifp;
723 	epair_setup_ifp(sca, name, unit);
724 	epair_generate_mac(sca, eaddr);
725 
726 	ether_ifattach(ifp, eaddr);
727 
728 	/* Swap the name and finish initialization of interface <n>b. */
729 	dp = name + strlen(name) - 1;
730 	*dp = 'b';
731 
732 	epair_setup_ifp(scb, name, unit);
733 
734 	ifp = scb->ifp;
735 	/* We need to play some tricks here for the second interface. */
736 	strlcpy(name, epairname, len);
737 	/* Correctly set the name for the cloner list. */
738 	strlcpy(name, scb->ifp->if_xname, len);
739 
740 	epair_clone_add(ifc, scb);
741 
742 	/*
743 	 * Restore name to <n>a as the ifp for this will go into the
744 	 * cloner list for the initial call.
745 	 */
746 	strlcpy(name, sca->ifp->if_xname, len);
747 
748 	/* Tell the world, that we are ready to rock. */
749 	epair_set_state(sca->ifp, true);
750 	epair_set_state(scb->ifp, true);
751 
752 	*ifpp = sca->ifp;
753 
754 	return (0);
755 }
756 
757 static void
758 epair_drain_rings(struct epair_softc *sc)
759 {
760 	for (int i = 0; i < sc->num_queues; i++) {
761 		struct epair_queue *q;
762 		struct mbuf *m, *n;
763 
764 		q = &sc->queues[i];
765 		mtx_lock(&q->mtx);
766 		m = mbufq_flush(&q->q);
767 		mtx_unlock(&q->mtx);
768 
769 		for (; m != NULL; m = n) {
770 			n = m->m_nextpkt;
771 			m_freem(m);
772 		}
773 	}
774 }
775 
776 static int
777 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp, uint32_t flags)
778 {
779 	struct ifnet *oifp;
780 	struct epair_softc *sca, *scb;
781 	int unit, error;
782 
783 	/*
784 	 * In case we called into if_clone_destroyif() ourselves
785 	 * again to remove the second interface, the softc will be
786 	 * NULL. In that case so not do anything but return success.
787 	 */
788 	if (ifp->if_softc == NULL)
789 		return (0);
790 
791 	unit = ifp->if_dunit;
792 	sca = ifp->if_softc;
793 	oifp = sca->oifp;
794 	scb = oifp->if_softc;
795 
796 	/* Frist get the interfaces down and detached. */
797 	epair_set_state(ifp, false);
798 	epair_set_state(oifp, false);
799 
800 	ether_ifdetach(ifp);
801 	ether_ifdetach(oifp);
802 
803 	/* Third free any queued packets and all the resources. */
804 	CURVNET_SET_QUIET(oifp->if_vnet);
805 	epair_drain_rings(scb);
806 	oifp->if_softc = NULL;
807 	error = if_clone_destroyif(ifc, oifp);
808 	if (error)
809 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
810 		    __func__, error);
811 	epair_free_sc(scb);
812 	CURVNET_RESTORE();
813 
814 	epair_drain_rings(sca);
815 	epair_free_sc(sca);
816 
817 	/* Last free the cloner unit. */
818 	ifc_free_unit(ifc, unit);
819 
820 	return (0);
821 }
822 
823 static void
824 vnet_epair_init(const void *unused __unused)
825 {
826 	struct if_clone_addreq req = {
827 		.match_f = epair_clone_match,
828 		.create_f = epair_clone_create,
829 		.destroy_f = epair_clone_destroy,
830 	};
831 	V_epair_cloner = ifc_attach_cloner(epairname, &req);
832 }
833 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
834     vnet_epair_init, NULL);
835 
836 static void
837 vnet_epair_uninit(const void *unused __unused)
838 {
839 
840 	ifc_detach_cloner(V_epair_cloner);
841 }
842 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
843     vnet_epair_uninit, NULL);
844 
845 static int
846 epair_mod_init(void)
847 {
848 	char name[32];
849 	epair_tasks.tasks = 0;
850 
851 #ifdef RSS
852 	int cpu;
853 
854 	CPU_FOREACH(cpu) {
855 		cpuset_t cpu_mask;
856 
857 		/* Pin to this CPU so we get appropriate NUMA allocations. */
858 		thread_lock(curthread);
859 		sched_bind(curthread, cpu);
860 		thread_unlock(curthread);
861 
862 		snprintf(name, sizeof(name), "epair_task_%d", cpu);
863 
864 		epair_tasks.tq[cpu] = taskqueue_create(name, M_WAITOK,
865 		    taskqueue_thread_enqueue,
866 		    &epair_tasks.tq[cpu]);
867 		CPU_SETOF(cpu, &cpu_mask);
868 		taskqueue_start_threads_cpuset(&epair_tasks.tq[cpu], 1, PI_NET,
869 		    &cpu_mask, "%s", name);
870 
871 		epair_tasks.tasks++;
872 	}
873 	thread_lock(curthread);
874 	sched_unbind(curthread);
875 	thread_unlock(curthread);
876 #else
877 	snprintf(name, sizeof(name), "epair_task");
878 
879 	epair_tasks.tq[0] = taskqueue_create(name, M_WAITOK,
880 	    taskqueue_thread_enqueue,
881 	    &epair_tasks.tq[0]);
882 	taskqueue_start_threads(&epair_tasks.tq[0], 1, PI_NET, "%s", name);
883 
884 	epair_tasks.tasks = 1;
885 #endif
886 
887 	return (0);
888 }
889 
890 static void
891 epair_mod_cleanup(void)
892 {
893 
894 	for (int i = 0; i < epair_tasks.tasks; i++) {
895 		taskqueue_drain_all(epair_tasks.tq[i]);
896 		taskqueue_free(epair_tasks.tq[i]);
897 	}
898 }
899 
900 static int
901 epair_modevent(module_t mod, int type, void *data)
902 {
903 	int ret;
904 
905 	switch (type) {
906 	case MOD_LOAD:
907 		EPAIR_LOCK_INIT();
908 		ret = epair_mod_init();
909 		if (ret != 0)
910 			return (ret);
911 		if (bootverbose)
912 			printf("%s: %s initialized.\n", __func__, epairname);
913 		break;
914 	case MOD_UNLOAD:
915 		epair_mod_cleanup();
916 		EPAIR_LOCK_DESTROY();
917 		if (bootverbose)
918 			printf("%s: %s unloaded.\n", __func__, epairname);
919 		break;
920 	default:
921 		return (EOPNOTSUPP);
922 	}
923 	return (0);
924 }
925 
926 static moduledata_t epair_mod = {
927 	"if_epair",
928 	epair_modevent,
929 	0
930 };
931 
932 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
933 MODULE_VERSION(if_epair, 3);
934