xref: /freebsd/sys/net/if_epair.c (revision 84823cc70824c8d842f503d8c2e6d7b0c2d95b61)
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/hash.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/taskqueue.h>
62 #include <sys/types.h>
63 #include <sys/buf_ring.h>
64 #include <sys/bus.h>
65 #include <sys/interrupt.h>
66 
67 #include <net/bpf.h>
68 #include <net/ethernet.h>
69 #include <net/if.h>
70 #include <net/if_var.h>
71 #include <net/if_clone.h>
72 #include <net/if_media.h>
73 #include <net/if_var.h>
74 #include <net/if_types.h>
75 #include <net/netisr.h>
76 #ifdef RSS
77 #include <net/rss_config.h>
78 #ifdef INET
79 #include <netinet/in_rss.h>
80 #endif
81 #ifdef INET6
82 #include <netinet6/in6_rss.h>
83 #endif
84 #endif
85 #include <net/vnet.h>
86 
87 static const char epairname[] = "epair";
88 #define	RXRSIZE	4096	/* Probably overkill by 4-8x. */
89 
90 static MALLOC_DEFINE(M_EPAIR, epairname,
91     "Pair of virtual cross-over connected Ethernet-like interfaces");
92 
93 VNET_DEFINE_STATIC(struct if_clone *, epair_cloner);
94 #define	V_epair_cloner	VNET(epair_cloner)
95 
96 static unsigned int next_index = 0;
97 #define	EPAIR_LOCK_INIT()		mtx_init(&epair_n_index_mtx, "epairidx", \
98 					    NULL, MTX_DEF)
99 #define	EPAIR_LOCK_DESTROY()		mtx_destroy(&epair_n_index_mtx)
100 #define	EPAIR_LOCK()			mtx_lock(&epair_n_index_mtx)
101 #define	EPAIR_UNLOCK()			mtx_unlock(&epair_n_index_mtx)
102 
103 #define BIT_QUEUE_TASK		0
104 #define BIT_MBUF_QUEUED		1
105 
106 struct epair_softc;
107 struct epair_queue {
108 	int			 id;
109 	struct buf_ring		*rxring[2];
110 	volatile int		 ridx;		/* 0 || 1 */
111 	volatile long		 state;		/* taskqueue coordination */
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_if_input(struct epair_softc *sc, struct epair_queue *q, int ridx)
148 {
149 	struct ifnet *ifp;
150 	struct mbuf *m;
151 
152 	ifp = sc->ifp;
153 	CURVNET_SET(ifp->if_vnet);
154 	while (! buf_ring_empty(q->rxring[ridx])) {
155 		m = buf_ring_dequeue_mc(q->rxring[ridx]);
156 		if (m == NULL)
157 			continue;
158 
159 		MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
160 		(*ifp->if_input)(ifp, m);
161 
162 	}
163 	CURVNET_RESTORE();
164 }
165 
166 static void
167 epair_tx_start_deferred(void *arg, int pending)
168 {
169 	struct epair_queue *q = (struct epair_queue *)arg;
170 	struct epair_softc *sc = q->sc;
171 	int ridx, nidx;
172 
173 	if_ref(sc->ifp);
174 	ridx = atomic_load_int(&q->ridx);
175 	do {
176 		nidx = (ridx == 0) ? 1 : 0;
177 	} while (!atomic_fcmpset_int(&q->ridx, &ridx, nidx));
178 	epair_if_input(sc, q, ridx);
179 
180 	atomic_clear_long(&q->state, (1 << BIT_QUEUE_TASK));
181 	if (atomic_testandclear_long(&q->state, BIT_MBUF_QUEUED))
182 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
183 
184 	if_rele(sc->ifp);
185 }
186 
187 static int
188 epair_menq(struct mbuf *m, struct epair_softc *osc)
189 {
190 	struct ifnet *ifp, *oifp;
191 	int len, ret;
192 	int ridx;
193 	short mflags;
194 	struct epair_queue *q = NULL;
195 	uint32_t bucket;
196 #ifdef RSS
197 	struct ether_header *eh;
198 #endif
199 
200 	/*
201 	 * I know this looks weird. We pass the "other sc" as we need that one
202 	 * and can get both ifps from it as well.
203 	 */
204 	oifp = osc->ifp;
205 	ifp = osc->oifp;
206 
207 	M_ASSERTPKTHDR(m);
208 	epair_clear_mbuf(m);
209 	if_setrcvif(m, oifp);
210 	M_SETFIB(m, oifp->if_fib);
211 
212 	/* Save values as once the mbuf is queued, it's not ours anymore. */
213 	len = m->m_pkthdr.len;
214 	mflags = m->m_flags;
215 
216 	MPASS(m->m_nextpkt == NULL);
217 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
218 
219 #ifdef RSS
220 	ret = rss_m2bucket(m, &bucket);
221 	if (ret) {
222 		/* Actually hash the packet. */
223 		eh = mtod(m, struct ether_header *);
224 
225 		switch (ntohs(eh->ether_type)) {
226 #ifdef INET
227 		case ETHERTYPE_IP:
228 			rss_soft_m2cpuid_v4(m, 0, &bucket);
229 			break;
230 #endif
231 #ifdef INET6
232 		case ETHERTYPE_IPV6:
233 			rss_soft_m2cpuid_v6(m, 0, &bucket);
234 			break;
235 #endif
236 		default:
237 			bucket = 0;
238 			break;
239 		}
240 	}
241 	bucket %= osc->num_queues;
242 #else
243 	bucket = 0;
244 #endif
245 	q = &osc->queues[bucket];
246 
247 	atomic_set_long(&q->state, (1 << BIT_MBUF_QUEUED));
248 	ridx = atomic_load_int(&q->ridx);
249 	ret = buf_ring_enqueue(q->rxring[ridx], m);
250 	if (ret != 0) {
251 		/* Ring is full. */
252 		if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
253 		m_freem(m);
254 		return (0);
255 	}
256 
257 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
258 	/*
259 	 * IFQ_HANDOFF_ADJ/ip_handoff() update statistics,
260 	 * but as we bypass all this we have to duplicate
261 	 * the logic another time.
262 	 */
263 	if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
264 	if (mflags & (M_BCAST|M_MCAST))
265 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
266 	/* Someone else received the packet. */
267 	if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
268 
269 	if (!atomic_testandset_long(&q->state, BIT_QUEUE_TASK))
270 		taskqueue_enqueue(epair_tasks.tq[bucket], &q->tx_task);
271 
272 	return (0);
273 }
274 
275 static void
276 epair_start(struct ifnet *ifp)
277 {
278 	struct mbuf *m;
279 	struct epair_softc *sc;
280 	struct ifnet *oifp;
281 
282 	/*
283 	 * We get packets here from ether_output via if_handoff()
284 	 * and need to put them into the input queue of the oifp
285 	 * and will put the packet into the receive-queue (rxq) of the
286 	 * other interface (oifp) of our pair.
287 	 */
288 	sc = ifp->if_softc;
289 	oifp = sc->oifp;
290 	sc = oifp->if_softc;
291 	for (;;) {
292 		IFQ_DEQUEUE(&ifp->if_snd, m);
293 		if (m == NULL)
294 			break;
295 		M_ASSERTPKTHDR(m);
296 		BPF_MTAP(ifp, m);
297 
298 		/* In case either interface is not usable drop the packet. */
299 		if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
300 		    (ifp->if_flags & IFF_UP) == 0 ||
301 		    (oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
302 		    (oifp->if_flags & IFF_UP) == 0) {
303 			m_freem(m);
304 			continue;
305 		}
306 
307 		(void) epair_menq(m, sc);
308 	}
309 }
310 
311 static int
312 epair_transmit(struct ifnet *ifp, struct mbuf *m)
313 {
314 	struct epair_softc *sc;
315 	struct ifnet *oifp;
316 	int error;
317 #ifdef ALTQ
318 	int len;
319 	short mflags;
320 #endif
321 
322 	if (m == NULL)
323 		return (0);
324 	M_ASSERTPKTHDR(m);
325 
326 	/*
327 	 * We are not going to use the interface en/dequeue mechanism
328 	 * on the TX side. We are called from ether_output_frame()
329 	 * and will put the packet into the receive-queue (rxq) of the
330 	 * other interface (oifp) of our pair.
331 	 */
332 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
333 		m_freem(m);
334 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
335 		return (ENXIO);
336 	}
337 	if ((ifp->if_flags & IFF_UP) == 0) {
338 		m_freem(m);
339 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
340 		return (ENETDOWN);
341 	}
342 
343 	BPF_MTAP(ifp, m);
344 
345 	/*
346 	 * In case the outgoing interface is not usable,
347 	 * drop the packet.
348 	 */
349 	sc = ifp->if_softc;
350 	oifp = sc->oifp;
351 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
352 	    (oifp->if_flags & IFF_UP) == 0) {
353 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
354 		m_freem(m);
355 		return (0);
356 	}
357 
358 #ifdef ALTQ
359 	len = m->m_pkthdr.len;
360 	mflags = m->m_flags;
361 
362 	/* Support ALTQ via the classic if_start() path. */
363 	IF_LOCK(&ifp->if_snd);
364 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
365 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
366 		if (error)
367 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
368 		IF_UNLOCK(&ifp->if_snd);
369 		if (!error) {
370 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
371 			if (mflags & (M_BCAST|M_MCAST))
372 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
373 			epair_start(ifp);
374 		}
375 		return (error);
376 	}
377 	IF_UNLOCK(&ifp->if_snd);
378 #endif
379 
380 	error = epair_menq(m, oifp->if_softc);
381 	return (error);
382 }
383 
384 static void
385 epair_qflush(struct ifnet *ifp __unused)
386 {
387 }
388 
389 static int
390 epair_media_change(struct ifnet *ifp __unused)
391 {
392 
393 	/* Do nothing. */
394 	return (0);
395 }
396 
397 static void
398 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
399 {
400 
401 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
402 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
403 }
404 
405 static int
406 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
407 {
408 	struct epair_softc *sc;
409 	struct ifreq *ifr;
410 	int error;
411 
412 	ifr = (struct ifreq *)data;
413 	switch (cmd) {
414 	case SIOCSIFFLAGS:
415 	case SIOCADDMULTI:
416 	case SIOCDELMULTI:
417 		error = 0;
418 		break;
419 
420 	case SIOCSIFMEDIA:
421 	case SIOCGIFMEDIA:
422 		sc = ifp->if_softc;
423 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
424 		break;
425 
426 	case SIOCSIFMTU:
427 		/* We basically allow all kinds of MTUs. */
428 		ifp->if_mtu = ifr->ifr_mtu;
429 		error = 0;
430 		break;
431 
432 	default:
433 		/* Let the common ethernet handler process this. */
434 		error = ether_ioctl(ifp, cmd, data);
435 		break;
436 	}
437 
438 	return (error);
439 }
440 
441 static void
442 epair_init(void *dummy __unused)
443 {
444 }
445 
446 /*
447  * Interface cloning functions.
448  * We use our private ones so that we can create/destroy our secondary
449  * device along with the primary one.
450  */
451 static int
452 epair_clone_match(struct if_clone *ifc, const char *name)
453 {
454 	const char *cp;
455 
456 	/*
457 	 * Our base name is epair.
458 	 * Our interfaces will be named epair<n>[ab].
459 	 * So accept anything of the following list:
460 	 * - epair
461 	 * - epair<n>
462 	 * but not the epair<n>[ab] versions.
463 	 */
464 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
465 		return (0);
466 
467 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
468 		if (*cp < '0' || *cp > '9')
469 			return (0);
470 	}
471 
472 	return (1);
473 }
474 
475 static void
476 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
477 {
478 	struct ifnet *ifp;
479 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
480 
481 	ifp = scb->ifp;
482 	/* Copy epairNa etheraddr and change the last byte. */
483 	memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
484 	eaddr[5] = 0x0b;
485 	ether_ifattach(ifp, eaddr);
486 
487 	if_clone_addif(ifc, ifp);
488 }
489 
490 static struct epair_softc *
491 epair_alloc_sc(struct if_clone *ifc)
492 {
493 	struct epair_softc *sc;
494 
495 	struct ifnet *ifp = if_alloc(IFT_ETHER);
496 	if (ifp == NULL)
497 		return (NULL);
498 
499 	sc = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
500 	sc->ifp = ifp;
501 	sc->num_queues = epair_tasks.tasks;
502 	sc->queues = mallocarray(sc->num_queues, sizeof(struct epair_queue),
503 	    M_EPAIR, M_WAITOK);
504 	for (int i = 0; i < sc->num_queues; i++) {
505 		struct epair_queue *q = &sc->queues[i];
506 		q->id = i;
507 		q->rxring[0] = buf_ring_alloc(RXRSIZE, M_EPAIR, M_WAITOK, NULL);
508 		q->rxring[1] = buf_ring_alloc(RXRSIZE, M_EPAIR, M_WAITOK, NULL);
509 		q->ridx = 0;
510 		q->state = 0;
511 		q->sc = sc;
512 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
513 	}
514 
515 	/* Initialise pseudo media types. */
516 	ifmedia_init(&sc->media, 0, epair_media_change, epair_media_status);
517 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T, 0, NULL);
518 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T);
519 
520 	return (sc);
521 }
522 
523 static void
524 epair_setup_ifp(struct epair_softc *sc, char *name, int unit)
525 {
526 	struct ifnet *ifp = sc->ifp;
527 
528 	ifp->if_softc = sc;
529 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
530 	ifp->if_dname = epairname;
531 	ifp->if_dunit = unit;
532 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
533 	ifp->if_flags |= IFF_KNOWSEPOCH;
534 	ifp->if_capabilities = IFCAP_VLAN_MTU;
535 	ifp->if_capenable = IFCAP_VLAN_MTU;
536 	ifp->if_transmit = epair_transmit;
537 	ifp->if_qflush = epair_qflush;
538 	ifp->if_start = epair_start;
539 	ifp->if_ioctl = epair_ioctl;
540 	ifp->if_init  = epair_init;
541 	if_setsendqlen(ifp, ifqmaxlen);
542 	if_setsendqready(ifp);
543 
544 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
545 }
546 
547 static void
548 epair_generate_mac(struct epair_softc *sc, uint8_t *eaddr)
549 {
550 	uint32_t key[3];
551 	uint32_t hash;
552 	uint64_t hostid;
553 
554 	EPAIR_LOCK();
555 #ifdef SMP
556 	/* Get an approximate distribution. */
557 	hash = next_index % mp_ncpus;
558 #else
559 	hash = 0;
560 #endif
561 	EPAIR_UNLOCK();
562 
563 	/*
564 	 * Calculate the etheraddr hashing the hostid and the
565 	 * interface index. The result would be hopefully unique.
566 	 * Note that the "a" component of an epair instance may get moved
567 	 * to a different VNET after creation. In that case its index
568 	 * will be freed and the index can get reused by new epair instance.
569 	 * Make sure we do not create same etheraddr again.
570 	 */
571 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
572 	if (hostid == 0)
573 		arc4rand(&hostid, sizeof(hostid), 0);
574 
575 	struct ifnet *ifp = sc->ifp;
576 	EPAIR_LOCK();
577 	if (ifp->if_index > next_index)
578 		next_index = ifp->if_index;
579 	else
580 		next_index++;
581 
582 	key[0] = (uint32_t)next_index;
583 	EPAIR_UNLOCK();
584 	key[1] = (uint32_t)(hostid & 0xffffffff);
585 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
586 	hash = jenkins_hash32(key, 3, 0);
587 
588 	eaddr[0] = 0x02;
589 	memcpy(&eaddr[1], &hash, 4);
590 	eaddr[5] = 0x0a;
591 }
592 
593 static void
594 epair_free_sc(struct epair_softc *sc)
595 {
596 	if (sc == NULL)
597 		return;
598 
599 	if_free(sc->ifp);
600 	ifmedia_removeall(&sc->media);
601 	for (int i = 0; i < sc->num_queues; i++) {
602 		struct epair_queue *q = &sc->queues[i];
603 		buf_ring_free(q->rxring[0], M_EPAIR);
604 		buf_ring_free(q->rxring[1], M_EPAIR);
605 	}
606 	free(sc->queues, M_EPAIR);
607 	free(sc, M_EPAIR);
608 }
609 
610 static int
611 epair_clone_create(struct if_clone *ifc, char *name, size_t len,
612     struct ifc_data *ifd, struct ifnet **ifpp)
613 {
614 	struct epair_softc *sca, *scb;
615 	struct ifnet *ifp;
616 	char *dp;
617 	int error, unit, wildcard;
618 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
619 
620 	/* Try to see if a special unit was requested. */
621 	error = ifc_name2unit(name, &unit);
622 	if (error != 0)
623 		return (error);
624 	wildcard = (unit < 0);
625 
626 	error = ifc_alloc_unit(ifc, &unit);
627 	if (error != 0)
628 		return (error);
629 
630 	/*
631 	 * If no unit had been given, we need to adjust the ifName.
632 	 * Also make sure there is space for our extra [ab] suffix.
633 	 */
634 	for (dp = name; *dp != '\0'; dp++);
635 	if (wildcard) {
636 		error = snprintf(dp, len - (dp - name), "%d", unit);
637 		if (error > len - (dp - name) - 1) {
638 			/* ifName too long. */
639 			ifc_free_unit(ifc, unit);
640 			return (ENOSPC);
641 		}
642 		dp += error;
643 	}
644 	if (len - (dp - name) - 1 < 1) {
645 		/* No space left for our [ab] suffix. */
646 		ifc_free_unit(ifc, unit);
647 		return (ENOSPC);
648 	}
649 	*dp = 'b';
650 	/* Must not change dp so we can replace 'a' by 'b' later. */
651 	*(dp+1) = '\0';
652 
653 	/* Check if 'a' and 'b' interfaces already exist. */
654 	if (ifunit(name) != NULL)
655 		return (EEXIST);
656 	*dp = 'a';
657 	if (ifunit(name) != NULL)
658 		return (EEXIST);
659 
660 	/* Allocate memory for both [ab] interfaces */
661 	sca = epair_alloc_sc(ifc);
662 	scb = epair_alloc_sc(ifc);
663 	if (sca == NULL || scb == NULL) {
664 		epair_free_sc(sca);
665 		epair_free_sc(scb);
666 		ifc_free_unit(ifc, unit);
667 		return (ENOSPC);
668 	}
669 
670 	/*
671 	 * Cross-reference the interfaces so we will be able to free both.
672 	 */
673 	sca->oifp = scb->ifp;
674 	scb->oifp = sca->ifp;
675 
676 	/* Finish initialization of interface <n>a. */
677 	ifp = sca->ifp;
678 	epair_setup_ifp(sca, name, unit);
679 	epair_generate_mac(sca, eaddr);
680 
681 	ether_ifattach(ifp, eaddr);
682 
683 	/* Swap the name and finish initialization of interface <n>b. */
684 	*dp = 'b';
685 
686 	epair_setup_ifp(scb, name, unit);
687 
688 	ifp = scb->ifp;
689 	/* We need to play some tricks here for the second interface. */
690 	strlcpy(name, epairname, len);
691 	/* Correctly set the name for the cloner list. */
692 	strlcpy(name, scb->ifp->if_xname, len);
693 
694 	epair_clone_add(ifc, scb);
695 
696 	/*
697 	 * Restore name to <n>a as the ifp for this will go into the
698 	 * cloner list for the initial call.
699 	 */
700 	strlcpy(name, sca->ifp->if_xname, len);
701 
702 	/* Tell the world, that we are ready to rock. */
703 	sca->ifp->if_drv_flags |= IFF_DRV_RUNNING;
704 	if_link_state_change(sca->ifp, LINK_STATE_UP);
705 	scb->ifp->if_drv_flags |= IFF_DRV_RUNNING;
706 	if_link_state_change(scb->ifp, LINK_STATE_UP);
707 
708 	*ifpp = sca->ifp;
709 
710 	return (0);
711 }
712 
713 static void
714 epair_drain_rings(struct epair_softc *sc)
715 {
716 	int ridx;
717 	struct mbuf *m;
718 
719 	for (ridx = 0; ridx < 2; ridx++) {
720 		for (int i = 0; i < sc->num_queues; i++) {
721 			struct epair_queue *q = &sc->queues[i];
722 			do {
723 				m = buf_ring_dequeue_sc(q->rxring[ridx]);
724 				if (m == NULL)
725 					break;
726 				m_freem(m);
727 			} while (1);
728 		}
729 	}
730 }
731 
732 static int
733 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp, uint32_t flags)
734 {
735 	struct ifnet *oifp;
736 	struct epair_softc *sca, *scb;
737 	int unit, error;
738 
739 	/*
740 	 * In case we called into if_clone_destroyif() ourselves
741 	 * again to remove the second interface, the softc will be
742 	 * NULL. In that case so not do anything but return success.
743 	 */
744 	if (ifp->if_softc == NULL)
745 		return (0);
746 
747 	unit = ifp->if_dunit;
748 	sca = ifp->if_softc;
749 	oifp = sca->oifp;
750 	scb = oifp->if_softc;
751 
752 	/* Frist get the interfaces down and detached. */
753 	if_link_state_change(ifp, LINK_STATE_DOWN);
754 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
755 	if_link_state_change(oifp, LINK_STATE_DOWN);
756 	oifp->if_drv_flags &= ~IFF_DRV_RUNNING;
757 
758 	ether_ifdetach(ifp);
759 	ether_ifdetach(oifp);
760 
761 	/* Third free any queued packets and all the resources. */
762 	CURVNET_SET_QUIET(oifp->if_vnet);
763 	epair_drain_rings(scb);
764 	oifp->if_softc = NULL;
765 	error = if_clone_destroyif(ifc, oifp);
766 	if (error)
767 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
768 		    __func__, error);
769 	epair_free_sc(scb);
770 	CURVNET_RESTORE();
771 
772 	epair_drain_rings(sca);
773 	epair_free_sc(sca);
774 
775 	/* Last free the cloner unit. */
776 	ifc_free_unit(ifc, unit);
777 
778 	return (0);
779 }
780 
781 static void
782 vnet_epair_init(const void *unused __unused)
783 {
784 	struct if_clone_addreq req = {
785 		.match_f = epair_clone_match,
786 		.create_f = epair_clone_create,
787 		.destroy_f = epair_clone_destroy,
788 	};
789 	V_epair_cloner = ifc_attach_cloner(epairname, &req);
790 }
791 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
792     vnet_epair_init, NULL);
793 
794 static void
795 vnet_epair_uninit(const void *unused __unused)
796 {
797 
798 	ifc_detach_cloner(V_epair_cloner);
799 }
800 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
801     vnet_epair_uninit, NULL);
802 
803 static int
804 epair_mod_init(void)
805 {
806 	char name[32];
807 	epair_tasks.tasks = 0;
808 
809 #ifdef RSS
810 	int cpu;
811 
812 	CPU_FOREACH(cpu) {
813 		cpuset_t cpu_mask;
814 
815 		/* Pin to this CPU so we get appropriate NUMA allocations. */
816 		thread_lock(curthread);
817 		sched_bind(curthread, cpu);
818 		thread_unlock(curthread);
819 
820 		snprintf(name, sizeof(name), "epair_task_%d", cpu);
821 
822 		epair_tasks.tq[cpu] = taskqueue_create(name, M_WAITOK,
823 		    taskqueue_thread_enqueue,
824 		    &epair_tasks.tq[cpu]);
825 		CPU_SETOF(cpu, &cpu_mask);
826 		taskqueue_start_threads_cpuset(&epair_tasks.tq[cpu], 1, PI_NET,
827 		    &cpu_mask, "%s", name);
828 
829 		epair_tasks.tasks++;
830 	}
831 	thread_lock(curthread);
832 	sched_unbind(curthread);
833 	thread_unlock(curthread);
834 #else
835 	snprintf(name, sizeof(name), "epair_task");
836 
837 	epair_tasks.tq[0] = taskqueue_create(name, M_WAITOK,
838 	    taskqueue_thread_enqueue,
839 	    &epair_tasks.tq[0]);
840 	taskqueue_start_threads(&epair_tasks.tq[0], 1, PI_NET, "%s", name);
841 
842 	epair_tasks.tasks = 1;
843 #endif
844 
845 	return (0);
846 }
847 
848 static void
849 epair_mod_cleanup(void)
850 {
851 
852 	for (int i = 0; i < epair_tasks.tasks; i++) {
853 		taskqueue_drain_all(epair_tasks.tq[i]);
854 		taskqueue_free(epair_tasks.tq[i]);
855 	}
856 }
857 
858 static int
859 epair_modevent(module_t mod, int type, void *data)
860 {
861 	int ret;
862 
863 	switch (type) {
864 	case MOD_LOAD:
865 		EPAIR_LOCK_INIT();
866 		ret = epair_mod_init();
867 		if (ret != 0)
868 			return (ret);
869 		if (bootverbose)
870 			printf("%s: %s initialized.\n", __func__, epairname);
871 		break;
872 	case MOD_UNLOAD:
873 		epair_mod_cleanup();
874 		EPAIR_LOCK_DESTROY();
875 		if (bootverbose)
876 			printf("%s: %s unloaded.\n", __func__, epairname);
877 		break;
878 	default:
879 		return (EOPNOTSUPP);
880 	}
881 	return (0);
882 }
883 
884 static moduledata_t epair_mod = {
885 	"if_epair",
886 	epair_modevent,
887 	0
888 };
889 
890 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
891 MODULE_VERSION(if_epair, 3);
892