xref: /freebsd/sys/net/if_epair.c (revision 242cd60a0a670ff7cc446436bedd129fbdce062c)
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 int epair_clone_match(struct if_clone *, const char *);
88 static int epair_clone_create(struct if_clone *, char *, size_t, caddr_t);
89 static int epair_clone_destroy(struct if_clone *, struct ifnet *);
90 
91 static const char epairname[] = "epair";
92 #define	RXRSIZE	4096	/* Probably overkill by 4-8x. */
93 
94 static MALLOC_DEFINE(M_EPAIR, epairname,
95     "Pair of virtual cross-over connected Ethernet-like interfaces");
96 
97 VNET_DEFINE_STATIC(struct if_clone *, epair_cloner);
98 #define	V_epair_cloner	VNET(epair_cloner)
99 
100 static unsigned int next_index = 0;
101 #define	EPAIR_LOCK_INIT()		mtx_init(&epair_n_index_mtx, "epairidx", \
102 					    NULL, MTX_DEF)
103 #define	EPAIR_LOCK_DESTROY()		mtx_destroy(&epair_n_index_mtx)
104 #define	EPAIR_LOCK()			mtx_lock(&epair_n_index_mtx)
105 #define	EPAIR_UNLOCK()			mtx_unlock(&epair_n_index_mtx)
106 
107 #define BIT_QUEUE_TASK		0
108 #define BIT_MBUF_QUEUED		1
109 
110 struct epair_softc;
111 struct epair_queue {
112 	int			 id;
113 	struct buf_ring		*rxring[2];
114 	volatile int		 ridx;		/* 0 || 1 */
115 	volatile long		 state;		/* taskqueue coordination */
116 	struct task		 tx_task;
117 	struct epair_softc	*sc;
118 };
119 
120 static struct mtx epair_n_index_mtx;
121 struct epair_softc {
122 	struct ifnet		*ifp;		/* This ifp. */
123 	struct ifnet		*oifp;		/* other ifp of pair. */
124 	int			 num_queues;
125 	struct epair_queue	*queues;
126 	struct ifmedia		 media;		/* Media config (fake). */
127 	STAILQ_ENTRY(epair_softc) entry;
128 };
129 
130 struct epair_tasks_t {
131 	int			 tasks;
132 	struct taskqueue	 *tq[MAXCPU];
133 };
134 
135 static struct epair_tasks_t epair_tasks;
136 
137 static void
138 epair_clear_mbuf(struct mbuf *m)
139 {
140 	/* Remove any CSUM_SND_TAG as ether_input will barf. */
141 	if (m->m_pkthdr.csum_flags & CSUM_SND_TAG) {
142 		m_snd_tag_rele(m->m_pkthdr.snd_tag);
143 		m->m_pkthdr.snd_tag = NULL;
144 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
145 	}
146 
147 	m_tag_delete_nonpersistent(m);
148 }
149 
150 static void
151 epair_if_input(struct epair_softc *sc, struct epair_queue *q, int ridx)
152 {
153 	struct ifnet *ifp;
154 	struct mbuf *m;
155 
156 	ifp = sc->ifp;
157 	CURVNET_SET(ifp->if_vnet);
158 	while (! buf_ring_empty(q->rxring[ridx])) {
159 		m = buf_ring_dequeue_mc(q->rxring[ridx]);
160 		if (m == NULL)
161 			continue;
162 
163 		MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
164 		(*ifp->if_input)(ifp, m);
165 
166 	}
167 	CURVNET_RESTORE();
168 }
169 
170 static void
171 epair_tx_start_deferred(void *arg, int pending)
172 {
173 	struct epair_queue *q = (struct epair_queue *)arg;
174 	struct epair_softc *sc = q->sc;
175 	int ridx, nidx;
176 
177 	if_ref(sc->ifp);
178 	ridx = atomic_load_int(&q->ridx);
179 	do {
180 		nidx = (ridx == 0) ? 1 : 0;
181 	} while (!atomic_fcmpset_int(&q->ridx, &ridx, nidx));
182 	epair_if_input(sc, q, ridx);
183 
184 	atomic_clear_long(&q->state, (1 << BIT_QUEUE_TASK));
185 	if (atomic_testandclear_long(&q->state, BIT_MBUF_QUEUED))
186 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
187 
188 	if_rele(sc->ifp);
189 }
190 
191 static int
192 epair_menq(struct mbuf *m, struct epair_softc *osc)
193 {
194 	struct ifnet *ifp, *oifp;
195 	int len, ret;
196 	int ridx;
197 	short mflags;
198 	struct epair_queue *q = NULL;
199 	uint32_t bucket;
200 #ifdef RSS
201 	struct ether_header *eh;
202 #endif
203 
204 	/*
205 	 * I know this looks weird. We pass the "other sc" as we need that one
206 	 * and can get both ifps from it as well.
207 	 */
208 	oifp = osc->ifp;
209 	ifp = osc->oifp;
210 
211 	M_ASSERTPKTHDR(m);
212 	epair_clear_mbuf(m);
213 	if_setrcvif(m, oifp);
214 	M_SETFIB(m, oifp->if_fib);
215 
216 	/* Save values as once the mbuf is queued, it's not ours anymore. */
217 	len = m->m_pkthdr.len;
218 	mflags = m->m_flags;
219 
220 	MPASS(m->m_nextpkt == NULL);
221 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
222 
223 #ifdef RSS
224 	ret = rss_m2bucket(m, &bucket);
225 	if (ret) {
226 		/* Actually hash the packet. */
227 		eh = mtod(m, struct ether_header *);
228 
229 		switch (ntohs(eh->ether_type)) {
230 #ifdef INET
231 		case ETHERTYPE_IP:
232 			rss_soft_m2cpuid_v4(m, 0, &bucket);
233 			break;
234 #endif
235 #ifdef INET6
236 		case ETHERTYPE_IPV6:
237 			rss_soft_m2cpuid_v6(m, 0, &bucket);
238 			break;
239 #endif
240 		default:
241 			bucket = 0;
242 			break;
243 		}
244 	}
245 	bucket %= osc->num_queues;
246 #else
247 	bucket = 0;
248 #endif
249 	q = &osc->queues[bucket];
250 
251 	atomic_set_long(&q->state, (1 << BIT_MBUF_QUEUED));
252 	ridx = atomic_load_int(&q->ridx);
253 	ret = buf_ring_enqueue(q->rxring[ridx], m);
254 	if (ret != 0) {
255 		/* Ring is full. */
256 		if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
257 		m_freem(m);
258 		return (0);
259 	}
260 
261 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
262 	/*
263 	 * IFQ_HANDOFF_ADJ/ip_handoff() update statistics,
264 	 * but as we bypass all this we have to duplicate
265 	 * the logic another time.
266 	 */
267 	if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
268 	if (mflags & (M_BCAST|M_MCAST))
269 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
270 	/* Someone else received the packet. */
271 	if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
272 
273 	if (!atomic_testandset_long(&q->state, BIT_QUEUE_TASK))
274 		taskqueue_enqueue(epair_tasks.tq[bucket], &q->tx_task);
275 
276 	return (0);
277 }
278 
279 static void
280 epair_start(struct ifnet *ifp)
281 {
282 	struct mbuf *m;
283 	struct epair_softc *sc;
284 	struct ifnet *oifp;
285 
286 	/*
287 	 * We get packets here from ether_output via if_handoff()
288 	 * and need to put them into the input queue of the oifp
289 	 * and will put the packet into the receive-queue (rxq) of the
290 	 * other interface (oifp) of our pair.
291 	 */
292 	sc = ifp->if_softc;
293 	oifp = sc->oifp;
294 	sc = oifp->if_softc;
295 	for (;;) {
296 		IFQ_DEQUEUE(&ifp->if_snd, m);
297 		if (m == NULL)
298 			break;
299 		M_ASSERTPKTHDR(m);
300 		BPF_MTAP(ifp, m);
301 
302 		/* In case either interface is not usable drop the packet. */
303 		if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
304 		    (ifp->if_flags & IFF_UP) == 0 ||
305 		    (oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
306 		    (oifp->if_flags & IFF_UP) == 0) {
307 			m_freem(m);
308 			continue;
309 		}
310 
311 		(void) epair_menq(m, sc);
312 	}
313 }
314 
315 static int
316 epair_transmit(struct ifnet *ifp, struct mbuf *m)
317 {
318 	struct epair_softc *sc;
319 	struct ifnet *oifp;
320 	int error;
321 #ifdef ALTQ
322 	int len;
323 	short mflags;
324 #endif
325 
326 	if (m == NULL)
327 		return (0);
328 	M_ASSERTPKTHDR(m);
329 
330 	/*
331 	 * We are not going to use the interface en/dequeue mechanism
332 	 * on the TX side. We are called from ether_output_frame()
333 	 * and will put the packet into the receive-queue (rxq) of the
334 	 * other interface (oifp) of our pair.
335 	 */
336 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
337 		m_freem(m);
338 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
339 		return (ENXIO);
340 	}
341 	if ((ifp->if_flags & IFF_UP) == 0) {
342 		m_freem(m);
343 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
344 		return (ENETDOWN);
345 	}
346 
347 	BPF_MTAP(ifp, m);
348 
349 	/*
350 	 * In case the outgoing interface is not usable,
351 	 * drop the packet.
352 	 */
353 	sc = ifp->if_softc;
354 	oifp = sc->oifp;
355 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
356 	    (oifp->if_flags & IFF_UP) == 0) {
357 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
358 		m_freem(m);
359 		return (0);
360 	}
361 
362 #ifdef ALTQ
363 	len = m->m_pkthdr.len;
364 	mflags = m->m_flags;
365 
366 	/* Support ALTQ via the classic if_start() path. */
367 	IF_LOCK(&ifp->if_snd);
368 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
369 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
370 		if (error)
371 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
372 		IF_UNLOCK(&ifp->if_snd);
373 		if (!error) {
374 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
375 			if (mflags & (M_BCAST|M_MCAST))
376 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
377 			epair_start(ifp);
378 		}
379 		return (error);
380 	}
381 	IF_UNLOCK(&ifp->if_snd);
382 #endif
383 
384 	error = epair_menq(m, oifp->if_softc);
385 	return (error);
386 }
387 
388 static int
389 epair_media_change(struct ifnet *ifp __unused)
390 {
391 
392 	/* Do nothing. */
393 	return (0);
394 }
395 
396 static void
397 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
398 {
399 
400 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
401 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
402 }
403 
404 static int
405 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
406 {
407 	struct epair_softc *sc;
408 	struct ifreq *ifr;
409 	int error;
410 
411 	ifr = (struct ifreq *)data;
412 	switch (cmd) {
413 	case SIOCSIFFLAGS:
414 	case SIOCADDMULTI:
415 	case SIOCDELMULTI:
416 		error = 0;
417 		break;
418 
419 	case SIOCSIFMEDIA:
420 	case SIOCGIFMEDIA:
421 		sc = ifp->if_softc;
422 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
423 		break;
424 
425 	case SIOCSIFMTU:
426 		/* We basically allow all kinds of MTUs. */
427 		ifp->if_mtu = ifr->ifr_mtu;
428 		error = 0;
429 		break;
430 
431 	default:
432 		/* Let the common ethernet handler process this. */
433 		error = ether_ioctl(ifp, cmd, data);
434 		break;
435 	}
436 
437 	return (error);
438 }
439 
440 static void
441 epair_init(void *dummy __unused)
442 {
443 }
444 
445 /*
446  * Interface cloning functions.
447  * We use our private ones so that we can create/destroy our secondary
448  * device along with the primary one.
449  */
450 static int
451 epair_clone_match(struct if_clone *ifc, const char *name)
452 {
453 	const char *cp;
454 
455 	/*
456 	 * Our base name is epair.
457 	 * Our interfaces will be named epair<n>[ab].
458 	 * So accept anything of the following list:
459 	 * - epair
460 	 * - epair<n>
461 	 * but not the epair<n>[ab] versions.
462 	 */
463 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
464 		return (0);
465 
466 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
467 		if (*cp < '0' || *cp > '9')
468 			return (0);
469 	}
470 
471 	return (1);
472 }
473 
474 static void
475 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
476 {
477 	struct ifnet *ifp;
478 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
479 
480 	ifp = scb->ifp;
481 	/* Copy epairNa etheraddr and change the last byte. */
482 	memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
483 	eaddr[5] = 0x0b;
484 	ether_ifattach(ifp, eaddr);
485 
486 	if_clone_addif(ifc, ifp);
487 }
488 
489 static int
490 epair_clone_create(struct if_clone *ifc, char *name, size_t len, caddr_t params)
491 {
492 	struct epair_softc *sca, *scb;
493 	struct ifnet *ifp;
494 	char *dp;
495 	int error, unit, wildcard;
496 	uint64_t hostid;
497 	uint32_t key[3];
498 	uint32_t hash;
499 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
500 
501 	/* Try to see if a special unit was requested. */
502 	error = ifc_name2unit(name, &unit);
503 	if (error != 0)
504 		return (error);
505 	wildcard = (unit < 0);
506 
507 	error = ifc_alloc_unit(ifc, &unit);
508 	if (error != 0)
509 		return (error);
510 
511 	/*
512 	 * If no unit had been given, we need to adjust the ifName.
513 	 * Also make sure there is space for our extra [ab] suffix.
514 	 */
515 	for (dp = name; *dp != '\0'; dp++);
516 	if (wildcard) {
517 		error = snprintf(dp, len - (dp - name), "%d", unit);
518 		if (error > len - (dp - name) - 1) {
519 			/* ifName too long. */
520 			ifc_free_unit(ifc, unit);
521 			return (ENOSPC);
522 		}
523 		dp += error;
524 	}
525 	if (len - (dp - name) - 1 < 1) {
526 		/* No space left for our [ab] suffix. */
527 		ifc_free_unit(ifc, unit);
528 		return (ENOSPC);
529 	}
530 	*dp = 'b';
531 	/* Must not change dp so we can replace 'a' by 'b' later. */
532 	*(dp+1) = '\0';
533 
534 	/* Check if 'a' and 'b' interfaces already exist. */
535 	if (ifunit(name) != NULL)
536 		return (EEXIST);
537 	*dp = 'a';
538 	if (ifunit(name) != NULL)
539 		return (EEXIST);
540 
541 	/* Allocate memory for both [ab] interfaces */
542 	sca = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
543 	sca->ifp = if_alloc(IFT_ETHER);
544 	sca->num_queues = epair_tasks.tasks;
545 	if (sca->ifp == NULL) {
546 		free(sca, M_EPAIR);
547 		ifc_free_unit(ifc, unit);
548 		return (ENOSPC);
549 	}
550 	sca->queues = mallocarray(sca->num_queues, sizeof(struct epair_queue),
551 	    M_EPAIR, M_WAITOK);
552 	for (int i = 0; i < sca->num_queues; i++) {
553 		struct epair_queue *q = &sca->queues[i];
554 		q->id = i;
555 		q->rxring[0] = buf_ring_alloc(RXRSIZE, M_EPAIR, M_WAITOK, NULL);
556 		q->rxring[1] = buf_ring_alloc(RXRSIZE, M_EPAIR, M_WAITOK, NULL);
557 		q->ridx = 0;
558 		q->state = 0;
559 		q->sc = sca;
560 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
561 	}
562 
563 	scb = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
564 	scb->ifp = if_alloc(IFT_ETHER);
565 	scb->num_queues = epair_tasks.tasks;
566 	if (scb->ifp == NULL) {
567 		free(scb, M_EPAIR);
568 		if_free(sca->ifp);
569 		free(sca, M_EPAIR);
570 		ifc_free_unit(ifc, unit);
571 		return (ENOSPC);
572 	}
573 	scb->queues = mallocarray(scb->num_queues, sizeof(struct epair_queue),
574 	    M_EPAIR, M_WAITOK);
575 	for (int i = 0; i < scb->num_queues; i++) {
576 		struct epair_queue *q = &scb->queues[i];
577 		q->id = i;
578 		q->rxring[0] = buf_ring_alloc(RXRSIZE, M_EPAIR, M_WAITOK, NULL);
579 		q->rxring[1] = buf_ring_alloc(RXRSIZE, M_EPAIR, M_WAITOK, NULL);
580 		q->ridx = 0;
581 		q->state = 0;
582 		q->sc = scb;
583 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
584 	}
585 
586 	/*
587 	 * Cross-reference the interfaces so we will be able to free both.
588 	 */
589 	sca->oifp = scb->ifp;
590 	scb->oifp = sca->ifp;
591 
592 	EPAIR_LOCK();
593 #ifdef SMP
594 	/* Get an approximate distribution. */
595 	hash = next_index % mp_ncpus;
596 #else
597 	hash = 0;
598 #endif
599 	EPAIR_UNLOCK();
600 
601 	/* Initialise pseudo media types. */
602 	ifmedia_init(&sca->media, 0, epair_media_change, epair_media_status);
603 	ifmedia_add(&sca->media, IFM_ETHER | IFM_10G_T, 0, NULL);
604 	ifmedia_set(&sca->media, IFM_ETHER | IFM_10G_T);
605 	ifmedia_init(&scb->media, 0, epair_media_change, epair_media_status);
606 	ifmedia_add(&scb->media, IFM_ETHER | IFM_10G_T, 0, NULL);
607 	ifmedia_set(&scb->media, IFM_ETHER | IFM_10G_T);
608 
609 	/* Finish initialization of interface <n>a. */
610 	ifp = sca->ifp;
611 	ifp->if_softc = sca;
612 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
613 	ifp->if_dname = epairname;
614 	ifp->if_dunit = unit;
615 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
616 	ifp->if_flags |= IFF_KNOWSEPOCH;
617 	ifp->if_capabilities = IFCAP_VLAN_MTU;
618 	ifp->if_capenable = IFCAP_VLAN_MTU;
619 	ifp->if_start = epair_start;
620 	ifp->if_ioctl = epair_ioctl;
621 	ifp->if_init  = epair_init;
622 	if_setsendqlen(ifp, ifqmaxlen);
623 	if_setsendqready(ifp);
624 
625 	/*
626 	 * Calculate the etheraddr hashing the hostid and the
627 	 * interface index. The result would be hopefully unique.
628 	 * Note that the "a" component of an epair instance may get moved
629 	 * to a different VNET after creation. In that case its index
630 	 * will be freed and the index can get reused by new epair instance.
631 	 * Make sure we do not create same etheraddr again.
632 	 */
633 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
634 	if (hostid == 0)
635 		arc4rand(&hostid, sizeof(hostid), 0);
636 
637 	EPAIR_LOCK();
638 	if (ifp->if_index > next_index)
639 		next_index = ifp->if_index;
640 	else
641 		next_index++;
642 
643 	key[0] = (uint32_t)next_index;
644 	EPAIR_UNLOCK();
645 	key[1] = (uint32_t)(hostid & 0xffffffff);
646 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
647 	hash = jenkins_hash32(key, 3, 0);
648 
649 	eaddr[0] = 0x02;
650 	memcpy(&eaddr[1], &hash, 4);
651 	eaddr[5] = 0x0a;
652 	ether_ifattach(ifp, eaddr);
653 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
654 	ifp->if_transmit = epair_transmit;
655 
656 	/* Swap the name and finish initialization of interface <n>b. */
657 	*dp = 'b';
658 
659 	ifp = scb->ifp;
660 	ifp->if_softc = scb;
661 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
662 	ifp->if_dname = epairname;
663 	ifp->if_dunit = unit;
664 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
665 	ifp->if_flags |= IFF_KNOWSEPOCH;
666 	ifp->if_capabilities = IFCAP_VLAN_MTU;
667 	ifp->if_capenable = IFCAP_VLAN_MTU;
668 	ifp->if_start = epair_start;
669 	ifp->if_ioctl = epair_ioctl;
670 	ifp->if_init  = epair_init;
671 	if_setsendqlen(ifp, ifqmaxlen);
672 	if_setsendqready(ifp);
673 	/* We need to play some tricks here for the second interface. */
674 	strlcpy(name, epairname, len);
675 
676 	/* Correctly set the name for the cloner list. */
677 	strlcpy(name, scb->ifp->if_xname, len);
678 	epair_clone_add(ifc, scb);
679 
680 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
681 	ifp->if_transmit = epair_transmit;
682 
683 	/*
684 	 * Restore name to <n>a as the ifp for this will go into the
685 	 * cloner list for the initial call.
686 	 */
687 	strlcpy(name, sca->ifp->if_xname, len);
688 
689 	/* Tell the world, that we are ready to rock. */
690 	sca->ifp->if_drv_flags |= IFF_DRV_RUNNING;
691 	if_link_state_change(sca->ifp, LINK_STATE_UP);
692 	scb->ifp->if_drv_flags |= IFF_DRV_RUNNING;
693 	if_link_state_change(scb->ifp, LINK_STATE_UP);
694 
695 	return (0);
696 }
697 
698 static void
699 epair_drain_rings(struct epair_softc *sc)
700 {
701 	int ridx;
702 	struct mbuf *m;
703 
704 	for (ridx = 0; ridx < 2; ridx++) {
705 		for (int i = 0; i < sc->num_queues; i++) {
706 			struct epair_queue *q = &sc->queues[i];
707 			do {
708 				m = buf_ring_dequeue_sc(q->rxring[ridx]);
709 				if (m == NULL)
710 					break;
711 				m_freem(m);
712 			} while (1);
713 		}
714 	}
715 }
716 
717 static int
718 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp)
719 {
720 	struct ifnet *oifp;
721 	struct epair_softc *sca, *scb;
722 	int unit, error;
723 
724 	/*
725 	 * In case we called into if_clone_destroyif() ourselves
726 	 * again to remove the second interface, the softc will be
727 	 * NULL. In that case so not do anything but return success.
728 	 */
729 	if (ifp->if_softc == NULL)
730 		return (0);
731 
732 	unit = ifp->if_dunit;
733 	sca = ifp->if_softc;
734 	oifp = sca->oifp;
735 	scb = oifp->if_softc;
736 
737 	/* Frist get the interfaces down and detached. */
738 	if_link_state_change(ifp, LINK_STATE_DOWN);
739 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
740 	if_link_state_change(oifp, LINK_STATE_DOWN);
741 	oifp->if_drv_flags &= ~IFF_DRV_RUNNING;
742 
743 	ether_ifdetach(ifp);
744 	ether_ifdetach(oifp);
745 
746 	/* Third free any queued packets and all the resources. */
747 	CURVNET_SET_QUIET(oifp->if_vnet);
748 	epair_drain_rings(scb);
749 	oifp->if_softc = NULL;
750 	error = if_clone_destroyif(ifc, oifp);
751 	if (error)
752 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
753 		    __func__, error);
754 	if_free(oifp);
755 	ifmedia_removeall(&scb->media);
756 	for (int i = 0; i < scb->num_queues; i++) {
757 		struct epair_queue *q = &scb->queues[i];
758 		buf_ring_free(q->rxring[0], M_EPAIR);
759 		buf_ring_free(q->rxring[1], M_EPAIR);
760 	}
761 	free(scb->queues, M_EPAIR);
762 	free(scb, M_EPAIR);
763 	CURVNET_RESTORE();
764 
765 	epair_drain_rings(sca);
766 	if_free(ifp);
767 	ifmedia_removeall(&sca->media);
768 	for (int i = 0; i < sca->num_queues; i++) {
769 		struct epair_queue *q = &sca->queues[i];
770 		buf_ring_free(q->rxring[0], M_EPAIR);
771 		buf_ring_free(q->rxring[1], M_EPAIR);
772 	}
773 	free(sca->queues, M_EPAIR);
774 	free(sca, M_EPAIR);
775 
776 	/* Last free the cloner unit. */
777 	ifc_free_unit(ifc, unit);
778 
779 	return (0);
780 }
781 
782 static void
783 vnet_epair_init(const void *unused __unused)
784 {
785 
786 	V_epair_cloner = if_clone_advanced(epairname, 0,
787 	    epair_clone_match, epair_clone_create, epair_clone_destroy);
788 }
789 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
790     vnet_epair_init, NULL);
791 
792 static void
793 vnet_epair_uninit(const void *unused __unused)
794 {
795 
796 	if_clone_detach(V_epair_cloner);
797 }
798 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
799     vnet_epair_uninit, NULL);
800 
801 static int
802 epair_mod_init(void)
803 {
804 	char name[32];
805 	epair_tasks.tasks = 0;
806 
807 #ifdef RSS
808 	struct pcpu *pcpu;
809 	int cpu;
810 
811 	CPU_FOREACH(cpu) {
812 		cpuset_t cpu_mask;
813 
814 		/* Pin to this CPU so we get appropriate NUMA allocations. */
815 		pcpu = pcpu_find(cpu);
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 #else
832 	snprintf(name, sizeof(name), "epair_task");
833 
834 	epair_tasks.tq[0] = taskqueue_create(name, M_WAITOK,
835 	    taskqueue_thread_enqueue,
836 	    &epair_tasks.tq[0]);
837 	taskqueue_start_threads(&epair_tasks.tq[0], 1, PI_NET, "%s", name);
838 
839 	epair_tasks.tasks = 1;
840 #endif
841 
842 	return (0);
843 }
844 
845 static void
846 epair_mod_cleanup(void)
847 {
848 
849 	for (int i = 0; i < epair_tasks.tasks; i++) {
850 		taskqueue_drain_all(epair_tasks.tq[i]);
851 		taskqueue_free(epair_tasks.tq[i]);
852 	}
853 }
854 
855 static int
856 epair_modevent(module_t mod, int type, void *data)
857 {
858 	int ret;
859 
860 	switch (type) {
861 	case MOD_LOAD:
862 		EPAIR_LOCK_INIT();
863 		ret = epair_mod_init();
864 		if (ret != 0)
865 			return (ret);
866 		if (bootverbose)
867 			printf("%s: %s initialized.\n", __func__, epairname);
868 		break;
869 	case MOD_UNLOAD:
870 		epair_mod_cleanup();
871 		EPAIR_LOCK_DESTROY();
872 		if (bootverbose)
873 			printf("%s: %s unloaded.\n", __func__, epairname);
874 		break;
875 	default:
876 		return (EOPNOTSUPP);
877 	}
878 	return (0);
879 }
880 
881 static moduledata_t epair_mod = {
882 	"if_epair",
883 	epair_modevent,
884 	0
885 };
886 
887 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
888 MODULE_VERSION(if_epair, 3);
889