xref: /freebsd/sys/net/if_epair.c (revision 64c2a712d661db9be31f02fe97c3b59710290ae3)
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_private.h>
75 #include <net/if_types.h>
76 #include <net/netisr.h>
77 #ifdef RSS
78 #include <net/rss_config.h>
79 #ifdef INET
80 #include <netinet/in_rss.h>
81 #endif
82 #ifdef INET6
83 #include <netinet6/in6_rss.h>
84 #endif
85 #endif
86 #include <net/vnet.h>
87 
88 static const char epairname[] = "epair";
89 #define	RXRSIZE	4096	/* Probably overkill by 4-8x. */
90 
91 static MALLOC_DEFINE(M_EPAIR, epairname,
92     "Pair of virtual cross-over connected Ethernet-like interfaces");
93 
94 VNET_DEFINE_STATIC(struct if_clone *, epair_cloner);
95 #define	V_epair_cloner	VNET(epair_cloner)
96 
97 static unsigned int next_index = 0;
98 #define	EPAIR_LOCK_INIT()		mtx_init(&epair_n_index_mtx, "epairidx", \
99 					    NULL, MTX_DEF)
100 #define	EPAIR_LOCK_DESTROY()		mtx_destroy(&epair_n_index_mtx)
101 #define	EPAIR_LOCK()			mtx_lock(&epair_n_index_mtx)
102 #define	EPAIR_UNLOCK()			mtx_unlock(&epair_n_index_mtx)
103 
104 struct epair_softc;
105 struct epair_queue {
106 	struct mtx		 mtx;
107 	struct mbufq		 q;
108 	int			 id;
109 	enum {
110 		EPAIR_QUEUE_IDLE,
111 		EPAIR_QUEUE_WAKING,
112 		EPAIR_QUEUE_RUNNING,
113 	}			 state;
114 	struct task		 tx_task;
115 	struct epair_softc	*sc;
116 };
117 
118 static struct mtx epair_n_index_mtx;
119 struct epair_softc {
120 	struct ifnet		*ifp;		/* This ifp. */
121 	struct ifnet		*oifp;		/* other ifp of pair. */
122 	int			 num_queues;
123 	struct epair_queue	*queues;
124 	struct ifmedia		 media;		/* Media config (fake). */
125 	STAILQ_ENTRY(epair_softc) entry;
126 };
127 
128 struct epair_tasks_t {
129 	int			 tasks;
130 	struct taskqueue	 *tq[MAXCPU];
131 };
132 
133 static struct epair_tasks_t epair_tasks;
134 
135 static void
136 epair_clear_mbuf(struct mbuf *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 	m_tag_delete_nonpersistent(m);
146 }
147 
148 static void
149 epair_tx_start_deferred(void *arg, int pending)
150 {
151 	struct epair_queue *q = (struct epair_queue *)arg;
152 	if_t ifp;
153 	struct mbuf *m, *n;
154 	bool resched;
155 
156 	ifp = q->sc->ifp;
157 
158 	if_ref(ifp);
159 	CURVNET_SET(ifp->if_vnet);
160 
161 	mtx_lock(&q->mtx);
162 	m = mbufq_flush(&q->q);
163 	q->state = EPAIR_QUEUE_RUNNING;
164 	mtx_unlock(&q->mtx);
165 
166 	while (m != NULL) {
167 		n = STAILQ_NEXT(m, m_stailqpkt);
168 		m->m_nextpkt = NULL;
169 		if_input(ifp, m);
170 		m = n;
171 	}
172 
173 	/*
174 	 * Avoid flushing the queue more than once per task.  We can otherwise
175 	 * end up starving ourselves in a multi-epair routing configuration.
176 	 */
177 	mtx_lock(&q->mtx);
178 	if (mbufq_len(&q->q) > 0) {
179 		resched = true;
180 		q->state = EPAIR_QUEUE_WAKING;
181 	} else {
182 		resched = false;
183 		q->state = EPAIR_QUEUE_IDLE;
184 	}
185 	mtx_unlock(&q->mtx);
186 
187 	if (resched)
188 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
189 
190 	CURVNET_RESTORE();
191 	if_rele(ifp);
192 }
193 
194 static struct epair_queue *
195 epair_select_queue(struct epair_softc *sc, struct mbuf *m)
196 {
197 	uint32_t bucket;
198 #ifdef RSS
199 	struct ether_header *eh;
200 	int ret;
201 
202 	ret = rss_m2bucket(m, &bucket);
203 	if (ret) {
204 		/* Actually hash the packet. */
205 		eh = mtod(m, struct ether_header *);
206 
207 		switch (ntohs(eh->ether_type)) {
208 #ifdef INET
209 		case ETHERTYPE_IP:
210 			rss_soft_m2cpuid_v4(m, 0, &bucket);
211 			break;
212 #endif
213 #ifdef INET6
214 		case ETHERTYPE_IPV6:
215 			rss_soft_m2cpuid_v6(m, 0, &bucket);
216 			break;
217 #endif
218 		default:
219 			bucket = 0;
220 			break;
221 		}
222 	}
223 	bucket %= sc->num_queues;
224 #else
225 	bucket = 0;
226 #endif
227 	return (&sc->queues[bucket]);
228 }
229 
230 static void
231 epair_prepare_mbuf(struct mbuf *m, struct ifnet *src_ifp)
232 {
233 	M_ASSERTPKTHDR(m);
234 	epair_clear_mbuf(m);
235 	if_setrcvif(m, src_ifp);
236 	M_SETFIB(m, src_ifp->if_fib);
237 
238 	MPASS(m->m_nextpkt == NULL);
239 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
240 }
241 
242 static void
243 epair_menq(struct mbuf *m, struct epair_softc *osc)
244 {
245 	struct epair_queue *q;
246 	struct ifnet *ifp, *oifp;
247 	int error, len;
248 	bool mcast;
249 
250 	/*
251 	 * I know this looks weird. We pass the "other sc" as we need that one
252 	 * and can get both ifps from it as well.
253 	 */
254 	oifp = osc->ifp;
255 	ifp = osc->oifp;
256 
257 	epair_prepare_mbuf(m, oifp);
258 
259 	/* Save values as once the mbuf is queued, it's not ours anymore. */
260 	len = m->m_pkthdr.len;
261 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
262 
263 	q = epair_select_queue(osc, m);
264 
265 	mtx_lock(&q->mtx);
266 	if (q->state == EPAIR_QUEUE_IDLE) {
267 		q->state = EPAIR_QUEUE_WAKING;
268 		taskqueue_enqueue(epair_tasks.tq[q->id], &q->tx_task);
269 	}
270 	error = mbufq_enqueue(&q->q, m);
271 	mtx_unlock(&q->mtx);
272 
273 	if (error != 0) {
274 		m_freem(m);
275 		if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
276 	} else {
277 		if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
278 		if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
279 		if (mcast)
280 			if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
281 		if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
282 	}
283 }
284 
285 static void
286 epair_start(struct ifnet *ifp)
287 {
288 	struct mbuf *m;
289 	struct epair_softc *sc;
290 	struct ifnet *oifp;
291 
292 	/*
293 	 * We get packets here from ether_output via if_handoff()
294 	 * and need to put them into the input queue of the oifp
295 	 * and will put the packet into the receive-queue (rxq) of the
296 	 * other interface (oifp) of our pair.
297 	 */
298 	sc = ifp->if_softc;
299 	oifp = sc->oifp;
300 	sc = oifp->if_softc;
301 	for (;;) {
302 		IFQ_DEQUEUE(&ifp->if_snd, m);
303 		if (m == NULL)
304 			break;
305 		M_ASSERTPKTHDR(m);
306 		BPF_MTAP(ifp, m);
307 
308 		/* In case either interface is not usable drop the packet. */
309 		if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
310 		    (ifp->if_flags & IFF_UP) == 0 ||
311 		    (oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
312 		    (oifp->if_flags & IFF_UP) == 0) {
313 			m_freem(m);
314 			continue;
315 		}
316 
317 		epair_menq(m, sc);
318 	}
319 }
320 
321 static int
322 epair_transmit(struct ifnet *ifp, struct mbuf *m)
323 {
324 	struct epair_softc *sc;
325 	struct ifnet *oifp;
326 #ifdef ALTQ
327 	int len;
328 	bool mcast;
329 #endif
330 
331 	if (m == NULL)
332 		return (0);
333 	M_ASSERTPKTHDR(m);
334 
335 	/*
336 	 * We are not going to use the interface en/dequeue mechanism
337 	 * on the TX side. We are called from ether_output_frame()
338 	 * and will put the packet into the receive-queue (rxq) of the
339 	 * other interface (oifp) of our pair.
340 	 */
341 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
342 		m_freem(m);
343 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
344 		return (ENXIO);
345 	}
346 	if ((ifp->if_flags & IFF_UP) == 0) {
347 		m_freem(m);
348 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
349 		return (ENETDOWN);
350 	}
351 
352 	BPF_MTAP(ifp, m);
353 
354 	/*
355 	 * In case the outgoing interface is not usable,
356 	 * drop the packet.
357 	 */
358 	sc = ifp->if_softc;
359 	oifp = sc->oifp;
360 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
361 	    (oifp->if_flags & IFF_UP) == 0) {
362 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
363 		m_freem(m);
364 		return (0);
365 	}
366 
367 #ifdef ALTQ
368 	len = m->m_pkthdr.len;
369 	mcast = (m->m_flags & (M_BCAST | M_MCAST)) != 0;
370 	int error = 0;
371 
372 	/* Support ALTQ via the classic if_start() path. */
373 	IF_LOCK(&ifp->if_snd);
374 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
375 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
376 		if (error)
377 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
378 		IF_UNLOCK(&ifp->if_snd);
379 		if (!error) {
380 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
381 			if (mcast)
382 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
383 			epair_start(ifp);
384 		}
385 		return (error);
386 	}
387 	IF_UNLOCK(&ifp->if_snd);
388 #endif
389 
390 	epair_menq(m, oifp->if_softc);
391 	return (0);
392 }
393 
394 static void
395 epair_qflush(struct ifnet *ifp __unused)
396 {
397 }
398 
399 static int
400 epair_media_change(struct ifnet *ifp __unused)
401 {
402 
403 	/* Do nothing. */
404 	return (0);
405 }
406 
407 static void
408 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
409 {
410 
411 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
412 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
413 }
414 
415 static int
416 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
417 {
418 	struct epair_softc *sc;
419 	struct ifreq *ifr;
420 	int error;
421 
422 	ifr = (struct ifreq *)data;
423 	switch (cmd) {
424 	case SIOCSIFFLAGS:
425 	case SIOCADDMULTI:
426 	case SIOCDELMULTI:
427 		error = 0;
428 		break;
429 
430 	case SIOCSIFMEDIA:
431 	case SIOCGIFMEDIA:
432 		sc = ifp->if_softc;
433 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
434 		break;
435 
436 	case SIOCSIFMTU:
437 		/* We basically allow all kinds of MTUs. */
438 		ifp->if_mtu = ifr->ifr_mtu;
439 		error = 0;
440 		break;
441 
442 	default:
443 		/* Let the common ethernet handler process this. */
444 		error = ether_ioctl(ifp, cmd, data);
445 		break;
446 	}
447 
448 	return (error);
449 }
450 
451 static void
452 epair_init(void *dummy __unused)
453 {
454 }
455 
456 /*
457  * Interface cloning functions.
458  * We use our private ones so that we can create/destroy our secondary
459  * device along with the primary one.
460  */
461 static int
462 epair_clone_match(struct if_clone *ifc, const char *name)
463 {
464 	const char *cp;
465 
466 	/*
467 	 * Our base name is epair.
468 	 * Our interfaces will be named epair<n>[ab].
469 	 * So accept anything of the following list:
470 	 * - epair
471 	 * - epair<n>
472 	 * but not the epair<n>[ab] versions.
473 	 */
474 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
475 		return (0);
476 
477 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
478 		if (*cp < '0' || *cp > '9')
479 			return (0);
480 	}
481 
482 	return (1);
483 }
484 
485 static void
486 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
487 {
488 	struct ifnet *ifp;
489 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
490 
491 	ifp = scb->ifp;
492 	/* Copy epairNa etheraddr and change the last byte. */
493 	memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
494 	eaddr[5] = 0x0b;
495 	ether_ifattach(ifp, eaddr);
496 
497 	if_clone_addif(ifc, ifp);
498 }
499 
500 static struct epair_softc *
501 epair_alloc_sc(struct if_clone *ifc)
502 {
503 	struct epair_softc *sc;
504 
505 	struct ifnet *ifp = if_alloc(IFT_ETHER);
506 	if (ifp == NULL)
507 		return (NULL);
508 
509 	sc = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
510 	sc->ifp = ifp;
511 	sc->num_queues = epair_tasks.tasks;
512 	sc->queues = mallocarray(sc->num_queues, sizeof(struct epair_queue),
513 	    M_EPAIR, M_WAITOK);
514 	for (int i = 0; i < sc->num_queues; i++) {
515 		struct epair_queue *q = &sc->queues[i];
516 		q->id = i;
517 		q->state = EPAIR_QUEUE_IDLE;
518 		mtx_init(&q->mtx, "epairq", NULL, MTX_DEF | MTX_NEW);
519 		mbufq_init(&q->q, RXRSIZE);
520 		q->sc = sc;
521 		NET_TASK_INIT(&q->tx_task, 0, epair_tx_start_deferred, q);
522 	}
523 
524 	/* Initialise pseudo media types. */
525 	ifmedia_init(&sc->media, 0, epair_media_change, epair_media_status);
526 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T, 0, NULL);
527 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T);
528 
529 	return (sc);
530 }
531 
532 static void
533 epair_setup_ifp(struct epair_softc *sc, char *name, int unit)
534 {
535 	struct ifnet *ifp = sc->ifp;
536 
537 	ifp->if_softc = sc;
538 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
539 	ifp->if_dname = epairname;
540 	ifp->if_dunit = unit;
541 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
542 	ifp->if_flags |= IFF_KNOWSEPOCH;
543 	ifp->if_capabilities = IFCAP_VLAN_MTU;
544 	ifp->if_capenable = IFCAP_VLAN_MTU;
545 	ifp->if_transmit = epair_transmit;
546 	ifp->if_qflush = epair_qflush;
547 	ifp->if_start = epair_start;
548 	ifp->if_ioctl = epair_ioctl;
549 	ifp->if_init  = epair_init;
550 	if_setsendqlen(ifp, ifqmaxlen);
551 	if_setsendqready(ifp);
552 
553 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
554 }
555 
556 static void
557 epair_generate_mac(struct epair_softc *sc, uint8_t *eaddr)
558 {
559 	uint32_t key[3];
560 	uint32_t hash;
561 	uint64_t hostid;
562 
563 	EPAIR_LOCK();
564 #ifdef SMP
565 	/* Get an approximate distribution. */
566 	hash = next_index % mp_ncpus;
567 #else
568 	hash = 0;
569 #endif
570 	EPAIR_UNLOCK();
571 
572 	/*
573 	 * Calculate the etheraddr hashing the hostid and the
574 	 * interface index. The result would be hopefully unique.
575 	 * Note that the "a" component of an epair instance may get moved
576 	 * to a different VNET after creation. In that case its index
577 	 * will be freed and the index can get reused by new epair instance.
578 	 * Make sure we do not create same etheraddr again.
579 	 */
580 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
581 	if (hostid == 0)
582 		arc4rand(&hostid, sizeof(hostid), 0);
583 
584 	struct ifnet *ifp = sc->ifp;
585 	EPAIR_LOCK();
586 	if (ifp->if_index > next_index)
587 		next_index = ifp->if_index;
588 	else
589 		next_index++;
590 
591 	key[0] = (uint32_t)next_index;
592 	EPAIR_UNLOCK();
593 	key[1] = (uint32_t)(hostid & 0xffffffff);
594 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
595 	hash = jenkins_hash32(key, 3, 0);
596 
597 	eaddr[0] = 0x02;
598 	memcpy(&eaddr[1], &hash, 4);
599 	eaddr[5] = 0x0a;
600 }
601 
602 static void
603 epair_free_sc(struct epair_softc *sc)
604 {
605 	if (sc == NULL)
606 		return;
607 
608 	if_free(sc->ifp);
609 	ifmedia_removeall(&sc->media);
610 	for (int i = 0; i < sc->num_queues; i++) {
611 		struct epair_queue *q = &sc->queues[i];
612 		mtx_destroy(&q->mtx);
613 	}
614 	free(sc->queues, M_EPAIR);
615 	free(sc, M_EPAIR);
616 }
617 
618 static void
619 epair_set_state(struct ifnet *ifp, bool running)
620 {
621 	if (running) {
622 		ifp->if_drv_flags |= IFF_DRV_RUNNING;
623 		if_link_state_change(ifp, LINK_STATE_UP);
624 	} else {
625 		if_link_state_change(ifp, LINK_STATE_DOWN);
626 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
627 	}
628 }
629 
630 static int
631 epair_handle_unit(struct if_clone *ifc, char *name, size_t len, int *punit)
632 {
633 	int error = 0, unit, wildcard;
634 	char *dp;
635 
636 	/* Try to see if a special unit was requested. */
637 	error = ifc_name2unit(name, &unit);
638 	if (error != 0)
639 		return (error);
640 	wildcard = (unit < 0);
641 
642 	error = ifc_alloc_unit(ifc, &unit);
643 	if (error != 0)
644 		return (error);
645 
646 	/*
647 	 * If no unit had been given, we need to adjust the ifName.
648 	 * Also make sure there is space for our extra [ab] suffix.
649 	 */
650 	for (dp = name; *dp != '\0'; dp++);
651 	if (wildcard) {
652 		int slen = snprintf(dp, len - (dp - name), "%d", unit);
653 		if (slen > len - (dp - name) - 1) {
654 			/* ifName too long. */
655 			error = ENOSPC;
656 			goto done;
657 		}
658 		dp += slen;
659 	}
660 	if (len - (dp - name) - 1 < 1) {
661 		/* No space left for our [ab] suffix. */
662 		error = ENOSPC;
663 		goto done;
664 	}
665 	*dp = 'b';
666 	/* Must not change dp so we can replace 'a' by 'b' later. */
667 	*(dp+1) = '\0';
668 
669 	/* Check if 'a' and 'b' interfaces already exist. */
670 	if (ifunit(name) != NULL) {
671 		error = EEXIST;
672 		goto done;
673 	}
674 
675 	*dp = 'a';
676 	if (ifunit(name) != NULL) {
677 		error = EEXIST;
678 		goto done;
679 	}
680 	*punit = unit;
681 done:
682 	if (error != 0)
683 		ifc_free_unit(ifc, unit);
684 
685 	return (error);
686 }
687 
688 static int
689 epair_clone_create(struct if_clone *ifc, char *name, size_t len,
690     struct ifc_data *ifd, struct ifnet **ifpp)
691 {
692 	struct epair_softc *sca, *scb;
693 	struct ifnet *ifp;
694 	char *dp;
695 	int error, unit;
696 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
697 
698 	error = epair_handle_unit(ifc, name, len, &unit);
699 	if (error != 0)
700 		return (error);
701 
702 	/* Allocate memory for both [ab] interfaces */
703 	sca = epair_alloc_sc(ifc);
704 	scb = epair_alloc_sc(ifc);
705 	if (sca == NULL || scb == NULL) {
706 		epair_free_sc(sca);
707 		epair_free_sc(scb);
708 		ifc_free_unit(ifc, unit);
709 		return (ENOSPC);
710 	}
711 
712 	/*
713 	 * Cross-reference the interfaces so we will be able to free both.
714 	 */
715 	sca->oifp = scb->ifp;
716 	scb->oifp = sca->ifp;
717 
718 	/* Finish initialization of interface <n>a. */
719 	ifp = sca->ifp;
720 	epair_setup_ifp(sca, name, unit);
721 	epair_generate_mac(sca, eaddr);
722 
723 	ether_ifattach(ifp, eaddr);
724 
725 	/* Swap the name and finish initialization of interface <n>b. */
726 	dp = name + strlen(name) - 1;
727 	*dp = 'b';
728 
729 	epair_setup_ifp(scb, name, unit);
730 
731 	ifp = scb->ifp;
732 	/* We need to play some tricks here for the second interface. */
733 	strlcpy(name, epairname, len);
734 	/* Correctly set the name for the cloner list. */
735 	strlcpy(name, scb->ifp->if_xname, len);
736 
737 	epair_clone_add(ifc, scb);
738 
739 	/*
740 	 * Restore name to <n>a as the ifp for this will go into the
741 	 * cloner list for the initial call.
742 	 */
743 	strlcpy(name, sca->ifp->if_xname, len);
744 
745 	/* Tell the world, that we are ready to rock. */
746 	epair_set_state(sca->ifp, true);
747 	epair_set_state(scb->ifp, true);
748 
749 	*ifpp = sca->ifp;
750 
751 	return (0);
752 }
753 
754 static void
755 epair_drain_rings(struct epair_softc *sc)
756 {
757 	for (int i = 0; i < sc->num_queues; i++) {
758 		struct epair_queue *q;
759 		struct mbuf *m, *n;
760 
761 		q = &sc->queues[i];
762 		mtx_lock(&q->mtx);
763 		m = mbufq_flush(&q->q);
764 		mtx_unlock(&q->mtx);
765 
766 		for (; m != NULL; m = n) {
767 			n = m->m_nextpkt;
768 			m_freem(m);
769 		}
770 	}
771 }
772 
773 static int
774 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp, uint32_t flags)
775 {
776 	struct ifnet *oifp;
777 	struct epair_softc *sca, *scb;
778 	int unit, error;
779 
780 	/*
781 	 * In case we called into if_clone_destroyif() ourselves
782 	 * again to remove the second interface, the softc will be
783 	 * NULL. In that case so not do anything but return success.
784 	 */
785 	if (ifp->if_softc == NULL)
786 		return (0);
787 
788 	unit = ifp->if_dunit;
789 	sca = ifp->if_softc;
790 	oifp = sca->oifp;
791 	scb = oifp->if_softc;
792 
793 	/* Frist get the interfaces down and detached. */
794 	epair_set_state(ifp, false);
795 	epair_set_state(oifp, false);
796 
797 	ether_ifdetach(ifp);
798 	ether_ifdetach(oifp);
799 
800 	/* Third free any queued packets and all the resources. */
801 	CURVNET_SET_QUIET(oifp->if_vnet);
802 	epair_drain_rings(scb);
803 	oifp->if_softc = NULL;
804 	error = if_clone_destroyif(ifc, oifp);
805 	if (error)
806 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
807 		    __func__, error);
808 	epair_free_sc(scb);
809 	CURVNET_RESTORE();
810 
811 	epair_drain_rings(sca);
812 	epair_free_sc(sca);
813 
814 	/* Last free the cloner unit. */
815 	ifc_free_unit(ifc, unit);
816 
817 	return (0);
818 }
819 
820 static void
821 vnet_epair_init(const void *unused __unused)
822 {
823 	struct if_clone_addreq req = {
824 		.match_f = epair_clone_match,
825 		.create_f = epair_clone_create,
826 		.destroy_f = epair_clone_destroy,
827 	};
828 	V_epair_cloner = ifc_attach_cloner(epairname, &req);
829 }
830 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
831     vnet_epair_init, NULL);
832 
833 static void
834 vnet_epair_uninit(const void *unused __unused)
835 {
836 
837 	ifc_detach_cloner(V_epair_cloner);
838 }
839 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
840     vnet_epair_uninit, NULL);
841 
842 static int
843 epair_mod_init(void)
844 {
845 	char name[32];
846 	epair_tasks.tasks = 0;
847 
848 #ifdef RSS
849 	int cpu;
850 
851 	CPU_FOREACH(cpu) {
852 		cpuset_t cpu_mask;
853 
854 		/* Pin to this CPU so we get appropriate NUMA allocations. */
855 		thread_lock(curthread);
856 		sched_bind(curthread, cpu);
857 		thread_unlock(curthread);
858 
859 		snprintf(name, sizeof(name), "epair_task_%d", cpu);
860 
861 		epair_tasks.tq[cpu] = taskqueue_create(name, M_WAITOK,
862 		    taskqueue_thread_enqueue,
863 		    &epair_tasks.tq[cpu]);
864 		CPU_SETOF(cpu, &cpu_mask);
865 		taskqueue_start_threads_cpuset(&epair_tasks.tq[cpu], 1, PI_NET,
866 		    &cpu_mask, "%s", name);
867 
868 		epair_tasks.tasks++;
869 	}
870 	thread_lock(curthread);
871 	sched_unbind(curthread);
872 	thread_unlock(curthread);
873 #else
874 	snprintf(name, sizeof(name), "epair_task");
875 
876 	epair_tasks.tq[0] = taskqueue_create(name, M_WAITOK,
877 	    taskqueue_thread_enqueue,
878 	    &epair_tasks.tq[0]);
879 	taskqueue_start_threads(&epair_tasks.tq[0], 1, PI_NET, "%s", name);
880 
881 	epair_tasks.tasks = 1;
882 #endif
883 
884 	return (0);
885 }
886 
887 static void
888 epair_mod_cleanup(void)
889 {
890 
891 	for (int i = 0; i < epair_tasks.tasks; i++) {
892 		taskqueue_drain_all(epair_tasks.tq[i]);
893 		taskqueue_free(epair_tasks.tq[i]);
894 	}
895 }
896 
897 static int
898 epair_modevent(module_t mod, int type, void *data)
899 {
900 	int ret;
901 
902 	switch (type) {
903 	case MOD_LOAD:
904 		EPAIR_LOCK_INIT();
905 		ret = epair_mod_init();
906 		if (ret != 0)
907 			return (ret);
908 		if (bootverbose)
909 			printf("%s: %s initialized.\n", __func__, epairname);
910 		break;
911 	case MOD_UNLOAD:
912 		epair_mod_cleanup();
913 		EPAIR_LOCK_DESTROY();
914 		if (bootverbose)
915 			printf("%s: %s unloaded.\n", __func__, epairname);
916 		break;
917 	default:
918 		return (EOPNOTSUPP);
919 	}
920 	return (0);
921 }
922 
923 static moduledata_t epair_mod = {
924 	"if_epair",
925 	epair_modevent,
926 	0
927 };
928 
929 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
930 MODULE_VERSION(if_epair, 3);
931