xref: /freebsd/sys/net/if_epair.c (revision 1670a1c2a47d10ecccd001970b859caf93cd3b6e)
1 /*-
2  * Copyright (c) 2008 The FreeBSD Foundation
3  * Copyright (c) 2009-2010 Bjoern A. Zeeb <bz@FreeBSD.org>
4  * All rights reserved.
5  *
6  * This software was developed by CK Software GmbH under sponsorship
7  * from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  * notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  * notice, this list of conditions and the following disclaimer in the
16  * documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 /*
32  * A pair of virtual back-to-back connected ethernet like interfaces
33  * (``two interfaces with a virtual cross-over cable'').
34  *
35  * This is mostly intended to be used to provide connectivity between
36  * different virtual network stack instances.
37  */
38 /*
39  * Things to re-think once we have more experience:
40  * - ifp->if_reassign function once we can test with vimage. Depending on
41  *   how if_vmove() is going to be improved.
42  * - Real random etheraddrs that are checked to be uniquish; we would need
43  *   to re-do them in case we move the interface between network stacks
44  *   in a private if_reassign function.
45  *   In case we bridge to a real interface/network or between indepedent
46  *   epairs on multiple stacks/machines, we may need this.
47  *   For now let the user handle that case.
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 #include <sys/param.h>
54 #include <sys/kernel.h>
55 #include <sys/mbuf.h>
56 #include <sys/module.h>
57 #include <sys/refcount.h>
58 #include <sys/queue.h>
59 #include <sys/smp.h>
60 #include <sys/socket.h>
61 #include <sys/sockio.h>
62 #include <sys/sysctl.h>
63 #include <sys/types.h>
64 
65 #include <net/bpf.h>
66 #include <net/ethernet.h>
67 #include <net/if.h>
68 #include <net/if_clone.h>
69 #include <net/if_var.h>
70 #include <net/if_types.h>
71 #include <net/netisr.h>
72 #include <net/vnet.h>
73 
74 #define	EPAIRNAME	"epair"
75 
76 SYSCTL_DECL(_net_link);
77 SYSCTL_NODE(_net_link, OID_AUTO, epair, CTLFLAG_RW, 0, "epair sysctl");
78 
79 #ifdef EPAIR_DEBUG
80 static int epair_debug = 0;
81 SYSCTL_XINT(_net_link_epair, OID_AUTO, epair_debug, CTLFLAG_RW,
82     &epair_debug, 0, "if_epair(4) debugging.");
83 #define	DPRINTF(fmt, arg...)						\
84 	if (epair_debug)						\
85 		printf("[%s:%d] " fmt, __func__, __LINE__, ##arg)
86 #else
87 #define	DPRINTF(fmt, arg...)
88 #endif
89 
90 static void epair_nh_sintr(struct mbuf *);
91 static struct mbuf *epair_nh_m2cpuid(struct mbuf *, uintptr_t, u_int *);
92 static void epair_nh_drainedcpu(u_int);
93 
94 static void epair_start_locked(struct ifnet *);
95 
96 static int epair_clone_match(struct if_clone *, const char *);
97 static int epair_clone_create(struct if_clone *, char *, size_t, caddr_t);
98 static int epair_clone_destroy(struct if_clone *, struct ifnet *);
99 
100 /* Netisr realted definitions and sysctl. */
101 static struct netisr_handler epair_nh = {
102 	.nh_name	= EPAIRNAME,
103 	.nh_proto	= NETISR_EPAIR,
104 	.nh_policy	= NETISR_POLICY_CPU,
105 	.nh_handler	= epair_nh_sintr,
106 	.nh_m2cpuid	= epair_nh_m2cpuid,
107 	.nh_drainedcpu	= epair_nh_drainedcpu,
108 };
109 
110 static int
111 sysctl_epair_netisr_maxqlen(SYSCTL_HANDLER_ARGS)
112 {
113 	int error, qlimit;
114 
115 	netisr_getqlimit(&epair_nh, &qlimit);
116 	error = sysctl_handle_int(oidp, &qlimit, 0, req);
117 	if (error || !req->newptr)
118 		return (error);
119 	if (qlimit < 1)
120 		return (EINVAL);
121 	return (netisr_setqlimit(&epair_nh, qlimit));
122 }
123 SYSCTL_PROC(_net_link_epair, OID_AUTO, netisr_maxqlen, CTLTYPE_INT|CTLFLAG_RW,
124     0, 0, sysctl_epair_netisr_maxqlen, "I",
125     "Maximum if_epair(4) netisr \"hw\" queue length");
126 
127 struct epair_softc {
128 	struct ifnet	*ifp;		/* This ifp. */
129 	struct ifnet	*oifp;		/* other ifp of pair. */
130 	u_int		refcount;	/* # of mbufs in flight. */
131 	u_int		cpuid;		/* CPU ID assigned upon creation. */
132 	void		(*if_qflush)(struct ifnet *);
133 					/* Original if_qflush routine. */
134 };
135 
136 /*
137  * Per-CPU list of ifps with data in the ifq that needs to be flushed
138  * to the netisr ``hw'' queue before we allow any further direct queuing
139  * to the ``hw'' queue.
140  */
141 struct epair_ifp_drain {
142 	STAILQ_ENTRY(epair_ifp_drain)	ifp_next;
143 	struct ifnet			*ifp;
144 };
145 STAILQ_HEAD(eid_list, epair_ifp_drain);
146 
147 #define	EPAIR_LOCK_INIT(dpcpu)		mtx_init(&(dpcpu)->if_epair_mtx, \
148 					    "if_epair", NULL, MTX_DEF)
149 #define	EPAIR_LOCK_DESTROY(dpcpu)	mtx_destroy(&(dpcpu)->if_epair_mtx)
150 #define	EPAIR_LOCK_ASSERT(dpcpu)	mtx_assert(&(dpcpu)->if_epair_mtx, \
151 					    MA_OWNED)
152 #define	EPAIR_LOCK(dpcpu)		mtx_lock(&(dpcpu)->if_epair_mtx)
153 #define	EPAIR_UNLOCK(dpcpu)		mtx_unlock(&(dpcpu)->if_epair_mtx)
154 
155 #ifdef INVARIANTS
156 #define	EPAIR_REFCOUNT_INIT(r, v)	refcount_init((r), (v))
157 #define	EPAIR_REFCOUNT_AQUIRE(r)	refcount_acquire((r))
158 #define	EPAIR_REFCOUNT_RELEASE(r)	refcount_release((r))
159 #define	EPAIR_REFCOUNT_ASSERT(a, p)	KASSERT(a, p)
160 #else
161 #define	EPAIR_REFCOUNT_INIT(r, v)
162 #define	EPAIR_REFCOUNT_AQUIRE(r)
163 #define	EPAIR_REFCOUNT_RELEASE(r)
164 #define	EPAIR_REFCOUNT_ASSERT(a, p)
165 #endif
166 
167 static MALLOC_DEFINE(M_EPAIR, EPAIRNAME,
168     "Pair of virtual cross-over connected Ethernet-like interfaces");
169 
170 static struct if_clone epair_cloner = IFC_CLONE_INITIALIZER(
171     EPAIRNAME, NULL, IF_MAXUNIT,
172     NULL, epair_clone_match, epair_clone_create, epair_clone_destroy);
173 
174 /*
175  * DPCPU area and functions.
176  */
177 struct epair_dpcpu {
178 	struct mtx	if_epair_mtx;		/* Per-CPU locking. */
179 	int		epair_drv_flags;	/* Per-CPU ``hw'' drv flags. */
180 	struct eid_list	epair_ifp_drain_list;	/* Per-CPU list of ifps with
181 						 * data in the ifq. */
182 };
183 DPCPU_DEFINE(struct epair_dpcpu, epair_dpcpu);
184 
185 static void
186 epair_dpcpu_init(void)
187 {
188 	struct epair_dpcpu *epair_dpcpu;
189 	struct eid_list *s;
190 	u_int cpuid;
191 
192 	CPU_FOREACH(cpuid) {
193 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
194 
195 		/* Initialize per-cpu lock. */
196 		EPAIR_LOCK_INIT(epair_dpcpu);
197 
198 		/* Driver flags are per-cpu as are our netisr "hw" queues. */
199 		epair_dpcpu->epair_drv_flags = 0;
200 
201 		/*
202 		 * Initialize per-cpu drain list.
203 		 * Manually do what STAILQ_HEAD_INITIALIZER would do.
204 		 */
205 		s = &epair_dpcpu->epair_ifp_drain_list;
206 		s->stqh_first = NULL;
207 		s->stqh_last = &s->stqh_first;
208 	}
209 }
210 
211 static void
212 epair_dpcpu_detach(void)
213 {
214 	struct epair_dpcpu *epair_dpcpu;
215 	u_int cpuid;
216 
217 	CPU_FOREACH(cpuid) {
218 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
219 
220 		/* Destroy per-cpu lock. */
221 		EPAIR_LOCK_DESTROY(epair_dpcpu);
222 	}
223 }
224 
225 /*
226  * Helper functions.
227  */
228 static u_int
229 cpuid_from_ifp(struct ifnet *ifp)
230 {
231 	struct epair_softc *sc;
232 
233 	if (ifp == NULL)
234 		return (0);
235 	sc = ifp->if_softc;
236 
237 	return (sc->cpuid);
238 }
239 
240 /*
241  * Netisr handler functions.
242  */
243 static void
244 epair_nh_sintr(struct mbuf *m)
245 {
246 	struct ifnet *ifp;
247 	struct epair_softc *sc;
248 
249 	ifp = m->m_pkthdr.rcvif;
250 	(*ifp->if_input)(ifp, m);
251 	sc = ifp->if_softc;
252 	EPAIR_REFCOUNT_RELEASE(&sc->refcount);
253 	EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
254 	    ("%s: ifp=%p sc->refcount not >= 1: %d",
255 	    __func__, ifp, sc->refcount));
256 	DPRINTF("ifp=%p refcount=%u\n", ifp, sc->refcount);
257 }
258 
259 static struct mbuf *
260 epair_nh_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
261 {
262 
263 	*cpuid = cpuid_from_ifp(m->m_pkthdr.rcvif);
264 
265 	return (m);
266 }
267 
268 static void
269 epair_nh_drainedcpu(u_int cpuid)
270 {
271 	struct epair_dpcpu *epair_dpcpu;
272 	struct epair_ifp_drain *elm, *tvar;
273 	struct ifnet *ifp;
274 
275 	epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
276 	EPAIR_LOCK(epair_dpcpu);
277 	/*
278 	 * Assume our "hw" queue and possibly ifq will be emptied
279 	 * again. In case we will overflow the "hw" queue while
280 	 * draining, epair_start_locked will set IFF_DRV_OACTIVE
281 	 * again and we will stop and return.
282 	 */
283 	STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
284 	    ifp_next, tvar) {
285 		ifp = elm->ifp;
286 		epair_dpcpu->epair_drv_flags &= ~IFF_DRV_OACTIVE;
287 		ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
288 		epair_start_locked(ifp);
289 
290 		IFQ_LOCK(&ifp->if_snd);
291 		if (IFQ_IS_EMPTY(&ifp->if_snd)) {
292 			struct epair_softc *sc;
293 
294 			STAILQ_REMOVE(&epair_dpcpu->epair_ifp_drain_list,
295 			    elm, epair_ifp_drain, ifp_next);
296 			/* The cached ifp goes off the list. */
297 			sc = ifp->if_softc;
298 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
299 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
300 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
301 			    __func__, ifp, sc->refcount));
302 			free(elm, M_EPAIR);
303 		}
304 		IFQ_UNLOCK(&ifp->if_snd);
305 
306 		if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) != 0) {
307 			/* Our "hw"q overflew again. */
308 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE
309 			DPRINTF("hw queue length overflow at %u\n",
310 			    epair_nh.nh_qlimit);
311 			break;
312 		}
313 	}
314 	EPAIR_UNLOCK(epair_dpcpu);
315 }
316 
317 /*
318  * Network interface (`if') related functions.
319  */
320 static void
321 epair_remove_ifp_from_draining(struct ifnet *ifp)
322 {
323 	struct epair_dpcpu *epair_dpcpu;
324 	struct epair_ifp_drain *elm, *tvar;
325 	u_int cpuid;
326 
327 	CPU_FOREACH(cpuid) {
328 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
329 		EPAIR_LOCK(epair_dpcpu);
330 		STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
331 		    ifp_next, tvar) {
332 			if (ifp == elm->ifp) {
333 				struct epair_softc *sc;
334 
335 				STAILQ_REMOVE(
336 				    &epair_dpcpu->epair_ifp_drain_list, elm,
337 				    epair_ifp_drain, ifp_next);
338 				/* The cached ifp goes off the list. */
339 				sc = ifp->if_softc;
340 				EPAIR_REFCOUNT_RELEASE(&sc->refcount);
341 				EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
342 				    ("%s: ifp=%p sc->refcount not >= 1: %d",
343 				    __func__, ifp, sc->refcount));
344 				free(elm, M_EPAIR);
345 			}
346 		}
347 		EPAIR_UNLOCK(epair_dpcpu);
348 	}
349 }
350 
351 static int
352 epair_add_ifp_for_draining(struct ifnet *ifp)
353 {
354 	struct epair_dpcpu *epair_dpcpu;
355 	struct epair_softc *sc;
356 	struct epair_ifp_drain *elm = NULL;
357 
358 	sc = ifp->if_softc;
359 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
360 	EPAIR_LOCK_ASSERT(epair_dpcpu);
361 	STAILQ_FOREACH(elm, &epair_dpcpu->epair_ifp_drain_list, ifp_next)
362 		if (elm->ifp == ifp)
363 			break;
364 	/* If the ifp is there already, return success. */
365 	if (elm != NULL)
366 		return (0);
367 
368 	elm = malloc(sizeof(struct epair_ifp_drain), M_EPAIR, M_NOWAIT|M_ZERO);
369 	if (elm == NULL)
370 		return (ENOMEM);
371 
372 	elm->ifp = ifp;
373 	/* Add a reference for the ifp pointer on the list. */
374 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
375 	STAILQ_INSERT_TAIL(&epair_dpcpu->epair_ifp_drain_list, elm, ifp_next);
376 
377 	return (0);
378 }
379 
380 static void
381 epair_start_locked(struct ifnet *ifp)
382 {
383 	struct epair_dpcpu *epair_dpcpu;
384 	struct mbuf *m;
385 	struct epair_softc *sc;
386 	struct ifnet *oifp;
387 	int error;
388 
389 	DPRINTF("ifp=%p\n", ifp);
390 	sc = ifp->if_softc;
391 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
392 	EPAIR_LOCK_ASSERT(epair_dpcpu);
393 
394 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
395 		return;
396 	if ((ifp->if_flags & IFF_UP) == 0)
397 		return;
398 
399 	/*
400 	 * We get patckets here from ether_output via if_handoff()
401 	 * and ned to put them into the input queue of the oifp
402 	 * and call oifp->if_input() via netisr/epair_sintr().
403 	 */
404 	oifp = sc->oifp;
405 	sc = oifp->if_softc;
406 	for (;;) {
407 		IFQ_DEQUEUE(&ifp->if_snd, m);
408 		if (m == NULL)
409 			break;
410 		BPF_MTAP(ifp, m);
411 
412 		/*
413 		 * In case the outgoing interface is not usable,
414 		 * drop the packet.
415 		 */
416 		if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
417 		    (oifp->if_flags & IFF_UP) ==0) {
418 			ifp->if_oerrors++;
419 			m_freem(m);
420 			continue;
421 		}
422 		DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
423 
424 		/*
425 		 * Add a reference so the interface cannot go while the
426 		 * packet is in transit as we rely on rcvif to stay valid.
427 		 */
428 		EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
429 		m->m_pkthdr.rcvif = oifp;
430 		CURVNET_SET_QUIET(oifp->if_vnet);
431 		error = netisr_queue(NETISR_EPAIR, m);
432 		CURVNET_RESTORE();
433 		if (!error) {
434 			ifp->if_opackets++;
435 			/* Someone else received the packet. */
436 			oifp->if_ipackets++;
437 		} else {
438 			/* The packet was freed already. */
439 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
440 			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
441 			(void) epair_add_ifp_for_draining(ifp);
442 			ifp->if_oerrors++;
443 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
444 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
445 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
446 			    __func__, oifp, sc->refcount));
447 		}
448 	}
449 }
450 
451 static void
452 epair_start(struct ifnet *ifp)
453 {
454 	struct epair_dpcpu *epair_dpcpu;
455 
456 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
457 	EPAIR_LOCK(epair_dpcpu);
458 	epair_start_locked(ifp);
459 	EPAIR_UNLOCK(epair_dpcpu);
460 }
461 
462 static int
463 epair_transmit_locked(struct ifnet *ifp, struct mbuf *m)
464 {
465 	struct epair_dpcpu *epair_dpcpu;
466 	struct epair_softc *sc;
467 	struct ifnet *oifp;
468 	int error, len;
469 	short mflags;
470 
471 	DPRINTF("ifp=%p m=%p\n", ifp, m);
472 	sc = ifp->if_softc;
473 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
474 	EPAIR_LOCK_ASSERT(epair_dpcpu);
475 
476 	if (m == NULL)
477 		return (0);
478 
479 	/*
480 	 * We are not going to use the interface en/dequeue mechanism
481 	 * on the TX side. We are called from ether_output_frame()
482 	 * and will put the packet into the incoming queue of the
483 	 * other interface of our pair via the netsir.
484 	 */
485 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
486 		m_freem(m);
487 		return (ENXIO);
488 	}
489 	if ((ifp->if_flags & IFF_UP) == 0) {
490 		m_freem(m);
491 		return (ENETDOWN);
492 	}
493 
494 	BPF_MTAP(ifp, m);
495 
496 	/*
497 	 * In case the outgoing interface is not usable,
498 	 * drop the packet.
499 	 */
500 	oifp = sc->oifp;
501 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
502 	    (oifp->if_flags & IFF_UP) ==0) {
503 		ifp->if_oerrors++;
504 		m_freem(m);
505 		return (0);
506 	}
507 	len = m->m_pkthdr.len;
508 	mflags = m->m_flags;
509 	DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
510 
511 #ifdef ALTQ
512 	/* Support ALTQ via the clasic if_start() path. */
513 	IF_LOCK(&ifp->if_snd);
514 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
515 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
516 		if (error)
517 			ifp->if_snd.ifq_drops++;
518 		IF_UNLOCK(&ifp->if_snd);
519 		if (!error) {
520 			ifp->if_obytes += len;
521 			if (mflags & (M_BCAST|M_MCAST))
522 				ifp->if_omcasts++;
523 
524 			if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) == 0)
525 				epair_start_locked(ifp);
526 			else
527 				(void)epair_add_ifp_for_draining(ifp);
528 		}
529 		return (error);
530 	}
531 	IF_UNLOCK(&ifp->if_snd);
532 #endif
533 
534 	if ((epair_dpcpu->epair_drv_flags & IFF_DRV_OACTIVE) != 0) {
535 		/*
536 		 * Our hardware queue is full, try to fall back
537 		 * queuing to the ifq but do not call ifp->if_start.
538 		 * Either we are lucky or the packet is gone.
539 		 */
540 		IFQ_ENQUEUE(&ifp->if_snd, m, error);
541 		if (!error)
542 			(void)epair_add_ifp_for_draining(ifp);
543 		return (error);
544 	}
545 	sc = oifp->if_softc;
546 	/*
547 	 * Add a reference so the interface cannot go while the
548 	 * packet is in transit as we rely on rcvif to stay valid.
549 	 */
550 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
551 	m->m_pkthdr.rcvif = oifp;
552 	CURVNET_SET_QUIET(oifp->if_vnet);
553 	error = netisr_queue(NETISR_EPAIR, m);
554 	CURVNET_RESTORE();
555 	if (!error) {
556 		ifp->if_opackets++;
557 		/*
558 		 * IFQ_HANDOFF_ADJ/ip_handoff() update statistics,
559 		 * but as we bypass all this we have to duplicate
560 		 * the logic another time.
561 		 */
562 		ifp->if_obytes += len;
563 		if (mflags & (M_BCAST|M_MCAST))
564 			ifp->if_omcasts++;
565 		/* Someone else received the packet. */
566 		oifp->if_ipackets++;
567 	} else {
568 		/* The packet was freed already. */
569 		epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
570 		ifp->if_drv_flags |= IFF_DRV_OACTIVE;
571 		ifp->if_oerrors++;
572 		EPAIR_REFCOUNT_RELEASE(&sc->refcount);
573 		EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
574 		    ("%s: ifp=%p sc->refcount not >= 1: %d",
575 		    __func__, oifp, sc->refcount));
576 	}
577 
578 	return (error);
579 }
580 
581 static int
582 epair_transmit(struct ifnet *ifp, struct mbuf *m)
583 {
584 	struct epair_dpcpu *epair_dpcpu;
585 	int error;
586 
587 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
588 	EPAIR_LOCK(epair_dpcpu);
589 	error = epair_transmit_locked(ifp, m);
590 	EPAIR_UNLOCK(epair_dpcpu);
591 	return (error);
592 }
593 
594 static void
595 epair_qflush(struct ifnet *ifp)
596 {
597 	struct epair_softc *sc;
598 
599 	sc = ifp->if_softc;
600 	KASSERT(sc != NULL, ("%s: ifp=%p, epair_softc gone? sc=%p\n",
601 	    __func__, ifp, sc));
602 	/*
603 	 * Remove this ifp from all backpointer lists. The interface will not
604 	 * usable for flushing anyway nor should it have anything to flush
605 	 * after if_qflush().
606 	 */
607 	epair_remove_ifp_from_draining(ifp);
608 
609 	if (sc->if_qflush)
610 		sc->if_qflush(ifp);
611 }
612 
613 static int
614 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
615 {
616 	struct ifreq *ifr;
617 	int error;
618 
619 	ifr = (struct ifreq *)data;
620 	switch (cmd) {
621 	case SIOCSIFFLAGS:
622 	case SIOCADDMULTI:
623 	case SIOCDELMULTI:
624 		error = 0;
625 		break;
626 
627 	case SIOCSIFMTU:
628 		/* We basically allow all kinds of MTUs. */
629 		ifp->if_mtu = ifr->ifr_mtu;
630 		error = 0;
631 		break;
632 
633 	default:
634 		/* Let the common ethernet handler process this. */
635 		error = ether_ioctl(ifp, cmd, data);
636 		break;
637 	}
638 
639 	return (error);
640 }
641 
642 static void
643 epair_init(void *dummy __unused)
644 {
645 }
646 
647 
648 /*
649  * Interface cloning functions.
650  * We use our private ones so that we can create/destroy our secondary
651  * device along with the primary one.
652  */
653 static int
654 epair_clone_match(struct if_clone *ifc, const char *name)
655 {
656 	const char *cp;
657 
658 	DPRINTF("name='%s'\n", name);
659 
660 	/*
661 	 * Our base name is epair.
662 	 * Our interfaces will be named epair<n>[ab].
663 	 * So accept anything of the following list:
664 	 * - epair
665 	 * - epair<n>
666 	 * but not the epair<n>[ab] versions.
667 	 */
668 	if (strncmp(EPAIRNAME, name, sizeof(EPAIRNAME)-1) != 0)
669 		return (0);
670 
671 	for (cp = name + sizeof(EPAIRNAME) - 1; *cp != '\0'; cp++) {
672 		if (*cp < '0' || *cp > '9')
673 			return (0);
674 	}
675 
676 	return (1);
677 }
678 
679 static int
680 epair_clone_create(struct if_clone *ifc, char *name, size_t len, caddr_t params)
681 {
682 	struct epair_softc *sca, *scb;
683 	struct ifnet *ifp;
684 	char *dp;
685 	int error, unit, wildcard;
686 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
687 
688 	/*
689 	 * We are abusing params to create our second interface.
690 	 * Actually we already created it and called if_clone_createif()
691 	 * for it to do the official insertion procedure the moment we knew
692 	 * it cannot fail anymore. So just do attach it here.
693 	 */
694 	if (params) {
695 		scb = (struct epair_softc *)params;
696 		ifp = scb->ifp;
697 		/* Assign a hopefully unique, locally administered etheraddr. */
698 		eaddr[0] = 0x02;
699 		eaddr[3] = (ifp->if_index >> 8) & 0xff;
700 		eaddr[4] = ifp->if_index & 0xff;
701 		eaddr[5] = 0x0b;
702 		ether_ifattach(ifp, eaddr);
703 		/* Correctly set the name for the cloner list. */
704 		strlcpy(name, scb->ifp->if_xname, len);
705 		return (0);
706 	}
707 
708 	/* Try to see if a special unit was requested. */
709 	error = ifc_name2unit(name, &unit);
710 	if (error != 0)
711 		return (error);
712 	wildcard = (unit < 0);
713 
714 	error = ifc_alloc_unit(ifc, &unit);
715 	if (error != 0)
716 		return (error);
717 
718 	/*
719 	 * If no unit had been given, we need to adjust the ifName.
720 	 * Also make sure there is space for our extra [ab] suffix.
721 	 */
722 	for (dp = name; *dp != '\0'; dp++);
723 	if (wildcard) {
724 		error = snprintf(dp, len - (dp - name), "%d", unit);
725 		if (error > len - (dp - name) - 1) {
726 			/* ifName too long. */
727 			ifc_free_unit(ifc, unit);
728 			return (ENOSPC);
729 		}
730 		dp += error;
731 	}
732 	if (len - (dp - name) - 1 < 1) {
733 		/* No space left for our [ab] suffix. */
734 		ifc_free_unit(ifc, unit);
735 		return (ENOSPC);
736 	}
737 	*dp = 'a';
738 	/* Must not change dp so we can replace 'a' by 'b' later. */
739 	*(dp+1) = '\0';
740 
741 	/* Allocate memory for both [ab] interfaces */
742 	sca = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
743 	EPAIR_REFCOUNT_INIT(&sca->refcount, 1);
744 	sca->ifp = if_alloc(IFT_ETHER);
745 	if (sca->ifp == NULL) {
746 		free(sca, M_EPAIR);
747 		ifc_free_unit(ifc, unit);
748 		return (ENOSPC);
749 	}
750 
751 	scb = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
752 	EPAIR_REFCOUNT_INIT(&scb->refcount, 1);
753 	scb->ifp = if_alloc(IFT_ETHER);
754 	if (scb->ifp == NULL) {
755 		free(scb, M_EPAIR);
756 		if_free(sca->ifp);
757 		free(sca, M_EPAIR);
758 		ifc_free_unit(ifc, unit);
759 		return (ENOSPC);
760 	}
761 
762 	/*
763 	 * Cross-reference the interfaces so we will be able to free both.
764 	 */
765 	sca->oifp = scb->ifp;
766 	scb->oifp = sca->ifp;
767 
768 	/*
769 	 * Calculate the cpuid for netisr queueing based on the
770 	 * ifIndex of the interfaces. As long as we cannot configure
771 	 * this or use cpuset information easily we cannot guarantee
772 	 * cache locality but we can at least allow parallelism.
773 	 */
774 	sca->cpuid =
775 	    netisr_get_cpuid(sca->ifp->if_index % netisr_get_cpucount());
776 	scb->cpuid =
777 	    netisr_get_cpuid(scb->ifp->if_index % netisr_get_cpucount());
778 
779 	/* Finish initialization of interface <n>a. */
780 	ifp = sca->ifp;
781 	ifp->if_softc = sca;
782 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
783 	ifp->if_dname = ifc->ifc_name;
784 	ifp->if_dunit = unit;
785 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
786 	ifp->if_start = epair_start;
787 	ifp->if_ioctl = epair_ioctl;
788 	ifp->if_init  = epair_init;
789 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
790 	/* Assign a hopefully unique, locally administered etheraddr. */
791 	eaddr[0] = 0x02;
792 	eaddr[3] = (ifp->if_index >> 8) & 0xff;
793 	eaddr[4] = ifp->if_index & 0xff;
794 	eaddr[5] = 0x0a;
795 	ether_ifattach(ifp, eaddr);
796 	sca->if_qflush = ifp->if_qflush;
797 	ifp->if_qflush = epair_qflush;
798 	ifp->if_transmit = epair_transmit;
799 	ifp->if_baudrate = IF_Gbps(10UL);	/* arbitrary maximum */
800 
801 	/* Swap the name and finish initialization of interface <n>b. */
802 	*dp = 'b';
803 
804 	ifp = scb->ifp;
805 	ifp->if_softc = scb;
806 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
807 	ifp->if_dname = ifc->ifc_name;
808 	ifp->if_dunit = unit;
809 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
810 	ifp->if_start = epair_start;
811 	ifp->if_ioctl = epair_ioctl;
812 	ifp->if_init  = epair_init;
813 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
814 	/* We need to play some tricks here for the second interface. */
815 	strlcpy(name, EPAIRNAME, len);
816 	error = if_clone_create(name, len, (caddr_t)scb);
817 	if (error)
818 		panic("%s: if_clone_createif() for our 2nd iface failed: %d",
819 		    __func__, error);
820 	scb->if_qflush = ifp->if_qflush;
821 	ifp->if_qflush = epair_qflush;
822 	ifp->if_transmit = epair_transmit;
823 	ifp->if_baudrate = IF_Gbps(10UL);	/* arbitrary maximum */
824 
825 	/*
826 	 * Restore name to <n>a as the ifp for this will go into the
827 	 * cloner list for the initial call.
828 	 */
829 	strlcpy(name, sca->ifp->if_xname, len);
830 	DPRINTF("name='%s/%db' created sca=%p scb=%p\n", name, unit, sca, scb);
831 
832 	/* Tell the world, that we are ready to rock. */
833 	sca->ifp->if_drv_flags |= IFF_DRV_RUNNING;
834 	scb->ifp->if_drv_flags |= IFF_DRV_RUNNING;
835 
836 	return (0);
837 }
838 
839 static int
840 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp)
841 {
842 	struct ifnet *oifp;
843 	struct epair_softc *sca, *scb;
844 	int unit, error;
845 
846 	DPRINTF("ifp=%p\n", ifp);
847 
848 	/*
849 	 * In case we called into if_clone_destroyif() ourselves
850 	 * again to remove the second interface, the softc will be
851 	 * NULL. In that case so not do anything but return success.
852 	 */
853 	if (ifp->if_softc == NULL)
854 		return (0);
855 
856 	unit = ifp->if_dunit;
857 	sca = ifp->if_softc;
858 	oifp = sca->oifp;
859 	scb = oifp->if_softc;
860 
861 	DPRINTF("ifp=%p oifp=%p\n", ifp, oifp);
862 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
863 	oifp->if_drv_flags &= ~IFF_DRV_RUNNING;
864 	ether_ifdetach(oifp);
865 	ether_ifdetach(ifp);
866 	/*
867 	 * Wait for all packets to be dispatched to if_input.
868 	 * The numbers can only go down as the interfaces are
869 	 * detached so there is no need to use atomics.
870 	 */
871 	DPRINTF("sca refcnt=%u scb refcnt=%u\n", sca->refcount, scb->refcount);
872 	EPAIR_REFCOUNT_ASSERT(sca->refcount == 1 && scb->refcount == 1,
873 	    ("%s: ifp=%p sca->refcount!=1: %d || ifp=%p scb->refcount!=1: %d",
874 	    __func__, ifp, sca->refcount, oifp, scb->refcount));
875 
876 	/*
877 	 * Get rid of our second half.
878 	 */
879 	oifp->if_softc = NULL;
880 	error = if_clone_destroyif(ifc, oifp);
881 	if (error)
882 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
883 		    __func__, error);
884 
885 	/*
886 	 * Finish cleaning up. Free them and release the unit.
887 	 * As the other of the two interfaces my reside in a different vnet,
888 	 * we need to switch before freeing them.
889 	 */
890 	CURVNET_SET_QUIET(oifp->if_vnet);
891 	if_free_type(oifp, IFT_ETHER);
892 	CURVNET_RESTORE();
893 	if_free_type(ifp, IFT_ETHER);
894 	free(scb, M_EPAIR);
895 	free(sca, M_EPAIR);
896 	ifc_free_unit(ifc, unit);
897 
898 	return (0);
899 }
900 
901 static int
902 epair_modevent(module_t mod, int type, void *data)
903 {
904 	int qlimit;
905 
906 	switch (type) {
907 	case MOD_LOAD:
908 		/* For now limit us to one global mutex and one inq. */
909 		epair_dpcpu_init();
910 		epair_nh.nh_qlimit = 42 * ifqmaxlen; /* 42 shall be the number. */
911 		if (TUNABLE_INT_FETCH("net.link.epair.netisr_maxqlen", &qlimit))
912 		    epair_nh.nh_qlimit = qlimit;
913 		netisr_register(&epair_nh);
914 		if_clone_attach(&epair_cloner);
915 		if (bootverbose)
916 			printf("%s initialized.\n", EPAIRNAME);
917 		break;
918 	case MOD_UNLOAD:
919 		if_clone_detach(&epair_cloner);
920 		netisr_unregister(&epair_nh);
921 		epair_dpcpu_detach();
922 		if (bootverbose)
923 			printf("%s unloaded.\n", EPAIRNAME);
924 		break;
925 	default:
926 		return (EOPNOTSUPP);
927 	}
928 	return (0);
929 }
930 
931 static moduledata_t epair_mod = {
932 	"if_epair",
933 	epair_modevent,
934 	0
935 };
936 
937 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);
938 MODULE_VERSION(if_epair, 1);
939