xref: /freebsd/sys/net/if_epair.c (revision f4f33ea0c752ff0f9bfad34991d5bbb54e71133d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008 The FreeBSD Foundation
5  * Copyright (c) 2009-2010 Bjoern A. Zeeb <bz@FreeBSD.org>
6  * All rights reserved.
7  *
8  * This software was developed by CK Software GmbH under sponsorship
9  * from the FreeBSD Foundation.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  * notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  * notice, this list of conditions and the following disclaimer in the
18  * documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 /*
34  * A pair of virtual back-to-back connected ethernet like interfaces
35  * (``two interfaces with a virtual cross-over cable'').
36  *
37  * This is mostly intended to be used to provide connectivity between
38  * different virtual network stack instances.
39  */
40 /*
41  * Things to re-think once we have more experience:
42  * - ifp->if_reassign function once we can test with vimage. Depending on
43  *   how if_vmove() is going to be improved.
44  * - Real random etheraddrs that are checked to be uniquish; we would need
45  *   to re-do them in case we move the interface between network stacks
46  *   in a private if_reassign function.
47  *   In case we bridge to a real interface/network or between indepedent
48  *   epairs on multiple stacks/machines, we may need this.
49  *   For now let the user handle that case.
50  */
51 
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
54 
55 #include <sys/param.h>
56 #include <sys/hash.h>
57 #include <sys/jail.h>
58 #include <sys/kernel.h>
59 #include <sys/libkern.h>
60 #include <sys/malloc.h>
61 #include <sys/mbuf.h>
62 #include <sys/module.h>
63 #include <sys/proc.h>
64 #include <sys/refcount.h>
65 #include <sys/queue.h>
66 #include <sys/smp.h>
67 #include <sys/socket.h>
68 #include <sys/sockio.h>
69 #include <sys/sysctl.h>
70 #include <sys/types.h>
71 
72 #include <net/bpf.h>
73 #include <net/ethernet.h>
74 #include <net/if.h>
75 #include <net/if_var.h>
76 #include <net/if_clone.h>
77 #include <net/if_media.h>
78 #include <net/if_var.h>
79 #include <net/if_types.h>
80 #include <net/netisr.h>
81 #include <net/vnet.h>
82 
83 SYSCTL_DECL(_net_link);
84 static SYSCTL_NODE(_net_link, OID_AUTO, epair, CTLFLAG_RW, 0, "epair sysctl");
85 
86 #ifdef EPAIR_DEBUG
87 static int epair_debug = 0;
88 SYSCTL_INT(_net_link_epair, OID_AUTO, epair_debug, CTLFLAG_RW,
89     &epair_debug, 0, "if_epair(4) debugging.");
90 #define	DPRINTF(fmt, arg...)						\
91 	if (epair_debug)						\
92 		printf("[%s:%d] " fmt, __func__, __LINE__, ##arg)
93 #else
94 #define	DPRINTF(fmt, arg...)
95 #endif
96 
97 static void epair_nh_sintr(struct mbuf *);
98 static struct mbuf *epair_nh_m2cpuid(struct mbuf *, uintptr_t, u_int *);
99 static void epair_nh_drainedcpu(u_int);
100 
101 static void epair_start_locked(struct ifnet *);
102 static int epair_media_change(struct ifnet *);
103 static void epair_media_status(struct ifnet *, struct ifmediareq *);
104 
105 static int epair_clone_match(struct if_clone *, const char *);
106 static int epair_clone_create(struct if_clone *, char *, size_t, caddr_t);
107 static int epair_clone_destroy(struct if_clone *, struct ifnet *);
108 
109 static const char epairname[] = "epair";
110 static unsigned int next_index = 0;
111 
112 /* Netisr related definitions and sysctl. */
113 static struct netisr_handler epair_nh = {
114 	.nh_name	= epairname,
115 	.nh_proto	= NETISR_EPAIR,
116 	.nh_policy	= NETISR_POLICY_CPU,
117 	.nh_handler	= epair_nh_sintr,
118 	.nh_m2cpuid	= epair_nh_m2cpuid,
119 	.nh_drainedcpu	= epair_nh_drainedcpu,
120 };
121 
122 static int
123 sysctl_epair_netisr_maxqlen(SYSCTL_HANDLER_ARGS)
124 {
125 	int error, qlimit;
126 
127 	netisr_getqlimit(&epair_nh, &qlimit);
128 	error = sysctl_handle_int(oidp, &qlimit, 0, req);
129 	if (error || !req->newptr)
130 		return (error);
131 	if (qlimit < 1)
132 		return (EINVAL);
133 	return (netisr_setqlimit(&epair_nh, qlimit));
134 }
135 SYSCTL_PROC(_net_link_epair, OID_AUTO, netisr_maxqlen, CTLTYPE_INT|CTLFLAG_RW,
136     0, 0, sysctl_epair_netisr_maxqlen, "I",
137     "Maximum if_epair(4) netisr \"hw\" queue length");
138 
139 struct epair_softc {
140 	struct ifnet	*ifp;		/* This ifp. */
141 	struct ifnet	*oifp;		/* other ifp of pair. */
142 	struct ifmedia	media;		/* Media config (fake). */
143 	u_int		refcount;	/* # of mbufs in flight. */
144 	u_int		cpuid;		/* CPU ID assigned upon creation. */
145 	void		(*if_qflush)(struct ifnet *);
146 					/* Original if_qflush routine. */
147 };
148 
149 /*
150  * Per-CPU list of ifps with data in the ifq that needs to be flushed
151  * to the netisr ``hw'' queue before we allow any further direct queuing
152  * to the ``hw'' queue.
153  */
154 struct epair_ifp_drain {
155 	STAILQ_ENTRY(epair_ifp_drain)	ifp_next;
156 	struct ifnet			*ifp;
157 };
158 STAILQ_HEAD(eid_list, epair_ifp_drain);
159 
160 #define	EPAIR_LOCK_INIT(dpcpu)		mtx_init(&(dpcpu)->if_epair_mtx, \
161 					    "if_epair", NULL, MTX_DEF)
162 #define	EPAIR_LOCK_DESTROY(dpcpu)	mtx_destroy(&(dpcpu)->if_epair_mtx)
163 #define	EPAIR_LOCK_ASSERT(dpcpu)	mtx_assert(&(dpcpu)->if_epair_mtx, \
164 					    MA_OWNED)
165 #define	EPAIR_LOCK(dpcpu)		mtx_lock(&(dpcpu)->if_epair_mtx)
166 #define	EPAIR_UNLOCK(dpcpu)		mtx_unlock(&(dpcpu)->if_epair_mtx)
167 
168 #ifdef INVARIANTS
169 #define	EPAIR_REFCOUNT_INIT(r, v)	refcount_init((r), (v))
170 #define	EPAIR_REFCOUNT_AQUIRE(r)	refcount_acquire((r))
171 #define	EPAIR_REFCOUNT_RELEASE(r)	refcount_release((r))
172 #define	EPAIR_REFCOUNT_ASSERT(a, p)	KASSERT(a, p)
173 #else
174 #define	EPAIR_REFCOUNT_INIT(r, v)
175 #define	EPAIR_REFCOUNT_AQUIRE(r)
176 #define	EPAIR_REFCOUNT_RELEASE(r)
177 #define	EPAIR_REFCOUNT_ASSERT(a, p)
178 #endif
179 
180 static MALLOC_DEFINE(M_EPAIR, epairname,
181     "Pair of virtual cross-over connected Ethernet-like interfaces");
182 
183 VNET_DEFINE_STATIC(struct if_clone *, epair_cloner);
184 #define	V_epair_cloner	VNET(epair_cloner)
185 
186 /*
187  * DPCPU area and functions.
188  */
189 struct epair_dpcpu {
190 	struct mtx	if_epair_mtx;		/* Per-CPU locking. */
191 	int		epair_drv_flags;	/* Per-CPU ``hw'' drv flags. */
192 	struct eid_list	epair_ifp_drain_list;	/* Per-CPU list of ifps with
193 						 * data in the ifq. */
194 };
195 DPCPU_DEFINE(struct epair_dpcpu, epair_dpcpu);
196 
197 static void
198 epair_dpcpu_init(void)
199 {
200 	struct epair_dpcpu *epair_dpcpu;
201 	struct eid_list *s;
202 	u_int cpuid;
203 
204 	CPU_FOREACH(cpuid) {
205 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
206 
207 		/* Initialize per-cpu lock. */
208 		EPAIR_LOCK_INIT(epair_dpcpu);
209 
210 		/* Driver flags are per-cpu as are our netisr "hw" queues. */
211 		epair_dpcpu->epair_drv_flags = 0;
212 
213 		/*
214 		 * Initialize per-cpu drain list.
215 		 * Manually do what STAILQ_HEAD_INITIALIZER would do.
216 		 */
217 		s = &epair_dpcpu->epair_ifp_drain_list;
218 		s->stqh_first = NULL;
219 		s->stqh_last = &s->stqh_first;
220 	}
221 }
222 
223 static void
224 epair_dpcpu_detach(void)
225 {
226 	struct epair_dpcpu *epair_dpcpu;
227 	u_int cpuid;
228 
229 	CPU_FOREACH(cpuid) {
230 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
231 
232 		/* Destroy per-cpu lock. */
233 		EPAIR_LOCK_DESTROY(epair_dpcpu);
234 	}
235 }
236 
237 /*
238  * Helper functions.
239  */
240 static u_int
241 cpuid_from_ifp(struct ifnet *ifp)
242 {
243 	struct epair_softc *sc;
244 
245 	if (ifp == NULL)
246 		return (0);
247 	sc = ifp->if_softc;
248 
249 	return (sc->cpuid);
250 }
251 
252 /*
253  * Netisr handler functions.
254  */
255 static void
256 epair_nh_sintr(struct mbuf *m)
257 {
258 	struct ifnet *ifp;
259 	struct epair_softc *sc __unused;
260 
261 	ifp = m->m_pkthdr.rcvif;
262 	(*ifp->if_input)(ifp, m);
263 	sc = ifp->if_softc;
264 	EPAIR_REFCOUNT_RELEASE(&sc->refcount);
265 	EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
266 	    ("%s: ifp=%p sc->refcount not >= 1: %d",
267 	    __func__, ifp, sc->refcount));
268 	DPRINTF("ifp=%p refcount=%u\n", ifp, sc->refcount);
269 }
270 
271 static struct mbuf *
272 epair_nh_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
273 {
274 
275 	*cpuid = cpuid_from_ifp(m->m_pkthdr.rcvif);
276 
277 	return (m);
278 }
279 
280 static void
281 epair_nh_drainedcpu(u_int cpuid)
282 {
283 	struct epair_dpcpu *epair_dpcpu;
284 	struct epair_ifp_drain *elm, *tvar;
285 	struct ifnet *ifp;
286 
287 	epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
288 	EPAIR_LOCK(epair_dpcpu);
289 	/*
290 	 * Assume our "hw" queue and possibly ifq will be emptied
291 	 * again. In case we will overflow the "hw" queue while
292 	 * draining, epair_start_locked will set IFF_DRV_OACTIVE
293 	 * again and we will stop and return.
294 	 */
295 	STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
296 	    ifp_next, tvar) {
297 		ifp = elm->ifp;
298 		epair_dpcpu->epair_drv_flags &= ~IFF_DRV_OACTIVE;
299 		ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
300 		epair_start_locked(ifp);
301 
302 		IFQ_LOCK(&ifp->if_snd);
303 		if (IFQ_IS_EMPTY(&ifp->if_snd)) {
304 			struct epair_softc *sc __unused;
305 
306 			STAILQ_REMOVE(&epair_dpcpu->epair_ifp_drain_list,
307 			    elm, epair_ifp_drain, ifp_next);
308 			/* The cached ifp goes off the list. */
309 			sc = ifp->if_softc;
310 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
311 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
312 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
313 			    __func__, ifp, sc->refcount));
314 			free(elm, M_EPAIR);
315 		}
316 		IFQ_UNLOCK(&ifp->if_snd);
317 
318 		if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) != 0) {
319 			/* Our "hw"q overflew again. */
320 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
321 			DPRINTF("hw queue length overflow at %u\n",
322 			    epair_nh.nh_qlimit);
323 			break;
324 		}
325 	}
326 	EPAIR_UNLOCK(epair_dpcpu);
327 }
328 
329 /*
330  * Network interface (`if') related functions.
331  */
332 static void
333 epair_remove_ifp_from_draining(struct ifnet *ifp)
334 {
335 	struct epair_dpcpu *epair_dpcpu;
336 	struct epair_ifp_drain *elm, *tvar;
337 	u_int cpuid;
338 
339 	CPU_FOREACH(cpuid) {
340 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
341 		EPAIR_LOCK(epair_dpcpu);
342 		STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
343 		    ifp_next, tvar) {
344 			if (ifp == elm->ifp) {
345 				struct epair_softc *sc __unused;
346 
347 				STAILQ_REMOVE(
348 				    &epair_dpcpu->epair_ifp_drain_list, elm,
349 				    epair_ifp_drain, ifp_next);
350 				/* The cached ifp goes off the list. */
351 				sc = ifp->if_softc;
352 				EPAIR_REFCOUNT_RELEASE(&sc->refcount);
353 				EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
354 				    ("%s: ifp=%p sc->refcount not >= 1: %d",
355 				    __func__, ifp, sc->refcount));
356 				free(elm, M_EPAIR);
357 			}
358 		}
359 		EPAIR_UNLOCK(epair_dpcpu);
360 	}
361 }
362 
363 static int
364 epair_add_ifp_for_draining(struct ifnet *ifp)
365 {
366 	struct epair_dpcpu *epair_dpcpu;
367 	struct epair_softc *sc;
368 	struct epair_ifp_drain *elm = NULL;
369 
370 	sc = ifp->if_softc;
371 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
372 	EPAIR_LOCK_ASSERT(epair_dpcpu);
373 	STAILQ_FOREACH(elm, &epair_dpcpu->epair_ifp_drain_list, ifp_next)
374 		if (elm->ifp == ifp)
375 			break;
376 	/* If the ifp is there already, return success. */
377 	if (elm != NULL)
378 		return (0);
379 
380 	elm = malloc(sizeof(struct epair_ifp_drain), M_EPAIR, M_NOWAIT|M_ZERO);
381 	if (elm == NULL)
382 		return (ENOMEM);
383 
384 	elm->ifp = ifp;
385 	/* Add a reference for the ifp pointer on the list. */
386 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
387 	STAILQ_INSERT_TAIL(&epair_dpcpu->epair_ifp_drain_list, elm, ifp_next);
388 
389 	return (0);
390 }
391 
392 static void
393 epair_start_locked(struct ifnet *ifp)
394 {
395 	struct epair_dpcpu *epair_dpcpu;
396 	struct mbuf *m;
397 	struct epair_softc *sc;
398 	struct ifnet *oifp;
399 	int error;
400 
401 	DPRINTF("ifp=%p\n", ifp);
402 	sc = ifp->if_softc;
403 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
404 	EPAIR_LOCK_ASSERT(epair_dpcpu);
405 
406 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
407 		return;
408 	if ((ifp->if_flags & IFF_UP) == 0)
409 		return;
410 
411 	/*
412 	 * We get packets here from ether_output via if_handoff()
413 	 * and need to put them into the input queue of the oifp
414 	 * and call oifp->if_input() via netisr/epair_sintr().
415 	 */
416 	oifp = sc->oifp;
417 	sc = oifp->if_softc;
418 	for (;;) {
419 		IFQ_DEQUEUE(&ifp->if_snd, m);
420 		if (m == NULL)
421 			break;
422 		BPF_MTAP(ifp, m);
423 
424 		/*
425 		 * In case the outgoing interface is not usable,
426 		 * drop the packet.
427 		 */
428 		if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
429 		    (oifp->if_flags & IFF_UP) ==0) {
430 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
431 			m_freem(m);
432 			continue;
433 		}
434 		DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
435 
436 		/*
437 		 * Add a reference so the interface cannot go while the
438 		 * packet is in transit as we rely on rcvif to stay valid.
439 		 */
440 		EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
441 		m->m_pkthdr.rcvif = oifp;
442 		CURVNET_SET_QUIET(oifp->if_vnet);
443 		error = netisr_queue(NETISR_EPAIR, m);
444 		CURVNET_RESTORE();
445 		if (!error) {
446 			if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
447 			/* Someone else received the packet. */
448 			if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
449 		} else {
450 			/* The packet was freed already. */
451 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
452 			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
453 			(void) epair_add_ifp_for_draining(ifp);
454 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
455 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
456 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
457 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
458 			    __func__, oifp, sc->refcount));
459 		}
460 	}
461 }
462 
463 static void
464 epair_start(struct ifnet *ifp)
465 {
466 	struct epair_dpcpu *epair_dpcpu;
467 
468 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
469 	EPAIR_LOCK(epair_dpcpu);
470 	epair_start_locked(ifp);
471 	EPAIR_UNLOCK(epair_dpcpu);
472 }
473 
474 static int
475 epair_transmit_locked(struct ifnet *ifp, struct mbuf *m)
476 {
477 	struct epair_dpcpu *epair_dpcpu;
478 	struct epair_softc *sc;
479 	struct ifnet *oifp;
480 	int error, len;
481 	short mflags;
482 
483 	DPRINTF("ifp=%p m=%p\n", ifp, m);
484 	sc = ifp->if_softc;
485 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
486 	EPAIR_LOCK_ASSERT(epair_dpcpu);
487 
488 	if (m == NULL)
489 		return (0);
490 
491 	/*
492 	 * We are not going to use the interface en/dequeue mechanism
493 	 * on the TX side. We are called from ether_output_frame()
494 	 * and will put the packet into the incoming queue of the
495 	 * other interface of our pair via the netsir.
496 	 */
497 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
498 		m_freem(m);
499 		return (ENXIO);
500 	}
501 	if ((ifp->if_flags & IFF_UP) == 0) {
502 		m_freem(m);
503 		return (ENETDOWN);
504 	}
505 
506 	BPF_MTAP(ifp, m);
507 
508 	/*
509 	 * In case the outgoing interface is not usable,
510 	 * drop the packet.
511 	 */
512 	oifp = sc->oifp;
513 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
514 	    (oifp->if_flags & IFF_UP) ==0) {
515 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
516 		m_freem(m);
517 		return (0);
518 	}
519 	len = m->m_pkthdr.len;
520 	mflags = m->m_flags;
521 	DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
522 
523 #ifdef ALTQ
524 	/* Support ALTQ via the classic if_start() path. */
525 	IF_LOCK(&ifp->if_snd);
526 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
527 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
528 		if (error)
529 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
530 		IF_UNLOCK(&ifp->if_snd);
531 		if (!error) {
532 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
533 			if (mflags & (M_BCAST|M_MCAST))
534 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
535 
536 			if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) == 0)
537 				epair_start_locked(ifp);
538 			else
539 				(void)epair_add_ifp_for_draining(ifp);
540 		}
541 		return (error);
542 	}
543 	IF_UNLOCK(&ifp->if_snd);
544 #endif
545 
546 	if ((epair_dpcpu->epair_drv_flags & IFF_DRV_OACTIVE) != 0) {
547 		/*
548 		 * Our hardware queue is full, try to fall back
549 		 * queuing to the ifq but do not call ifp->if_start.
550 		 * Either we are lucky or the packet is gone.
551 		 */
552 		IFQ_ENQUEUE(&ifp->if_snd, m, error);
553 		if (!error)
554 			(void)epair_add_ifp_for_draining(ifp);
555 		return (error);
556 	}
557 	sc = oifp->if_softc;
558 	/*
559 	 * Add a reference so the interface cannot go while the
560 	 * packet is in transit as we rely on rcvif to stay valid.
561 	 */
562 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
563 	m->m_pkthdr.rcvif = oifp;
564 	CURVNET_SET_QUIET(oifp->if_vnet);
565 	error = netisr_queue(NETISR_EPAIR, m);
566 	CURVNET_RESTORE();
567 	if (!error) {
568 		if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
569 		/*
570 		 * IFQ_HANDOFF_ADJ/ip_handoff() update statistics,
571 		 * but as we bypass all this we have to duplicate
572 		 * the logic another time.
573 		 */
574 		if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
575 		if (mflags & (M_BCAST|M_MCAST))
576 			if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
577 		/* Someone else received the packet. */
578 		if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
579 	} else {
580 		/* The packet was freed already. */
581 		epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
582 		ifp->if_drv_flags |= IFF_DRV_OACTIVE;
583 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
584 		EPAIR_REFCOUNT_RELEASE(&sc->refcount);
585 		EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
586 		    ("%s: ifp=%p sc->refcount not >= 1: %d",
587 		    __func__, oifp, sc->refcount));
588 	}
589 
590 	return (error);
591 }
592 
593 static int
594 epair_transmit(struct ifnet *ifp, struct mbuf *m)
595 {
596 	struct epair_dpcpu *epair_dpcpu;
597 	int error;
598 
599 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
600 	EPAIR_LOCK(epair_dpcpu);
601 	error = epair_transmit_locked(ifp, m);
602 	EPAIR_UNLOCK(epair_dpcpu);
603 	return (error);
604 }
605 
606 static void
607 epair_qflush(struct ifnet *ifp)
608 {
609 	struct epair_softc *sc;
610 
611 	sc = ifp->if_softc;
612 	KASSERT(sc != NULL, ("%s: ifp=%p, epair_softc gone? sc=%p\n",
613 	    __func__, ifp, sc));
614 	/*
615 	 * Remove this ifp from all backpointer lists. The interface will not
616 	 * usable for flushing anyway nor should it have anything to flush
617 	 * after if_qflush().
618 	 */
619 	epair_remove_ifp_from_draining(ifp);
620 
621 	if (sc->if_qflush)
622 		sc->if_qflush(ifp);
623 }
624 
625 static int
626 epair_media_change(struct ifnet *ifp __unused)
627 {
628 
629 	/* Do nothing. */
630 	return (0);
631 }
632 
633 static void
634 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
635 {
636 
637 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
638 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
639 }
640 
641 static int
642 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
643 {
644 	struct epair_softc *sc;
645 	struct ifreq *ifr;
646 	int error;
647 
648 	ifr = (struct ifreq *)data;
649 	switch (cmd) {
650 	case SIOCSIFFLAGS:
651 	case SIOCADDMULTI:
652 	case SIOCDELMULTI:
653 		error = 0;
654 		break;
655 
656 	case SIOCSIFMEDIA:
657 	case SIOCGIFMEDIA:
658 		sc = ifp->if_softc;
659 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
660 		break;
661 
662 	case SIOCSIFMTU:
663 		/* We basically allow all kinds of MTUs. */
664 		ifp->if_mtu = ifr->ifr_mtu;
665 		error = 0;
666 		break;
667 
668 	default:
669 		/* Let the common ethernet handler process this. */
670 		error = ether_ioctl(ifp, cmd, data);
671 		break;
672 	}
673 
674 	return (error);
675 }
676 
677 static void
678 epair_init(void *dummy __unused)
679 {
680 }
681 
682 
683 /*
684  * Interface cloning functions.
685  * We use our private ones so that we can create/destroy our secondary
686  * device along with the primary one.
687  */
688 static int
689 epair_clone_match(struct if_clone *ifc, const char *name)
690 {
691 	const char *cp;
692 
693 	DPRINTF("name='%s'\n", name);
694 
695 	/*
696 	 * Our base name is epair.
697 	 * Our interfaces will be named epair<n>[ab].
698 	 * So accept anything of the following list:
699 	 * - epair
700 	 * - epair<n>
701 	 * but not the epair<n>[ab] versions.
702 	 */
703 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
704 		return (0);
705 
706 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
707 		if (*cp < '0' || *cp > '9')
708 			return (0);
709 	}
710 
711 	return (1);
712 }
713 
714 static int
715 epair_clone_create(struct if_clone *ifc, char *name, size_t len, caddr_t params)
716 {
717 	struct epair_softc *sca, *scb;
718 	struct ifnet *ifp;
719 	char *dp;
720 	int error, unit, wildcard;
721 	uint64_t hostid;
722 	uint32_t key[3];
723 	uint32_t hash;
724 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
725 
726 	/*
727 	 * We are abusing params to create our second interface.
728 	 * Actually we already created it and called if_clone_create()
729 	 * for it to do the official insertion procedure the moment we knew
730 	 * it cannot fail anymore. So just do attach it here.
731 	 */
732 	if (params) {
733 		scb = (struct epair_softc *)params;
734 		ifp = scb->ifp;
735 		/* Copy epairNa etheraddr and change the last byte. */
736 		memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
737 		eaddr[5] = 0x0b;
738 		ether_ifattach(ifp, eaddr);
739 		/* Correctly set the name for the cloner list. */
740 		strlcpy(name, ifp->if_xname, len);
741 		return (0);
742 	}
743 
744 	/* Try to see if a special unit was requested. */
745 	error = ifc_name2unit(name, &unit);
746 	if (error != 0)
747 		return (error);
748 	wildcard = (unit < 0);
749 
750 	error = ifc_alloc_unit(ifc, &unit);
751 	if (error != 0)
752 		return (error);
753 
754 	/*
755 	 * If no unit had been given, we need to adjust the ifName.
756 	 * Also make sure there is space for our extra [ab] suffix.
757 	 */
758 	for (dp = name; *dp != '\0'; dp++);
759 	if (wildcard) {
760 		error = snprintf(dp, len - (dp - name), "%d", unit);
761 		if (error > len - (dp - name) - 1) {
762 			/* ifName too long. */
763 			ifc_free_unit(ifc, unit);
764 			return (ENOSPC);
765 		}
766 		dp += error;
767 	}
768 	if (len - (dp - name) - 1 < 1) {
769 		/* No space left for our [ab] suffix. */
770 		ifc_free_unit(ifc, unit);
771 		return (ENOSPC);
772 	}
773 	*dp = 'b';
774 	/* Must not change dp so we can replace 'a' by 'b' later. */
775 	*(dp+1) = '\0';
776 
777 	/* Check if 'a' and 'b' interfaces already exist. */
778 	if (ifunit(name) != NULL)
779 		return (EEXIST);
780 	*dp = 'a';
781 	if (ifunit(name) != NULL)
782 		return (EEXIST);
783 
784 	/* Allocate memory for both [ab] interfaces */
785 	sca = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
786 	EPAIR_REFCOUNT_INIT(&sca->refcount, 1);
787 	sca->ifp = if_alloc(IFT_ETHER);
788 	if (sca->ifp == NULL) {
789 		free(sca, M_EPAIR);
790 		ifc_free_unit(ifc, unit);
791 		return (ENOSPC);
792 	}
793 
794 	scb = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
795 	EPAIR_REFCOUNT_INIT(&scb->refcount, 1);
796 	scb->ifp = if_alloc(IFT_ETHER);
797 	if (scb->ifp == NULL) {
798 		free(scb, M_EPAIR);
799 		if_free(sca->ifp);
800 		free(sca, M_EPAIR);
801 		ifc_free_unit(ifc, unit);
802 		return (ENOSPC);
803 	}
804 
805 	/*
806 	 * Cross-reference the interfaces so we will be able to free both.
807 	 */
808 	sca->oifp = scb->ifp;
809 	scb->oifp = sca->ifp;
810 
811 	/*
812 	 * Calculate the cpuid for netisr queueing based on the
813 	 * ifIndex of the interfaces. As long as we cannot configure
814 	 * this or use cpuset information easily we cannot guarantee
815 	 * cache locality but we can at least allow parallelism.
816 	 */
817 	sca->cpuid =
818 	    netisr_get_cpuid(sca->ifp->if_index);
819 	scb->cpuid =
820 	    netisr_get_cpuid(scb->ifp->if_index);
821 
822 	/* Initialise pseudo media types. */
823 	ifmedia_init(&sca->media, 0, epair_media_change, epair_media_status);
824 	ifmedia_add(&sca->media, IFM_ETHER | IFM_10G_T, 0, NULL);
825 	ifmedia_set(&sca->media, IFM_ETHER | IFM_10G_T);
826 	ifmedia_init(&scb->media, 0, epair_media_change, epair_media_status);
827 	ifmedia_add(&scb->media, IFM_ETHER | IFM_10G_T, 0, NULL);
828 	ifmedia_set(&scb->media, IFM_ETHER | IFM_10G_T);
829 
830 	/* Finish initialization of interface <n>a. */
831 	ifp = sca->ifp;
832 	ifp->if_softc = sca;
833 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
834 	ifp->if_dname = epairname;
835 	ifp->if_dunit = unit;
836 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
837 	ifp->if_capabilities = IFCAP_VLAN_MTU;
838 	ifp->if_capenable = IFCAP_VLAN_MTU;
839 	ifp->if_start = epair_start;
840 	ifp->if_ioctl = epair_ioctl;
841 	ifp->if_init  = epair_init;
842 	if_setsendqlen(ifp, ifqmaxlen);
843 	if_setsendqready(ifp);
844 
845 	/*
846 	 * Calculate the etheraddr hashing the hostid and the
847 	 * interface index. The result would be hopefully unique.
848 	 * Note that the "a" component of an epair instance may get moved
849 	 * to a different VNET after creation. In that case its index
850 	 * will be freed and the index can get reused by new epair instance.
851 	 * Make sure we do not create same etheraddr again.
852 	 */
853 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
854 	if (hostid == 0)
855 		arc4rand(&hostid, sizeof(hostid), 0);
856 
857 	if (ifp->if_index > next_index)
858 		next_index = ifp->if_index;
859 	else
860 		next_index++;
861 
862 	key[0] = (uint32_t)next_index;
863 	key[1] = (uint32_t)(hostid & 0xffffffff);
864 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
865 	hash = jenkins_hash32(key, 3, 0);
866 
867 	eaddr[0] = 0x02;
868 	memcpy(&eaddr[1], &hash, 4);
869 	eaddr[5] = 0x0a;
870 	ether_ifattach(ifp, eaddr);
871 	sca->if_qflush = ifp->if_qflush;
872 	ifp->if_qflush = epair_qflush;
873 	ifp->if_transmit = epair_transmit;
874 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
875 
876 	/* Swap the name and finish initialization of interface <n>b. */
877 	*dp = 'b';
878 
879 	ifp = scb->ifp;
880 	ifp->if_softc = scb;
881 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
882 	ifp->if_dname = epairname;
883 	ifp->if_dunit = unit;
884 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
885 	ifp->if_capabilities = IFCAP_VLAN_MTU;
886 	ifp->if_capenable = IFCAP_VLAN_MTU;
887 	ifp->if_start = epair_start;
888 	ifp->if_ioctl = epair_ioctl;
889 	ifp->if_init  = epair_init;
890 	if_setsendqlen(ifp, ifqmaxlen);
891 	if_setsendqready(ifp);
892 	/* We need to play some tricks here for the second interface. */
893 	strlcpy(name, epairname, len);
894 	error = if_clone_create(name, len, (caddr_t)scb);
895 	if (error)
896 		panic("%s: if_clone_create() for our 2nd iface failed: %d",
897 		    __func__, error);
898 	scb->if_qflush = ifp->if_qflush;
899 	ifp->if_qflush = epair_qflush;
900 	ifp->if_transmit = epair_transmit;
901 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
902 
903 	/*
904 	 * Restore name to <n>a as the ifp for this will go into the
905 	 * cloner list for the initial call.
906 	 */
907 	strlcpy(name, sca->ifp->if_xname, len);
908 	DPRINTF("name='%s/%db' created sca=%p scb=%p\n", name, unit, sca, scb);
909 
910 	/* Tell the world, that we are ready to rock. */
911 	sca->ifp->if_drv_flags |= IFF_DRV_RUNNING;
912 	scb->ifp->if_drv_flags |= IFF_DRV_RUNNING;
913 	if_link_state_change(sca->ifp, LINK_STATE_UP);
914 	if_link_state_change(scb->ifp, LINK_STATE_UP);
915 
916 	return (0);
917 }
918 
919 static int
920 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp)
921 {
922 	struct ifnet *oifp;
923 	struct epair_softc *sca, *scb;
924 	int unit, error;
925 
926 	DPRINTF("ifp=%p\n", ifp);
927 
928 	/*
929 	 * In case we called into if_clone_destroyif() ourselves
930 	 * again to remove the second interface, the softc will be
931 	 * NULL. In that case so not do anything but return success.
932 	 */
933 	if (ifp->if_softc == NULL)
934 		return (0);
935 
936 	unit = ifp->if_dunit;
937 	sca = ifp->if_softc;
938 	oifp = sca->oifp;
939 	scb = oifp->if_softc;
940 
941 	DPRINTF("ifp=%p oifp=%p\n", ifp, oifp);
942 	if_link_state_change(ifp, LINK_STATE_DOWN);
943 	if_link_state_change(oifp, LINK_STATE_DOWN);
944 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
945 	oifp->if_drv_flags &= ~IFF_DRV_RUNNING;
946 
947 	/*
948 	 * Get rid of our second half. As the other of the two
949 	 * interfaces may reside in a different vnet, we need to
950 	 * switch before freeing them.
951 	 */
952 	CURVNET_SET_QUIET(oifp->if_vnet);
953 	ether_ifdetach(oifp);
954 	/*
955 	 * Wait for all packets to be dispatched to if_input.
956 	 * The numbers can only go down as the interface is
957 	 * detached so there is no need to use atomics.
958 	 */
959 	DPRINTF("scb refcnt=%u\n", scb->refcount);
960 	EPAIR_REFCOUNT_ASSERT(scb->refcount == 1,
961 	    ("%s: ifp=%p scb->refcount!=1: %d", __func__, oifp, scb->refcount));
962 	oifp->if_softc = NULL;
963 	error = if_clone_destroyif(ifc, oifp);
964 	if (error)
965 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
966 		    __func__, error);
967 	if_free(oifp);
968 	ifmedia_removeall(&scb->media);
969 	free(scb, M_EPAIR);
970 	CURVNET_RESTORE();
971 
972 	ether_ifdetach(ifp);
973 	/*
974 	 * Wait for all packets to be dispatched to if_input.
975 	 */
976 	DPRINTF("sca refcnt=%u\n", sca->refcount);
977 	EPAIR_REFCOUNT_ASSERT(sca->refcount == 1,
978 	    ("%s: ifp=%p sca->refcount!=1: %d", __func__, ifp, sca->refcount));
979 	if_free(ifp);
980 	ifmedia_removeall(&sca->media);
981 	free(sca, M_EPAIR);
982 	ifc_free_unit(ifc, unit);
983 
984 	return (0);
985 }
986 
987 static void
988 vnet_epair_init(const void *unused __unused)
989 {
990 
991 	V_epair_cloner = if_clone_advanced(epairname, 0,
992 	    epair_clone_match, epair_clone_create, epair_clone_destroy);
993 #ifdef VIMAGE
994 	netisr_register_vnet(&epair_nh);
995 #endif
996 }
997 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
998     vnet_epair_init, NULL);
999 
1000 static void
1001 vnet_epair_uninit(const void *unused __unused)
1002 {
1003 
1004 #ifdef VIMAGE
1005 	netisr_unregister_vnet(&epair_nh);
1006 #endif
1007 	if_clone_detach(V_epair_cloner);
1008 }
1009 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
1010     vnet_epair_uninit, NULL);
1011 
1012 static void
1013 epair_uninit(const void *unused __unused)
1014 {
1015 	netisr_unregister(&epair_nh);
1016 	epair_dpcpu_detach();
1017 	if (bootverbose)
1018 		printf("%s unloaded.\n", epairname);
1019 }
1020 SYSUNINIT(epair_uninit, SI_SUB_INIT_IF, SI_ORDER_MIDDLE,
1021     epair_uninit, NULL);
1022 
1023 static int
1024 epair_modevent(module_t mod, int type, void *data)
1025 {
1026 	int qlimit;
1027 
1028 	switch (type) {
1029 	case MOD_LOAD:
1030 		/* For now limit us to one global mutex and one inq. */
1031 		epair_dpcpu_init();
1032 		epair_nh.nh_qlimit = 42 * ifqmaxlen; /* 42 shall be the number. */
1033 		if (TUNABLE_INT_FETCH("net.link.epair.netisr_maxqlen", &qlimit))
1034 		    epair_nh.nh_qlimit = qlimit;
1035 		netisr_register(&epair_nh);
1036 		if (bootverbose)
1037 			printf("%s initialized.\n", epairname);
1038 		break;
1039 	case MOD_UNLOAD:
1040 		/* Handled in epair_uninit() */
1041 		break;
1042 	default:
1043 		return (EOPNOTSUPP);
1044 	}
1045 	return (0);
1046 }
1047 
1048 static moduledata_t epair_mod = {
1049 	"if_epair",
1050 	epair_modevent,
1051 	0
1052 };
1053 
1054 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
1055 MODULE_VERSION(if_epair, 1);
1056