xref: /freebsd/sys/net/if_epair.c (revision c6989859ae9388eeb46a24fe88f9b8d07101c710)
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 | CTLFLAG_MPSAFE, 0,
85     "epair sysctl");
86 
87 #ifdef EPAIR_DEBUG
88 static int epair_debug = 0;
89 SYSCTL_INT(_net_link_epair, OID_AUTO, epair_debug, CTLFLAG_RW,
90     &epair_debug, 0, "if_epair(4) debugging.");
91 #define	DPRINTF(fmt, arg...)						\
92 	if (epair_debug)						\
93 		printf("[%s:%d] " fmt, __func__, __LINE__, ##arg)
94 #else
95 #define	DPRINTF(fmt, arg...)
96 #endif
97 
98 static void epair_nh_sintr(struct mbuf *);
99 static struct mbuf *epair_nh_m2cpuid(struct mbuf *, uintptr_t, u_int *);
100 static void epair_nh_drainedcpu(u_int);
101 
102 static void epair_start_locked(struct ifnet *);
103 static int epair_media_change(struct ifnet *);
104 static void epair_media_status(struct ifnet *, struct ifmediareq *);
105 
106 static int epair_clone_match(struct if_clone *, const char *);
107 static int epair_clone_create(struct if_clone *, char *, size_t, caddr_t);
108 static int epair_clone_destroy(struct if_clone *, struct ifnet *);
109 
110 static const char epairname[] = "epair";
111 static unsigned int next_index = 0;
112 
113 /* Netisr related definitions and sysctl. */
114 static struct netisr_handler epair_nh = {
115 	.nh_name	= epairname,
116 	.nh_proto	= NETISR_EPAIR,
117 	.nh_policy	= NETISR_POLICY_CPU,
118 	.nh_handler	= epair_nh_sintr,
119 	.nh_m2cpuid	= epair_nh_m2cpuid,
120 	.nh_drainedcpu	= epair_nh_drainedcpu,
121 };
122 
123 static int
124 sysctl_epair_netisr_maxqlen(SYSCTL_HANDLER_ARGS)
125 {
126 	int error, qlimit;
127 
128 	netisr_getqlimit(&epair_nh, &qlimit);
129 	error = sysctl_handle_int(oidp, &qlimit, 0, req);
130 	if (error || !req->newptr)
131 		return (error);
132 	if (qlimit < 1)
133 		return (EINVAL);
134 	return (netisr_setqlimit(&epair_nh, qlimit));
135 }
136 SYSCTL_PROC(_net_link_epair, OID_AUTO, netisr_maxqlen,
137     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
138     sysctl_epair_netisr_maxqlen, "I",
139     "Maximum if_epair(4) netisr \"hw\" queue length");
140 
141 struct epair_softc {
142 	struct ifnet	*ifp;		/* This ifp. */
143 	struct ifnet	*oifp;		/* other ifp of pair. */
144 	struct ifmedia	media;		/* Media config (fake). */
145 	u_int		refcount;	/* # of mbufs in flight. */
146 	u_int		cpuid;		/* CPU ID assigned upon creation. */
147 	void		(*if_qflush)(struct ifnet *);
148 					/* Original if_qflush routine. */
149 };
150 
151 /*
152  * Per-CPU list of ifps with data in the ifq that needs to be flushed
153  * to the netisr ``hw'' queue before we allow any further direct queuing
154  * to the ``hw'' queue.
155  */
156 struct epair_ifp_drain {
157 	STAILQ_ENTRY(epair_ifp_drain)	ifp_next;
158 	struct ifnet			*ifp;
159 };
160 STAILQ_HEAD(eid_list, epair_ifp_drain);
161 
162 #define	EPAIR_LOCK_INIT(dpcpu)		mtx_init(&(dpcpu)->if_epair_mtx, \
163 					    "if_epair", NULL, MTX_DEF)
164 #define	EPAIR_LOCK_DESTROY(dpcpu)	mtx_destroy(&(dpcpu)->if_epair_mtx)
165 #define	EPAIR_LOCK_ASSERT(dpcpu)	mtx_assert(&(dpcpu)->if_epair_mtx, \
166 					    MA_OWNED)
167 #define	EPAIR_LOCK(dpcpu)		mtx_lock(&(dpcpu)->if_epair_mtx)
168 #define	EPAIR_UNLOCK(dpcpu)		mtx_unlock(&(dpcpu)->if_epair_mtx)
169 
170 #ifdef INVARIANTS
171 #define	EPAIR_REFCOUNT_INIT(r, v)	refcount_init((r), (v))
172 #define	EPAIR_REFCOUNT_AQUIRE(r)	refcount_acquire((r))
173 #define	EPAIR_REFCOUNT_RELEASE(r)	refcount_release((r))
174 #define	EPAIR_REFCOUNT_ASSERT(a, p)	KASSERT(a, p)
175 #else
176 #define	EPAIR_REFCOUNT_INIT(r, v)
177 #define	EPAIR_REFCOUNT_AQUIRE(r)
178 #define	EPAIR_REFCOUNT_RELEASE(r)
179 #define	EPAIR_REFCOUNT_ASSERT(a, p)
180 #endif
181 
182 static MALLOC_DEFINE(M_EPAIR, epairname,
183     "Pair of virtual cross-over connected Ethernet-like interfaces");
184 
185 VNET_DEFINE_STATIC(struct if_clone *, epair_cloner);
186 #define	V_epair_cloner	VNET(epair_cloner)
187 
188 /*
189  * DPCPU area and functions.
190  */
191 struct epair_dpcpu {
192 	struct mtx	if_epair_mtx;		/* Per-CPU locking. */
193 	int		epair_drv_flags;	/* Per-CPU ``hw'' drv flags. */
194 	struct eid_list	epair_ifp_drain_list;	/* Per-CPU list of ifps with
195 						 * data in the ifq. */
196 };
197 DPCPU_DEFINE(struct epair_dpcpu, epair_dpcpu);
198 
199 static void
200 epair_dpcpu_init(void)
201 {
202 	struct epair_dpcpu *epair_dpcpu;
203 	struct eid_list *s;
204 	u_int cpuid;
205 
206 	CPU_FOREACH(cpuid) {
207 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
208 
209 		/* Initialize per-cpu lock. */
210 		EPAIR_LOCK_INIT(epair_dpcpu);
211 
212 		/* Driver flags are per-cpu as are our netisr "hw" queues. */
213 		epair_dpcpu->epair_drv_flags = 0;
214 
215 		/*
216 		 * Initialize per-cpu drain list.
217 		 * Manually do what STAILQ_HEAD_INITIALIZER would do.
218 		 */
219 		s = &epair_dpcpu->epair_ifp_drain_list;
220 		s->stqh_first = NULL;
221 		s->stqh_last = &s->stqh_first;
222 	}
223 }
224 
225 static void
226 epair_dpcpu_detach(void)
227 {
228 	struct epair_dpcpu *epair_dpcpu;
229 	u_int cpuid;
230 
231 	CPU_FOREACH(cpuid) {
232 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
233 
234 		/* Destroy per-cpu lock. */
235 		EPAIR_LOCK_DESTROY(epair_dpcpu);
236 	}
237 }
238 
239 /*
240  * Helper functions.
241  */
242 static u_int
243 cpuid_from_ifp(struct ifnet *ifp)
244 {
245 	struct epair_softc *sc;
246 
247 	if (ifp == NULL)
248 		return (0);
249 	sc = ifp->if_softc;
250 
251 	return (sc->cpuid);
252 }
253 
254 /*
255  * Netisr handler functions.
256  */
257 static void
258 epair_nh_sintr(struct mbuf *m)
259 {
260 	struct ifnet *ifp;
261 	struct epair_softc *sc __unused;
262 
263 	ifp = m->m_pkthdr.rcvif;
264 	(*ifp->if_input)(ifp, m);
265 	sc = ifp->if_softc;
266 	EPAIR_REFCOUNT_RELEASE(&sc->refcount);
267 	EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
268 	    ("%s: ifp=%p sc->refcount not >= 1: %d",
269 	    __func__, ifp, sc->refcount));
270 	DPRINTF("ifp=%p refcount=%u\n", ifp, sc->refcount);
271 }
272 
273 static struct mbuf *
274 epair_nh_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
275 {
276 
277 	*cpuid = cpuid_from_ifp(m->m_pkthdr.rcvif);
278 
279 	return (m);
280 }
281 
282 static void
283 epair_nh_drainedcpu(u_int cpuid)
284 {
285 	struct epair_dpcpu *epair_dpcpu;
286 	struct epair_ifp_drain *elm, *tvar;
287 	struct ifnet *ifp;
288 
289 	epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
290 	EPAIR_LOCK(epair_dpcpu);
291 	/*
292 	 * Assume our "hw" queue and possibly ifq will be emptied
293 	 * again. In case we will overflow the "hw" queue while
294 	 * draining, epair_start_locked will set IFF_DRV_OACTIVE
295 	 * again and we will stop and return.
296 	 */
297 	STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
298 	    ifp_next, tvar) {
299 		ifp = elm->ifp;
300 		epair_dpcpu->epair_drv_flags &= ~IFF_DRV_OACTIVE;
301 		ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
302 		epair_start_locked(ifp);
303 
304 		IFQ_LOCK(&ifp->if_snd);
305 		if (IFQ_IS_EMPTY(&ifp->if_snd)) {
306 			struct epair_softc *sc __unused;
307 
308 			STAILQ_REMOVE(&epair_dpcpu->epair_ifp_drain_list,
309 			    elm, epair_ifp_drain, ifp_next);
310 			/* The cached ifp goes off the list. */
311 			sc = ifp->if_softc;
312 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
313 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
314 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
315 			    __func__, ifp, sc->refcount));
316 			free(elm, M_EPAIR);
317 		}
318 		IFQ_UNLOCK(&ifp->if_snd);
319 
320 		if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) != 0) {
321 			/* Our "hw"q overflew again. */
322 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
323 			DPRINTF("hw queue length overflow at %u\n",
324 			    epair_nh.nh_qlimit);
325 			break;
326 		}
327 	}
328 	EPAIR_UNLOCK(epair_dpcpu);
329 }
330 
331 /*
332  * Network interface (`if') related functions.
333  */
334 static void
335 epair_remove_ifp_from_draining(struct ifnet *ifp)
336 {
337 	struct epair_dpcpu *epair_dpcpu;
338 	struct epair_ifp_drain *elm, *tvar;
339 	u_int cpuid;
340 
341 	CPU_FOREACH(cpuid) {
342 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
343 		EPAIR_LOCK(epair_dpcpu);
344 		STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
345 		    ifp_next, tvar) {
346 			if (ifp == elm->ifp) {
347 				struct epair_softc *sc __unused;
348 
349 				STAILQ_REMOVE(
350 				    &epair_dpcpu->epair_ifp_drain_list, elm,
351 				    epair_ifp_drain, ifp_next);
352 				/* The cached ifp goes off the list. */
353 				sc = ifp->if_softc;
354 				EPAIR_REFCOUNT_RELEASE(&sc->refcount);
355 				EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
356 				    ("%s: ifp=%p sc->refcount not >= 1: %d",
357 				    __func__, ifp, sc->refcount));
358 				free(elm, M_EPAIR);
359 			}
360 		}
361 		EPAIR_UNLOCK(epair_dpcpu);
362 	}
363 }
364 
365 static int
366 epair_add_ifp_for_draining(struct ifnet *ifp)
367 {
368 	struct epair_dpcpu *epair_dpcpu;
369 	struct epair_softc *sc;
370 	struct epair_ifp_drain *elm = NULL;
371 
372 	sc = ifp->if_softc;
373 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
374 	EPAIR_LOCK_ASSERT(epair_dpcpu);
375 	STAILQ_FOREACH(elm, &epair_dpcpu->epair_ifp_drain_list, ifp_next)
376 		if (elm->ifp == ifp)
377 			break;
378 	/* If the ifp is there already, return success. */
379 	if (elm != NULL)
380 		return (0);
381 
382 	elm = malloc(sizeof(struct epair_ifp_drain), M_EPAIR, M_NOWAIT|M_ZERO);
383 	if (elm == NULL)
384 		return (ENOMEM);
385 
386 	elm->ifp = ifp;
387 	/* Add a reference for the ifp pointer on the list. */
388 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
389 	STAILQ_INSERT_TAIL(&epair_dpcpu->epair_ifp_drain_list, elm, ifp_next);
390 
391 	return (0);
392 }
393 
394 static void
395 epair_start_locked(struct ifnet *ifp)
396 {
397 	struct epair_dpcpu *epair_dpcpu;
398 	struct mbuf *m;
399 	struct epair_softc *sc;
400 	struct ifnet *oifp;
401 	int error;
402 
403 	DPRINTF("ifp=%p\n", ifp);
404 	sc = ifp->if_softc;
405 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
406 	EPAIR_LOCK_ASSERT(epair_dpcpu);
407 
408 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
409 		return;
410 	if ((ifp->if_flags & IFF_UP) == 0)
411 		return;
412 
413 	/*
414 	 * We get packets here from ether_output via if_handoff()
415 	 * and need to put them into the input queue of the oifp
416 	 * and call oifp->if_input() via netisr/epair_sintr().
417 	 */
418 	oifp = sc->oifp;
419 	sc = oifp->if_softc;
420 	for (;;) {
421 		IFQ_DEQUEUE(&ifp->if_snd, m);
422 		if (m == NULL)
423 			break;
424 		BPF_MTAP(ifp, m);
425 
426 		/*
427 		 * In case the outgoing interface is not usable,
428 		 * drop the packet.
429 		 */
430 		if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
431 		    (oifp->if_flags & IFF_UP) ==0) {
432 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
433 			m_freem(m);
434 			continue;
435 		}
436 		DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
437 
438 		/*
439 		 * Add a reference so the interface cannot go while the
440 		 * packet is in transit as we rely on rcvif to stay valid.
441 		 */
442 		EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
443 		m->m_pkthdr.rcvif = oifp;
444 		CURVNET_SET_QUIET(oifp->if_vnet);
445 		error = netisr_queue(NETISR_EPAIR, m);
446 		CURVNET_RESTORE();
447 		if (!error) {
448 			if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
449 			/* Someone else received the packet. */
450 			if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
451 		} else {
452 			/* The packet was freed already. */
453 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
454 			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
455 			(void) epair_add_ifp_for_draining(ifp);
456 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
457 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
458 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
459 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
460 			    __func__, oifp, sc->refcount));
461 		}
462 	}
463 }
464 
465 static void
466 epair_start(struct ifnet *ifp)
467 {
468 	struct epair_dpcpu *epair_dpcpu;
469 
470 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
471 	EPAIR_LOCK(epair_dpcpu);
472 	epair_start_locked(ifp);
473 	EPAIR_UNLOCK(epair_dpcpu);
474 }
475 
476 static int
477 epair_transmit_locked(struct ifnet *ifp, struct mbuf *m)
478 {
479 	struct epair_dpcpu *epair_dpcpu;
480 	struct epair_softc *sc;
481 	struct ifnet *oifp;
482 	int error, len;
483 	short mflags;
484 
485 	DPRINTF("ifp=%p m=%p\n", ifp, m);
486 	sc = ifp->if_softc;
487 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
488 	EPAIR_LOCK_ASSERT(epair_dpcpu);
489 
490 	if (m == NULL)
491 		return (0);
492 
493 	/*
494 	 * We are not going to use the interface en/dequeue mechanism
495 	 * on the TX side. We are called from ether_output_frame()
496 	 * and will put the packet into the incoming queue of the
497 	 * other interface of our pair via the netsir.
498 	 */
499 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
500 		m_freem(m);
501 		return (ENXIO);
502 	}
503 	if ((ifp->if_flags & IFF_UP) == 0) {
504 		m_freem(m);
505 		return (ENETDOWN);
506 	}
507 
508 	BPF_MTAP(ifp, m);
509 
510 	/*
511 	 * In case the outgoing interface is not usable,
512 	 * drop the packet.
513 	 */
514 	oifp = sc->oifp;
515 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
516 	    (oifp->if_flags & IFF_UP) ==0) {
517 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
518 		m_freem(m);
519 		return (0);
520 	}
521 	len = m->m_pkthdr.len;
522 	mflags = m->m_flags;
523 	DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
524 
525 #ifdef ALTQ
526 	/* Support ALTQ via the classic if_start() path. */
527 	IF_LOCK(&ifp->if_snd);
528 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
529 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
530 		if (error)
531 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
532 		IF_UNLOCK(&ifp->if_snd);
533 		if (!error) {
534 			if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
535 			if (mflags & (M_BCAST|M_MCAST))
536 				if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
537 
538 			if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) == 0)
539 				epair_start_locked(ifp);
540 			else
541 				(void)epair_add_ifp_for_draining(ifp);
542 		}
543 		return (error);
544 	}
545 	IF_UNLOCK(&ifp->if_snd);
546 #endif
547 
548 	if ((epair_dpcpu->epair_drv_flags & IFF_DRV_OACTIVE) != 0) {
549 		/*
550 		 * Our hardware queue is full, try to fall back
551 		 * queuing to the ifq but do not call ifp->if_start.
552 		 * Either we are lucky or the packet is gone.
553 		 */
554 		IFQ_ENQUEUE(&ifp->if_snd, m, error);
555 		if (!error)
556 			(void)epair_add_ifp_for_draining(ifp);
557 		return (error);
558 	}
559 	sc = oifp->if_softc;
560 	/*
561 	 * Add a reference so the interface cannot go while the
562 	 * packet is in transit as we rely on rcvif to stay valid.
563 	 */
564 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
565 	m->m_pkthdr.rcvif = oifp;
566 	CURVNET_SET_QUIET(oifp->if_vnet);
567 	error = netisr_queue(NETISR_EPAIR, m);
568 	CURVNET_RESTORE();
569 	if (!error) {
570 		if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
571 		/*
572 		 * IFQ_HANDOFF_ADJ/ip_handoff() update statistics,
573 		 * but as we bypass all this we have to duplicate
574 		 * the logic another time.
575 		 */
576 		if_inc_counter(ifp, IFCOUNTER_OBYTES, len);
577 		if (mflags & (M_BCAST|M_MCAST))
578 			if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
579 		/* Someone else received the packet. */
580 		if_inc_counter(oifp, IFCOUNTER_IPACKETS, 1);
581 	} else {
582 		/* The packet was freed already. */
583 		epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
584 		ifp->if_drv_flags |= IFF_DRV_OACTIVE;
585 		if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
586 		EPAIR_REFCOUNT_RELEASE(&sc->refcount);
587 		EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
588 		    ("%s: ifp=%p sc->refcount not >= 1: %d",
589 		    __func__, oifp, sc->refcount));
590 	}
591 
592 	return (error);
593 }
594 
595 static int
596 epair_transmit(struct ifnet *ifp, struct mbuf *m)
597 {
598 	struct epair_dpcpu *epair_dpcpu;
599 	int error;
600 
601 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
602 	EPAIR_LOCK(epair_dpcpu);
603 	error = epair_transmit_locked(ifp, m);
604 	EPAIR_UNLOCK(epair_dpcpu);
605 	return (error);
606 }
607 
608 static void
609 epair_qflush(struct ifnet *ifp)
610 {
611 	struct epair_softc *sc;
612 
613 	sc = ifp->if_softc;
614 
615 	/*
616 	 * See epair_clone_destroy(), we can end up getting called twice.
617 	 * Don't do anything on the second call.
618 	 */
619 	if (sc == NULL)
620 		return;
621 
622 	/*
623 	 * Remove this ifp from all backpointer lists. The interface will not
624 	 * usable for flushing anyway nor should it have anything to flush
625 	 * after if_qflush().
626 	 */
627 	epair_remove_ifp_from_draining(ifp);
628 
629 	if (sc->if_qflush)
630 		sc->if_qflush(ifp);
631 }
632 
633 static int
634 epair_media_change(struct ifnet *ifp __unused)
635 {
636 
637 	/* Do nothing. */
638 	return (0);
639 }
640 
641 static void
642 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
643 {
644 
645 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
646 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
647 }
648 
649 static int
650 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
651 {
652 	struct epair_softc *sc;
653 	struct ifreq *ifr;
654 	int error;
655 
656 	ifr = (struct ifreq *)data;
657 	switch (cmd) {
658 	case SIOCSIFFLAGS:
659 	case SIOCADDMULTI:
660 	case SIOCDELMULTI:
661 		error = 0;
662 		break;
663 
664 	case SIOCSIFMEDIA:
665 	case SIOCGIFMEDIA:
666 		sc = ifp->if_softc;
667 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
668 		break;
669 
670 	case SIOCSIFMTU:
671 		/* We basically allow all kinds of MTUs. */
672 		ifp->if_mtu = ifr->ifr_mtu;
673 		error = 0;
674 		break;
675 
676 	default:
677 		/* Let the common ethernet handler process this. */
678 		error = ether_ioctl(ifp, cmd, data);
679 		break;
680 	}
681 
682 	return (error);
683 }
684 
685 static void
686 epair_init(void *dummy __unused)
687 {
688 }
689 
690 /*
691  * Interface cloning functions.
692  * We use our private ones so that we can create/destroy our secondary
693  * device along with the primary one.
694  */
695 static int
696 epair_clone_match(struct if_clone *ifc, const char *name)
697 {
698 	const char *cp;
699 
700 	DPRINTF("name='%s'\n", name);
701 
702 	/*
703 	 * Our base name is epair.
704 	 * Our interfaces will be named epair<n>[ab].
705 	 * So accept anything of the following list:
706 	 * - epair
707 	 * - epair<n>
708 	 * but not the epair<n>[ab] versions.
709 	 */
710 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
711 		return (0);
712 
713 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
714 		if (*cp < '0' || *cp > '9')
715 			return (0);
716 	}
717 
718 	return (1);
719 }
720 
721 static void
722 epair_clone_add(struct if_clone *ifc, struct epair_softc *scb)
723 {
724 	struct ifnet *ifp;
725 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
726 
727 	ifp = scb->ifp;
728 	/* Copy epairNa etheraddr and change the last byte. */
729 	memcpy(eaddr, scb->oifp->if_hw_addr, ETHER_ADDR_LEN);
730 	eaddr[5] = 0x0b;
731 	ether_ifattach(ifp, eaddr);
732 
733 	if_clone_addif(ifc, ifp);
734 }
735 
736 static int
737 epair_clone_create(struct if_clone *ifc, char *name, size_t len, caddr_t params)
738 {
739 	struct epair_softc *sca, *scb;
740 	struct ifnet *ifp;
741 	char *dp;
742 	int error, unit, wildcard;
743 	uint64_t hostid;
744 	uint32_t key[3];
745 	uint32_t hash;
746 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
747 
748 	/* Try to see if a special unit was requested. */
749 	error = ifc_name2unit(name, &unit);
750 	if (error != 0)
751 		return (error);
752 	wildcard = (unit < 0);
753 
754 	error = ifc_alloc_unit(ifc, &unit);
755 	if (error != 0)
756 		return (error);
757 
758 	/*
759 	 * If no unit had been given, we need to adjust the ifName.
760 	 * Also make sure there is space for our extra [ab] suffix.
761 	 */
762 	for (dp = name; *dp != '\0'; dp++);
763 	if (wildcard) {
764 		error = snprintf(dp, len - (dp - name), "%d", unit);
765 		if (error > len - (dp - name) - 1) {
766 			/* ifName too long. */
767 			ifc_free_unit(ifc, unit);
768 			return (ENOSPC);
769 		}
770 		dp += error;
771 	}
772 	if (len - (dp - name) - 1 < 1) {
773 		/* No space left for our [ab] suffix. */
774 		ifc_free_unit(ifc, unit);
775 		return (ENOSPC);
776 	}
777 	*dp = 'b';
778 	/* Must not change dp so we can replace 'a' by 'b' later. */
779 	*(dp+1) = '\0';
780 
781 	/* Check if 'a' and 'b' interfaces already exist. */
782 	if (ifunit(name) != NULL)
783 		return (EEXIST);
784 	*dp = 'a';
785 	if (ifunit(name) != NULL)
786 		return (EEXIST);
787 
788 	/* Allocate memory for both [ab] interfaces */
789 	sca = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
790 	EPAIR_REFCOUNT_INIT(&sca->refcount, 1);
791 	sca->ifp = if_alloc(IFT_ETHER);
792 	if (sca->ifp == NULL) {
793 		free(sca, M_EPAIR);
794 		ifc_free_unit(ifc, unit);
795 		return (ENOSPC);
796 	}
797 
798 	scb = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
799 	EPAIR_REFCOUNT_INIT(&scb->refcount, 1);
800 	scb->ifp = if_alloc(IFT_ETHER);
801 	if (scb->ifp == NULL) {
802 		free(scb, M_EPAIR);
803 		if_free(sca->ifp);
804 		free(sca, M_EPAIR);
805 		ifc_free_unit(ifc, unit);
806 		return (ENOSPC);
807 	}
808 
809 	/*
810 	 * Cross-reference the interfaces so we will be able to free both.
811 	 */
812 	sca->oifp = scb->ifp;
813 	scb->oifp = sca->ifp;
814 
815 	/*
816 	 * Calculate the cpuid for netisr queueing based on the
817 	 * ifIndex of the interfaces. As long as we cannot configure
818 	 * this or use cpuset information easily we cannot guarantee
819 	 * cache locality but we can at least allow parallelism.
820 	 */
821 	sca->cpuid =
822 	    netisr_get_cpuid(sca->ifp->if_index);
823 	scb->cpuid =
824 	    netisr_get_cpuid(scb->ifp->if_index);
825 
826 	/* Initialise pseudo media types. */
827 	ifmedia_init(&sca->media, 0, epair_media_change, epair_media_status);
828 	ifmedia_add(&sca->media, IFM_ETHER | IFM_10G_T, 0, NULL);
829 	ifmedia_set(&sca->media, IFM_ETHER | IFM_10G_T);
830 	ifmedia_init(&scb->media, 0, epair_media_change, epair_media_status);
831 	ifmedia_add(&scb->media, IFM_ETHER | IFM_10G_T, 0, NULL);
832 	ifmedia_set(&scb->media, IFM_ETHER | IFM_10G_T);
833 
834 	/* Finish initialization of interface <n>a. */
835 	ifp = sca->ifp;
836 	ifp->if_softc = sca;
837 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
838 	ifp->if_dname = epairname;
839 	ifp->if_dunit = unit;
840 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
841 	ifp->if_capabilities = IFCAP_VLAN_MTU;
842 	ifp->if_capenable = IFCAP_VLAN_MTU;
843 	ifp->if_start = epair_start;
844 	ifp->if_ioctl = epair_ioctl;
845 	ifp->if_init  = epair_init;
846 	if_setsendqlen(ifp, ifqmaxlen);
847 	if_setsendqready(ifp);
848 
849 	/*
850 	 * Calculate the etheraddr hashing the hostid and the
851 	 * interface index. The result would be hopefully unique.
852 	 * Note that the "a" component of an epair instance may get moved
853 	 * to a different VNET after creation. In that case its index
854 	 * will be freed and the index can get reused by new epair instance.
855 	 * Make sure we do not create same etheraddr again.
856 	 */
857 	getcredhostid(curthread->td_ucred, (unsigned long *)&hostid);
858 	if (hostid == 0)
859 		arc4rand(&hostid, sizeof(hostid), 0);
860 
861 	if (ifp->if_index > next_index)
862 		next_index = ifp->if_index;
863 	else
864 		next_index++;
865 
866 	key[0] = (uint32_t)next_index;
867 	key[1] = (uint32_t)(hostid & 0xffffffff);
868 	key[2] = (uint32_t)((hostid >> 32) & 0xfffffffff);
869 	hash = jenkins_hash32(key, 3, 0);
870 
871 	eaddr[0] = 0x02;
872 	memcpy(&eaddr[1], &hash, 4);
873 	eaddr[5] = 0x0a;
874 	ether_ifattach(ifp, eaddr);
875 	sca->if_qflush = ifp->if_qflush;
876 	ifp->if_qflush = epair_qflush;
877 	ifp->if_transmit = epair_transmit;
878 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
879 
880 	/* Swap the name and finish initialization of interface <n>b. */
881 	*dp = 'b';
882 
883 	ifp = scb->ifp;
884 	ifp->if_softc = scb;
885 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
886 	ifp->if_dname = epairname;
887 	ifp->if_dunit = unit;
888 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
889 	ifp->if_capabilities = IFCAP_VLAN_MTU;
890 	ifp->if_capenable = IFCAP_VLAN_MTU;
891 	ifp->if_start = epair_start;
892 	ifp->if_ioctl = epair_ioctl;
893 	ifp->if_init  = epair_init;
894 	if_setsendqlen(ifp, ifqmaxlen);
895 	if_setsendqready(ifp);
896 	/* We need to play some tricks here for the second interface. */
897 	strlcpy(name, epairname, len);
898 
899 	/* Correctly set the name for the cloner list. */
900 	strlcpy(name, scb->ifp->if_xname, len);
901 	epair_clone_add(ifc, scb);
902 
903 	scb->if_qflush = ifp->if_qflush;
904 	ifp->if_qflush = epair_qflush;
905 	ifp->if_transmit = epair_transmit;
906 	ifp->if_baudrate = IF_Gbps(10);	/* arbitrary maximum */
907 
908 	/*
909 	 * Restore name to <n>a as the ifp for this will go into the
910 	 * cloner list for the initial call.
911 	 */
912 	strlcpy(name, sca->ifp->if_xname, len);
913 	DPRINTF("name='%s/%db' created sca=%p scb=%p\n", name, unit, sca, scb);
914 
915 	/* Tell the world, that we are ready to rock. */
916 	sca->ifp->if_drv_flags |= IFF_DRV_RUNNING;
917 	scb->ifp->if_drv_flags |= IFF_DRV_RUNNING;
918 	if_link_state_change(sca->ifp, LINK_STATE_UP);
919 	if_link_state_change(scb->ifp, LINK_STATE_UP);
920 
921 	return (0);
922 }
923 
924 static int
925 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp)
926 {
927 	struct ifnet *oifp;
928 	struct epair_softc *sca, *scb;
929 	int unit, error;
930 
931 	DPRINTF("ifp=%p\n", ifp);
932 
933 	/*
934 	 * In case we called into if_clone_destroyif() ourselves
935 	 * again to remove the second interface, the softc will be
936 	 * NULL. In that case so not do anything but return success.
937 	 */
938 	if (ifp->if_softc == NULL)
939 		return (0);
940 
941 	unit = ifp->if_dunit;
942 	sca = ifp->if_softc;
943 	oifp = sca->oifp;
944 	scb = oifp->if_softc;
945 
946 	DPRINTF("ifp=%p oifp=%p\n", ifp, oifp);
947 	if_link_state_change(ifp, LINK_STATE_DOWN);
948 	if_link_state_change(oifp, LINK_STATE_DOWN);
949 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
950 	oifp->if_drv_flags &= ~IFF_DRV_RUNNING;
951 
952 	/*
953 	 * Get rid of our second half. As the other of the two
954 	 * interfaces may reside in a different vnet, we need to
955 	 * switch before freeing them.
956 	 */
957 	CURVNET_SET_QUIET(oifp->if_vnet);
958 	ether_ifdetach(oifp);
959 	/*
960 	 * Wait for all packets to be dispatched to if_input.
961 	 * The numbers can only go down as the interface is
962 	 * detached so there is no need to use atomics.
963 	 */
964 	DPRINTF("scb refcnt=%u\n", scb->refcount);
965 	EPAIR_REFCOUNT_ASSERT(scb->refcount == 1,
966 	    ("%s: ifp=%p scb->refcount!=1: %d", __func__, oifp, scb->refcount));
967 	oifp->if_softc = NULL;
968 	error = if_clone_destroyif(ifc, oifp);
969 	if (error)
970 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
971 		    __func__, error);
972 	if_free(oifp);
973 	ifmedia_removeall(&scb->media);
974 	free(scb, M_EPAIR);
975 	CURVNET_RESTORE();
976 
977 	ether_ifdetach(ifp);
978 	/*
979 	 * Wait for all packets to be dispatched to if_input.
980 	 */
981 	DPRINTF("sca refcnt=%u\n", sca->refcount);
982 	EPAIR_REFCOUNT_ASSERT(sca->refcount == 1,
983 	    ("%s: ifp=%p sca->refcount!=1: %d", __func__, ifp, sca->refcount));
984 	if_free(ifp);
985 	ifmedia_removeall(&sca->media);
986 	free(sca, M_EPAIR);
987 	ifc_free_unit(ifc, unit);
988 
989 	return (0);
990 }
991 
992 static void
993 vnet_epair_init(const void *unused __unused)
994 {
995 
996 	V_epair_cloner = if_clone_advanced(epairname, 0,
997 	    epair_clone_match, epair_clone_create, epair_clone_destroy);
998 #ifdef VIMAGE
999 	netisr_register_vnet(&epair_nh);
1000 #endif
1001 }
1002 VNET_SYSINIT(vnet_epair_init, SI_SUB_PSEUDO, SI_ORDER_ANY,
1003     vnet_epair_init, NULL);
1004 
1005 static void
1006 vnet_epair_uninit(const void *unused __unused)
1007 {
1008 
1009 #ifdef VIMAGE
1010 	netisr_unregister_vnet(&epair_nh);
1011 #endif
1012 	if_clone_detach(V_epair_cloner);
1013 }
1014 VNET_SYSUNINIT(vnet_epair_uninit, SI_SUB_INIT_IF, SI_ORDER_ANY,
1015     vnet_epair_uninit, NULL);
1016 
1017 static void
1018 epair_uninit(const void *unused __unused)
1019 {
1020 	netisr_unregister(&epair_nh);
1021 	epair_dpcpu_detach();
1022 	if (bootverbose)
1023 		printf("%s unloaded.\n", epairname);
1024 }
1025 SYSUNINIT(epair_uninit, SI_SUB_INIT_IF, SI_ORDER_MIDDLE,
1026     epair_uninit, NULL);
1027 
1028 static int
1029 epair_modevent(module_t mod, int type, void *data)
1030 {
1031 	int qlimit;
1032 
1033 	switch (type) {
1034 	case MOD_LOAD:
1035 		/* For now limit us to one global mutex and one inq. */
1036 		epair_dpcpu_init();
1037 		epair_nh.nh_qlimit = 42 * ifqmaxlen; /* 42 shall be the number. */
1038 		if (TUNABLE_INT_FETCH("net.link.epair.netisr_maxqlen", &qlimit))
1039 		    epair_nh.nh_qlimit = qlimit;
1040 		netisr_register(&epair_nh);
1041 		if (bootverbose)
1042 			printf("%s initialized.\n", epairname);
1043 		break;
1044 	case MOD_UNLOAD:
1045 		/* Handled in epair_uninit() */
1046 		break;
1047 	default:
1048 		return (EOPNOTSUPP);
1049 	}
1050 	return (0);
1051 }
1052 
1053 static moduledata_t epair_mod = {
1054 	"if_epair",
1055 	epair_modevent,
1056 	0
1057 };
1058 
1059 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_MIDDLE);
1060 MODULE_VERSION(if_epair, 1);
1061