xref: /freebsd/sys/net/if_epair.c (revision 724b4bfdf1306e4f2c451b6d146fe0fe0353b2c8)
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_media.h>
70 #include <net/if_var.h>
71 #include <net/if_types.h>
72 #include <net/netisr.h>
73 #include <net/vnet.h>
74 
75 SYSCTL_DECL(_net_link);
76 static SYSCTL_NODE(_net_link, OID_AUTO, epair, CTLFLAG_RW, 0, "epair sysctl");
77 
78 #ifdef EPAIR_DEBUG
79 static int epair_debug = 0;
80 SYSCTL_INT(_net_link_epair, OID_AUTO, epair_debug, CTLFLAG_RW,
81     &epair_debug, 0, "if_epair(4) debugging.");
82 #define	DPRINTF(fmt, arg...)						\
83 	if (epair_debug)						\
84 		printf("[%s:%d] " fmt, __func__, __LINE__, ##arg)
85 #else
86 #define	DPRINTF(fmt, arg...)
87 #endif
88 
89 static void epair_nh_sintr(struct mbuf *);
90 static struct mbuf *epair_nh_m2cpuid(struct mbuf *, uintptr_t, u_int *);
91 static void epair_nh_drainedcpu(u_int);
92 
93 static void epair_start_locked(struct ifnet *);
94 static int epair_media_change(struct ifnet *);
95 static void epair_media_status(struct ifnet *, struct ifmediareq *);
96 
97 static int epair_clone_match(struct if_clone *, const char *);
98 static int epair_clone_create(struct if_clone *, char *, size_t, caddr_t);
99 static int epair_clone_destroy(struct if_clone *, struct ifnet *);
100 
101 static const char epairname[] = "epair";
102 
103 /* Netisr realted definitions and sysctl. */
104 static struct netisr_handler epair_nh = {
105 	.nh_name	= epairname,
106 	.nh_proto	= NETISR_EPAIR,
107 	.nh_policy	= NETISR_POLICY_CPU,
108 	.nh_handler	= epair_nh_sintr,
109 	.nh_m2cpuid	= epair_nh_m2cpuid,
110 	.nh_drainedcpu	= epair_nh_drainedcpu,
111 };
112 
113 static int
114 sysctl_epair_netisr_maxqlen(SYSCTL_HANDLER_ARGS)
115 {
116 	int error, qlimit;
117 
118 	netisr_getqlimit(&epair_nh, &qlimit);
119 	error = sysctl_handle_int(oidp, &qlimit, 0, req);
120 	if (error || !req->newptr)
121 		return (error);
122 	if (qlimit < 1)
123 		return (EINVAL);
124 	return (netisr_setqlimit(&epair_nh, qlimit));
125 }
126 SYSCTL_PROC(_net_link_epair, OID_AUTO, netisr_maxqlen, CTLTYPE_INT|CTLFLAG_RW,
127     0, 0, sysctl_epair_netisr_maxqlen, "I",
128     "Maximum if_epair(4) netisr \"hw\" queue length");
129 
130 struct epair_softc {
131 	struct ifnet	*ifp;		/* This ifp. */
132 	struct ifnet	*oifp;		/* other ifp of pair. */
133 	struct ifmedia	media;		/* Media config (fake). */
134 	u_int		refcount;	/* # of mbufs in flight. */
135 	u_int		cpuid;		/* CPU ID assigned upon creation. */
136 	void		(*if_qflush)(struct ifnet *);
137 					/* Original if_qflush routine. */
138 };
139 
140 /*
141  * Per-CPU list of ifps with data in the ifq that needs to be flushed
142  * to the netisr ``hw'' queue before we allow any further direct queuing
143  * to the ``hw'' queue.
144  */
145 struct epair_ifp_drain {
146 	STAILQ_ENTRY(epair_ifp_drain)	ifp_next;
147 	struct ifnet			*ifp;
148 };
149 STAILQ_HEAD(eid_list, epair_ifp_drain);
150 
151 #define	EPAIR_LOCK_INIT(dpcpu)		mtx_init(&(dpcpu)->if_epair_mtx, \
152 					    "if_epair", NULL, MTX_DEF)
153 #define	EPAIR_LOCK_DESTROY(dpcpu)	mtx_destroy(&(dpcpu)->if_epair_mtx)
154 #define	EPAIR_LOCK_ASSERT(dpcpu)	mtx_assert(&(dpcpu)->if_epair_mtx, \
155 					    MA_OWNED)
156 #define	EPAIR_LOCK(dpcpu)		mtx_lock(&(dpcpu)->if_epair_mtx)
157 #define	EPAIR_UNLOCK(dpcpu)		mtx_unlock(&(dpcpu)->if_epair_mtx)
158 
159 #ifdef INVARIANTS
160 #define	EPAIR_REFCOUNT_INIT(r, v)	refcount_init((r), (v))
161 #define	EPAIR_REFCOUNT_AQUIRE(r)	refcount_acquire((r))
162 #define	EPAIR_REFCOUNT_RELEASE(r)	refcount_release((r))
163 #define	EPAIR_REFCOUNT_ASSERT(a, p)	KASSERT(a, p)
164 #else
165 #define	EPAIR_REFCOUNT_INIT(r, v)
166 #define	EPAIR_REFCOUNT_AQUIRE(r)
167 #define	EPAIR_REFCOUNT_RELEASE(r)
168 #define	EPAIR_REFCOUNT_ASSERT(a, p)
169 #endif
170 
171 static MALLOC_DEFINE(M_EPAIR, epairname,
172     "Pair of virtual cross-over connected Ethernet-like interfaces");
173 
174 static struct if_clone *epair_cloner;
175 
176 /*
177  * DPCPU area and functions.
178  */
179 struct epair_dpcpu {
180 	struct mtx	if_epair_mtx;		/* Per-CPU locking. */
181 	int		epair_drv_flags;	/* Per-CPU ``hw'' drv flags. */
182 	struct eid_list	epair_ifp_drain_list;	/* Per-CPU list of ifps with
183 						 * data in the ifq. */
184 };
185 DPCPU_DEFINE(struct epair_dpcpu, epair_dpcpu);
186 
187 static void
188 epair_dpcpu_init(void)
189 {
190 	struct epair_dpcpu *epair_dpcpu;
191 	struct eid_list *s;
192 	u_int cpuid;
193 
194 	CPU_FOREACH(cpuid) {
195 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
196 
197 		/* Initialize per-cpu lock. */
198 		EPAIR_LOCK_INIT(epair_dpcpu);
199 
200 		/* Driver flags are per-cpu as are our netisr "hw" queues. */
201 		epair_dpcpu->epair_drv_flags = 0;
202 
203 		/*
204 		 * Initialize per-cpu drain list.
205 		 * Manually do what STAILQ_HEAD_INITIALIZER would do.
206 		 */
207 		s = &epair_dpcpu->epair_ifp_drain_list;
208 		s->stqh_first = NULL;
209 		s->stqh_last = &s->stqh_first;
210 	}
211 }
212 
213 static void
214 epair_dpcpu_detach(void)
215 {
216 	struct epair_dpcpu *epair_dpcpu;
217 	u_int cpuid;
218 
219 	CPU_FOREACH(cpuid) {
220 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
221 
222 		/* Destroy per-cpu lock. */
223 		EPAIR_LOCK_DESTROY(epair_dpcpu);
224 	}
225 }
226 
227 /*
228  * Helper functions.
229  */
230 static u_int
231 cpuid_from_ifp(struct ifnet *ifp)
232 {
233 	struct epair_softc *sc;
234 
235 	if (ifp == NULL)
236 		return (0);
237 	sc = ifp->if_softc;
238 
239 	return (sc->cpuid);
240 }
241 
242 /*
243  * Netisr handler functions.
244  */
245 static void
246 epair_nh_sintr(struct mbuf *m)
247 {
248 	struct ifnet *ifp;
249 	struct epair_softc *sc;
250 
251 	ifp = m->m_pkthdr.rcvif;
252 	(*ifp->if_input)(ifp, m);
253 	sc = ifp->if_softc;
254 	EPAIR_REFCOUNT_RELEASE(&sc->refcount);
255 	EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
256 	    ("%s: ifp=%p sc->refcount not >= 1: %d",
257 	    __func__, ifp, sc->refcount));
258 	DPRINTF("ifp=%p refcount=%u\n", ifp, sc->refcount);
259 }
260 
261 static struct mbuf *
262 epair_nh_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
263 {
264 
265 	*cpuid = cpuid_from_ifp(m->m_pkthdr.rcvif);
266 
267 	return (m);
268 }
269 
270 static void
271 epair_nh_drainedcpu(u_int cpuid)
272 {
273 	struct epair_dpcpu *epair_dpcpu;
274 	struct epair_ifp_drain *elm, *tvar;
275 	struct ifnet *ifp;
276 
277 	epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
278 	EPAIR_LOCK(epair_dpcpu);
279 	/*
280 	 * Assume our "hw" queue and possibly ifq will be emptied
281 	 * again. In case we will overflow the "hw" queue while
282 	 * draining, epair_start_locked will set IFF_DRV_OACTIVE
283 	 * again and we will stop and return.
284 	 */
285 	STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
286 	    ifp_next, tvar) {
287 		ifp = elm->ifp;
288 		epair_dpcpu->epair_drv_flags &= ~IFF_DRV_OACTIVE;
289 		ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
290 		epair_start_locked(ifp);
291 
292 		IFQ_LOCK(&ifp->if_snd);
293 		if (IFQ_IS_EMPTY(&ifp->if_snd)) {
294 			struct epair_softc *sc;
295 
296 			STAILQ_REMOVE(&epair_dpcpu->epair_ifp_drain_list,
297 			    elm, epair_ifp_drain, ifp_next);
298 			/* The cached ifp goes off the list. */
299 			sc = ifp->if_softc;
300 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
301 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
302 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
303 			    __func__, ifp, sc->refcount));
304 			free(elm, M_EPAIR);
305 		}
306 		IFQ_UNLOCK(&ifp->if_snd);
307 
308 		if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) != 0) {
309 			/* Our "hw"q overflew again. */
310 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
311 			DPRINTF("hw queue length overflow at %u\n",
312 			    epair_nh.nh_qlimit);
313 			break;
314 		}
315 	}
316 	EPAIR_UNLOCK(epair_dpcpu);
317 }
318 
319 /*
320  * Network interface (`if') related functions.
321  */
322 static void
323 epair_remove_ifp_from_draining(struct ifnet *ifp)
324 {
325 	struct epair_dpcpu *epair_dpcpu;
326 	struct epair_ifp_drain *elm, *tvar;
327 	u_int cpuid;
328 
329 	CPU_FOREACH(cpuid) {
330 		epair_dpcpu = DPCPU_ID_PTR(cpuid, epair_dpcpu);
331 		EPAIR_LOCK(epair_dpcpu);
332 		STAILQ_FOREACH_SAFE(elm, &epair_dpcpu->epair_ifp_drain_list,
333 		    ifp_next, tvar) {
334 			if (ifp == elm->ifp) {
335 				struct epair_softc *sc;
336 
337 				STAILQ_REMOVE(
338 				    &epair_dpcpu->epair_ifp_drain_list, elm,
339 				    epair_ifp_drain, ifp_next);
340 				/* The cached ifp goes off the list. */
341 				sc = ifp->if_softc;
342 				EPAIR_REFCOUNT_RELEASE(&sc->refcount);
343 				EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
344 				    ("%s: ifp=%p sc->refcount not >= 1: %d",
345 				    __func__, ifp, sc->refcount));
346 				free(elm, M_EPAIR);
347 			}
348 		}
349 		EPAIR_UNLOCK(epair_dpcpu);
350 	}
351 }
352 
353 static int
354 epair_add_ifp_for_draining(struct ifnet *ifp)
355 {
356 	struct epair_dpcpu *epair_dpcpu;
357 	struct epair_softc *sc;
358 	struct epair_ifp_drain *elm = NULL;
359 
360 	sc = ifp->if_softc;
361 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
362 	EPAIR_LOCK_ASSERT(epair_dpcpu);
363 	STAILQ_FOREACH(elm, &epair_dpcpu->epair_ifp_drain_list, ifp_next)
364 		if (elm->ifp == ifp)
365 			break;
366 	/* If the ifp is there already, return success. */
367 	if (elm != NULL)
368 		return (0);
369 
370 	elm = malloc(sizeof(struct epair_ifp_drain), M_EPAIR, M_NOWAIT|M_ZERO);
371 	if (elm == NULL)
372 		return (ENOMEM);
373 
374 	elm->ifp = ifp;
375 	/* Add a reference for the ifp pointer on the list. */
376 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
377 	STAILQ_INSERT_TAIL(&epair_dpcpu->epair_ifp_drain_list, elm, ifp_next);
378 
379 	return (0);
380 }
381 
382 static void
383 epair_start_locked(struct ifnet *ifp)
384 {
385 	struct epair_dpcpu *epair_dpcpu;
386 	struct mbuf *m;
387 	struct epair_softc *sc;
388 	struct ifnet *oifp;
389 	int error;
390 
391 	DPRINTF("ifp=%p\n", ifp);
392 	sc = ifp->if_softc;
393 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
394 	EPAIR_LOCK_ASSERT(epair_dpcpu);
395 
396 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
397 		return;
398 	if ((ifp->if_flags & IFF_UP) == 0)
399 		return;
400 
401 	/*
402 	 * We get patckets here from ether_output via if_handoff()
403 	 * and ned to put them into the input queue of the oifp
404 	 * and call oifp->if_input() via netisr/epair_sintr().
405 	 */
406 	oifp = sc->oifp;
407 	sc = oifp->if_softc;
408 	for (;;) {
409 		IFQ_DEQUEUE(&ifp->if_snd, m);
410 		if (m == NULL)
411 			break;
412 		BPF_MTAP(ifp, m);
413 
414 		/*
415 		 * In case the outgoing interface is not usable,
416 		 * drop the packet.
417 		 */
418 		if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
419 		    (oifp->if_flags & IFF_UP) ==0) {
420 			ifp->if_oerrors++;
421 			m_freem(m);
422 			continue;
423 		}
424 		DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
425 
426 		/*
427 		 * Add a reference so the interface cannot go while the
428 		 * packet is in transit as we rely on rcvif to stay valid.
429 		 */
430 		EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
431 		m->m_pkthdr.rcvif = oifp;
432 		CURVNET_SET_QUIET(oifp->if_vnet);
433 		error = netisr_queue(NETISR_EPAIR, m);
434 		CURVNET_RESTORE();
435 		if (!error) {
436 			ifp->if_opackets++;
437 			/* Someone else received the packet. */
438 			oifp->if_ipackets++;
439 		} else {
440 			/* The packet was freed already. */
441 			epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
442 			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
443 			(void) epair_add_ifp_for_draining(ifp);
444 			ifp->if_oerrors++;
445 			EPAIR_REFCOUNT_RELEASE(&sc->refcount);
446 			EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
447 			    ("%s: ifp=%p sc->refcount not >= 1: %d",
448 			    __func__, oifp, sc->refcount));
449 		}
450 	}
451 }
452 
453 static void
454 epair_start(struct ifnet *ifp)
455 {
456 	struct epair_dpcpu *epair_dpcpu;
457 
458 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
459 	EPAIR_LOCK(epair_dpcpu);
460 	epair_start_locked(ifp);
461 	EPAIR_UNLOCK(epair_dpcpu);
462 }
463 
464 static int
465 epair_transmit_locked(struct ifnet *ifp, struct mbuf *m)
466 {
467 	struct epair_dpcpu *epair_dpcpu;
468 	struct epair_softc *sc;
469 	struct ifnet *oifp;
470 	int error, len;
471 	short mflags;
472 
473 	DPRINTF("ifp=%p m=%p\n", ifp, m);
474 	sc = ifp->if_softc;
475 	epair_dpcpu = DPCPU_ID_PTR(sc->cpuid, epair_dpcpu);
476 	EPAIR_LOCK_ASSERT(epair_dpcpu);
477 
478 	if (m == NULL)
479 		return (0);
480 
481 	/*
482 	 * We are not going to use the interface en/dequeue mechanism
483 	 * on the TX side. We are called from ether_output_frame()
484 	 * and will put the packet into the incoming queue of the
485 	 * other interface of our pair via the netsir.
486 	 */
487 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
488 		m_freem(m);
489 		return (ENXIO);
490 	}
491 	if ((ifp->if_flags & IFF_UP) == 0) {
492 		m_freem(m);
493 		return (ENETDOWN);
494 	}
495 
496 	BPF_MTAP(ifp, m);
497 
498 	/*
499 	 * In case the outgoing interface is not usable,
500 	 * drop the packet.
501 	 */
502 	oifp = sc->oifp;
503 	if ((oifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
504 	    (oifp->if_flags & IFF_UP) ==0) {
505 		ifp->if_oerrors++;
506 		m_freem(m);
507 		return (0);
508 	}
509 	len = m->m_pkthdr.len;
510 	mflags = m->m_flags;
511 	DPRINTF("packet %s -> %s\n", ifp->if_xname, oifp->if_xname);
512 
513 #ifdef ALTQ
514 	/* Support ALTQ via the clasic if_start() path. */
515 	IF_LOCK(&ifp->if_snd);
516 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
517 		ALTQ_ENQUEUE(&ifp->if_snd, m, NULL, error);
518 		if (error)
519 			ifp->if_snd.ifq_drops++;
520 		IF_UNLOCK(&ifp->if_snd);
521 		if (!error) {
522 			ifp->if_obytes += len;
523 			if (mflags & (M_BCAST|M_MCAST))
524 				ifp->if_omcasts++;
525 
526 			if ((ifp->if_drv_flags & IFF_DRV_OACTIVE) == 0)
527 				epair_start_locked(ifp);
528 			else
529 				(void)epair_add_ifp_for_draining(ifp);
530 		}
531 		return (error);
532 	}
533 	IF_UNLOCK(&ifp->if_snd);
534 #endif
535 
536 	if ((epair_dpcpu->epair_drv_flags & IFF_DRV_OACTIVE) != 0) {
537 		/*
538 		 * Our hardware queue is full, try to fall back
539 		 * queuing to the ifq but do not call ifp->if_start.
540 		 * Either we are lucky or the packet is gone.
541 		 */
542 		IFQ_ENQUEUE(&ifp->if_snd, m, error);
543 		if (!error)
544 			(void)epair_add_ifp_for_draining(ifp);
545 		return (error);
546 	}
547 	sc = oifp->if_softc;
548 	/*
549 	 * Add a reference so the interface cannot go while the
550 	 * packet is in transit as we rely on rcvif to stay valid.
551 	 */
552 	EPAIR_REFCOUNT_AQUIRE(&sc->refcount);
553 	m->m_pkthdr.rcvif = oifp;
554 	CURVNET_SET_QUIET(oifp->if_vnet);
555 	error = netisr_queue(NETISR_EPAIR, m);
556 	CURVNET_RESTORE();
557 	if (!error) {
558 		ifp->if_opackets++;
559 		/*
560 		 * IFQ_HANDOFF_ADJ/ip_handoff() update statistics,
561 		 * but as we bypass all this we have to duplicate
562 		 * the logic another time.
563 		 */
564 		ifp->if_obytes += len;
565 		if (mflags & (M_BCAST|M_MCAST))
566 			ifp->if_omcasts++;
567 		/* Someone else received the packet. */
568 		oifp->if_ipackets++;
569 	} else {
570 		/* The packet was freed already. */
571 		epair_dpcpu->epair_drv_flags |= IFF_DRV_OACTIVE;
572 		ifp->if_drv_flags |= IFF_DRV_OACTIVE;
573 		ifp->if_oerrors++;
574 		EPAIR_REFCOUNT_RELEASE(&sc->refcount);
575 		EPAIR_REFCOUNT_ASSERT((int)sc->refcount >= 1,
576 		    ("%s: ifp=%p sc->refcount not >= 1: %d",
577 		    __func__, oifp, sc->refcount));
578 	}
579 
580 	return (error);
581 }
582 
583 static int
584 epair_transmit(struct ifnet *ifp, struct mbuf *m)
585 {
586 	struct epair_dpcpu *epair_dpcpu;
587 	int error;
588 
589 	epair_dpcpu = DPCPU_ID_PTR(cpuid_from_ifp(ifp), epair_dpcpu);
590 	EPAIR_LOCK(epair_dpcpu);
591 	error = epair_transmit_locked(ifp, m);
592 	EPAIR_UNLOCK(epair_dpcpu);
593 	return (error);
594 }
595 
596 static void
597 epair_qflush(struct ifnet *ifp)
598 {
599 	struct epair_softc *sc;
600 
601 	sc = ifp->if_softc;
602 	KASSERT(sc != NULL, ("%s: ifp=%p, epair_softc gone? sc=%p\n",
603 	    __func__, ifp, sc));
604 	/*
605 	 * Remove this ifp from all backpointer lists. The interface will not
606 	 * usable for flushing anyway nor should it have anything to flush
607 	 * after if_qflush().
608 	 */
609 	epair_remove_ifp_from_draining(ifp);
610 
611 	if (sc->if_qflush)
612 		sc->if_qflush(ifp);
613 }
614 
615 static int
616 epair_media_change(struct ifnet *ifp __unused)
617 {
618 
619 	/* Do nothing. */
620 	return (0);
621 }
622 
623 static void
624 epair_media_status(struct ifnet *ifp __unused, struct ifmediareq *imr)
625 {
626 
627 	imr->ifm_status = IFM_AVALID | IFM_ACTIVE;
628 	imr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
629 }
630 
631 static int
632 epair_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
633 {
634 	struct epair_softc *sc;
635 	struct ifreq *ifr;
636 	int error;
637 
638 	ifr = (struct ifreq *)data;
639 	switch (cmd) {
640 	case SIOCSIFFLAGS:
641 	case SIOCADDMULTI:
642 	case SIOCDELMULTI:
643 		error = 0;
644 		break;
645 
646 	case SIOCSIFMEDIA:
647 	case SIOCGIFMEDIA:
648 		sc = ifp->if_softc;
649 		error = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
650 		break;
651 
652 	case SIOCSIFMTU:
653 		/* We basically allow all kinds of MTUs. */
654 		ifp->if_mtu = ifr->ifr_mtu;
655 		error = 0;
656 		break;
657 
658 	default:
659 		/* Let the common ethernet handler process this. */
660 		error = ether_ioctl(ifp, cmd, data);
661 		break;
662 	}
663 
664 	return (error);
665 }
666 
667 static void
668 epair_init(void *dummy __unused)
669 {
670 }
671 
672 
673 /*
674  * Interface cloning functions.
675  * We use our private ones so that we can create/destroy our secondary
676  * device along with the primary one.
677  */
678 static int
679 epair_clone_match(struct if_clone *ifc, const char *name)
680 {
681 	const char *cp;
682 
683 	DPRINTF("name='%s'\n", name);
684 
685 	/*
686 	 * Our base name is epair.
687 	 * Our interfaces will be named epair<n>[ab].
688 	 * So accept anything of the following list:
689 	 * - epair
690 	 * - epair<n>
691 	 * but not the epair<n>[ab] versions.
692 	 */
693 	if (strncmp(epairname, name, sizeof(epairname)-1) != 0)
694 		return (0);
695 
696 	for (cp = name + sizeof(epairname) - 1; *cp != '\0'; cp++) {
697 		if (*cp < '0' || *cp > '9')
698 			return (0);
699 	}
700 
701 	return (1);
702 }
703 
704 static int
705 epair_clone_create(struct if_clone *ifc, char *name, size_t len, caddr_t params)
706 {
707 	struct epair_softc *sca, *scb;
708 	struct ifnet *ifp;
709 	char *dp;
710 	int error, unit, wildcard;
711 	uint8_t eaddr[ETHER_ADDR_LEN];	/* 00:00:00:00:00:00 */
712 
713 	/*
714 	 * We are abusing params to create our second interface.
715 	 * Actually we already created it and called if_clone_create()
716 	 * for it to do the official insertion procedure the moment we knew
717 	 * it cannot fail anymore. So just do attach it here.
718 	 */
719 	if (params) {
720 		scb = (struct epair_softc *)params;
721 		ifp = scb->ifp;
722 		/* Assign a hopefully unique, locally administered etheraddr. */
723 		eaddr[0] = 0x02;
724 		eaddr[3] = (ifp->if_index >> 8) & 0xff;
725 		eaddr[4] = ifp->if_index & 0xff;
726 		eaddr[5] = 0x0b;
727 		ether_ifattach(ifp, eaddr);
728 		/* Correctly set the name for the cloner list. */
729 		strlcpy(name, scb->ifp->if_xname, len);
730 		return (0);
731 	}
732 
733 	/* Try to see if a special unit was requested. */
734 	error = ifc_name2unit(name, &unit);
735 	if (error != 0)
736 		return (error);
737 	wildcard = (unit < 0);
738 
739 	error = ifc_alloc_unit(ifc, &unit);
740 	if (error != 0)
741 		return (error);
742 
743 	/*
744 	 * If no unit had been given, we need to adjust the ifName.
745 	 * Also make sure there is space for our extra [ab] suffix.
746 	 */
747 	for (dp = name; *dp != '\0'; dp++);
748 	if (wildcard) {
749 		error = snprintf(dp, len - (dp - name), "%d", unit);
750 		if (error > len - (dp - name) - 1) {
751 			/* ifName too long. */
752 			ifc_free_unit(ifc, unit);
753 			return (ENOSPC);
754 		}
755 		dp += error;
756 	}
757 	if (len - (dp - name) - 1 < 1) {
758 		/* No space left for our [ab] suffix. */
759 		ifc_free_unit(ifc, unit);
760 		return (ENOSPC);
761 	}
762 	*dp = 'a';
763 	/* Must not change dp so we can replace 'a' by 'b' later. */
764 	*(dp+1) = '\0';
765 
766 	/* Allocate memory for both [ab] interfaces */
767 	sca = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
768 	EPAIR_REFCOUNT_INIT(&sca->refcount, 1);
769 	sca->ifp = if_alloc(IFT_ETHER);
770 	if (sca->ifp == NULL) {
771 		free(sca, M_EPAIR);
772 		ifc_free_unit(ifc, unit);
773 		return (ENOSPC);
774 	}
775 
776 	scb = malloc(sizeof(struct epair_softc), M_EPAIR, M_WAITOK | M_ZERO);
777 	EPAIR_REFCOUNT_INIT(&scb->refcount, 1);
778 	scb->ifp = if_alloc(IFT_ETHER);
779 	if (scb->ifp == NULL) {
780 		free(scb, M_EPAIR);
781 		if_free(sca->ifp);
782 		free(sca, M_EPAIR);
783 		ifc_free_unit(ifc, unit);
784 		return (ENOSPC);
785 	}
786 
787 	/*
788 	 * Cross-reference the interfaces so we will be able to free both.
789 	 */
790 	sca->oifp = scb->ifp;
791 	scb->oifp = sca->ifp;
792 
793 	/*
794 	 * Calculate the cpuid for netisr queueing based on the
795 	 * ifIndex of the interfaces. As long as we cannot configure
796 	 * this or use cpuset information easily we cannot guarantee
797 	 * cache locality but we can at least allow parallelism.
798 	 */
799 	sca->cpuid =
800 	    netisr_get_cpuid(sca->ifp->if_index % netisr_get_cpucount());
801 	scb->cpuid =
802 	    netisr_get_cpuid(scb->ifp->if_index % netisr_get_cpucount());
803 
804 	/* Finish initialization of interface <n>a. */
805 	ifp = sca->ifp;
806 	ifp->if_softc = sca;
807 	strlcpy(ifp->if_xname, name, IFNAMSIZ);
808 	ifp->if_dname = epairname;
809 	ifp->if_dunit = unit;
810 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
811 	ifp->if_capabilities = IFCAP_VLAN_MTU;
812 	ifp->if_capenable = IFCAP_VLAN_MTU;
813 	ifp->if_start = epair_start;
814 	ifp->if_ioctl = epair_ioctl;
815 	ifp->if_init  = epair_init;
816 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
817 	/* Assign a hopefully unique, locally administered etheraddr. */
818 	eaddr[0] = 0x02;
819 	eaddr[3] = (ifp->if_index >> 8) & 0xff;
820 	eaddr[4] = ifp->if_index & 0xff;
821 	eaddr[5] = 0x0a;
822 	ether_ifattach(ifp, eaddr);
823 	sca->if_qflush = ifp->if_qflush;
824 	ifp->if_qflush = epair_qflush;
825 	ifp->if_transmit = epair_transmit;
826 	if_initbaudrate(ifp, IF_Gbps(10));	/* arbitrary maximum */
827 
828 	/* Swap the name and finish initialization of interface <n>b. */
829 	*dp = 'b';
830 
831 	ifp = scb->ifp;
832 	ifp->if_softc = scb;
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 	ifp->if_snd.ifq_maxlen = ifqmaxlen;
843 	/* We need to play some tricks here for the second interface. */
844 	strlcpy(name, epairname, len);
845 	error = if_clone_create(name, len, (caddr_t)scb);
846 	if (error)
847 		panic("%s: if_clone_create() for our 2nd iface failed: %d",
848 		    __func__, error);
849 	scb->if_qflush = ifp->if_qflush;
850 	ifp->if_qflush = epair_qflush;
851 	ifp->if_transmit = epair_transmit;
852 	if_initbaudrate(ifp, IF_Gbps(10));	/* arbitrary maximum */
853 
854 	/*
855 	 * Restore name to <n>a as the ifp for this will go into the
856 	 * cloner list for the initial call.
857 	 */
858 	strlcpy(name, sca->ifp->if_xname, len);
859 	DPRINTF("name='%s/%db' created sca=%p scb=%p\n", name, unit, sca, scb);
860 
861 	/* Initialise pseudo media types. */
862 	ifmedia_init(&sca->media, 0, epair_media_change, epair_media_status);
863 	ifmedia_add(&sca->media, IFM_ETHER | IFM_10G_T, 0, NULL);
864 	ifmedia_set(&sca->media, IFM_ETHER | IFM_10G_T);
865 	ifmedia_init(&scb->media, 0, epair_media_change, epair_media_status);
866 	ifmedia_add(&scb->media, IFM_ETHER | IFM_10G_T, 0, NULL);
867 	ifmedia_set(&scb->media, IFM_ETHER | IFM_10G_T);
868 
869 	/* Tell the world, that we are ready to rock. */
870 	sca->ifp->if_drv_flags |= IFF_DRV_RUNNING;
871 	scb->ifp->if_drv_flags |= IFF_DRV_RUNNING;
872 	if_link_state_change(sca->ifp, LINK_STATE_UP);
873 	if_link_state_change(scb->ifp, LINK_STATE_UP);
874 
875 	return (0);
876 }
877 
878 static int
879 epair_clone_destroy(struct if_clone *ifc, struct ifnet *ifp)
880 {
881 	struct ifnet *oifp;
882 	struct epair_softc *sca, *scb;
883 	int unit, error;
884 
885 	DPRINTF("ifp=%p\n", ifp);
886 
887 	/*
888 	 * In case we called into if_clone_destroyif() ourselves
889 	 * again to remove the second interface, the softc will be
890 	 * NULL. In that case so not do anything but return success.
891 	 */
892 	if (ifp->if_softc == NULL)
893 		return (0);
894 
895 	unit = ifp->if_dunit;
896 	sca = ifp->if_softc;
897 	oifp = sca->oifp;
898 	scb = oifp->if_softc;
899 
900 	DPRINTF("ifp=%p oifp=%p\n", ifp, oifp);
901 	if_link_state_change(ifp, LINK_STATE_DOWN);
902 	if_link_state_change(oifp, LINK_STATE_DOWN);
903 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
904 	oifp->if_drv_flags &= ~IFF_DRV_RUNNING;
905 
906 	/*
907 	 * Get rid of our second half. As the other of the two
908 	 * interfaces may reside in a different vnet, we need to
909 	 * switch before freeing them.
910 	 */
911 	CURVNET_SET_QUIET(oifp->if_vnet);
912 	ether_ifdetach(oifp);
913 	/*
914 	 * Wait for all packets to be dispatched to if_input.
915 	 * The numbers can only go down as the interface is
916 	 * detached so there is no need to use atomics.
917 	 */
918 	DPRINTF("scb refcnt=%u\n", scb->refcount);
919 	EPAIR_REFCOUNT_ASSERT(scb->refcount == 1,
920 	    ("%s: ifp=%p scb->refcount!=1: %d", __func__, oifp, scb->refcount));
921 	oifp->if_softc = NULL;
922 	error = if_clone_destroyif(ifc, oifp);
923 	if (error)
924 		panic("%s: if_clone_destroyif() for our 2nd iface failed: %d",
925 		    __func__, error);
926 	if_free(oifp);
927 	ifmedia_removeall(&scb->media);
928 	free(scb, M_EPAIR);
929 	CURVNET_RESTORE();
930 
931 	ether_ifdetach(ifp);
932 	/*
933 	 * Wait for all packets to be dispatched to if_input.
934 	 */
935 	DPRINTF("sca refcnt=%u\n", sca->refcount);
936 	EPAIR_REFCOUNT_ASSERT(sca->refcount == 1,
937 	    ("%s: ifp=%p sca->refcount!=1: %d", __func__, ifp, sca->refcount));
938 	if_free(ifp);
939 	ifmedia_removeall(&sca->media);
940 	free(sca, M_EPAIR);
941 	ifc_free_unit(ifc, unit);
942 
943 	return (0);
944 }
945 
946 static int
947 epair_modevent(module_t mod, int type, void *data)
948 {
949 	int qlimit;
950 
951 	switch (type) {
952 	case MOD_LOAD:
953 		/* For now limit us to one global mutex and one inq. */
954 		epair_dpcpu_init();
955 		epair_nh.nh_qlimit = 42 * ifqmaxlen; /* 42 shall be the number. */
956 		if (TUNABLE_INT_FETCH("net.link.epair.netisr_maxqlen", &qlimit))
957 		    epair_nh.nh_qlimit = qlimit;
958 		netisr_register(&epair_nh);
959 		epair_cloner = if_clone_advanced(epairname, 0,
960 		    epair_clone_match, epair_clone_create, epair_clone_destroy);
961 		if (bootverbose)
962 			printf("%s initialized.\n", epairname);
963 		break;
964 	case MOD_UNLOAD:
965 		if_clone_detach(epair_cloner);
966 		netisr_unregister(&epair_nh);
967 		epair_dpcpu_detach();
968 		if (bootverbose)
969 			printf("%s unloaded.\n", epairname);
970 		break;
971 	default:
972 		return (EOPNOTSUPP);
973 	}
974 	return (0);
975 }
976 
977 static moduledata_t epair_mod = {
978 	"if_epair",
979 	epair_modevent,
980 	0
981 };
982 
983 DECLARE_MODULE(if_epair, epair_mod, SI_SUB_PSEUDO, SI_ORDER_ANY);
984 MODULE_VERSION(if_epair, 1);
985