xref: /netbsd-src/sys/rump/net/lib/libshmif/if_shmem.c (revision 46f5119e40af2e51998f686b2fdcc76b5488f7f3)
1 /*	$NetBSD: if_shmem.c,v 1.39 2011/03/21 16:41:09 pooka Exp $	*/
2 
3 /*
4  * Copyright (c) 2009, 2010 Antti Kantee.  All Rights Reserved.
5  *
6  * Development of this software was supported by The Nokia Foundation.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
18  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
19  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
20  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __KERNEL_RCSID(0, "$NetBSD: if_shmem.c,v 1.39 2011/03/21 16:41:09 pooka Exp $");
32 
33 #include <sys/param.h>
34 #include <sys/atomic.h>
35 #include <sys/fcntl.h>
36 #include <sys/kmem.h>
37 #include <sys/kthread.h>
38 #include <sys/lock.h>
39 #include <sys/vmem.h>
40 
41 #include <net/bpf.h>
42 #include <net/if.h>
43 #include <net/if_dl.h>
44 #include <net/if_ether.h>
45 
46 #include <netinet/in.h>
47 #include <netinet/in_var.h>
48 
49 #include <rump/rump.h>
50 #include <rump/rumpuser.h>
51 
52 #include "rump_private.h"
53 #include "rump_net_private.h"
54 
55 static int shmif_clone(struct if_clone *, int);
56 static int shmif_unclone(struct ifnet *);
57 
58 struct if_clone shmif_cloner =
59     IF_CLONE_INITIALIZER("shmif", shmif_clone, shmif_unclone);
60 
61 /*
62  * Do r/w prefault for backend pages when attaching the interface.
63  * At least logically thinking improves performance (although no
64  * mlocking is done, so they might go away).
65  */
66 #define PREFAULT_RW
67 
68 /*
69  * A virtual ethernet interface which uses shared memory from a
70  * memory mapped file as the bus.
71  */
72 
73 static int	shmif_init(struct ifnet *);
74 static int	shmif_ioctl(struct ifnet *, u_long, void *);
75 static void	shmif_start(struct ifnet *);
76 static void	shmif_stop(struct ifnet *, int);
77 
78 #include "shmifvar.h"
79 
80 struct shmif_sc {
81 	struct ethercom sc_ec;
82 	struct shmif_mem *sc_busmem;
83 	int sc_memfd;
84 	int sc_kq;
85 	int sc_unit;
86 
87 	char *sc_backfile;
88 	size_t sc_backfilelen;
89 
90 	uint64_t sc_devgen;
91 	uint32_t sc_nextpacket;
92 
93 	kmutex_t sc_mtx;
94 	kcondvar_t sc_cv;
95 
96 	struct lwp *sc_rcvl;
97 	bool sc_dying;
98 };
99 
100 static const uint32_t busversion = SHMIF_VERSION;
101 
102 static void shmif_rcv(void *);
103 
104 #define LOCK_UNLOCKED	0
105 #define LOCK_LOCKED	1
106 #define LOCK_COOLDOWN	1001
107 
108 vmem_t *shmif_units;
109 
110 /*
111  * This locking needs work and will misbehave severely if:
112  * 1) the backing memory has to be paged in
113  * 2) some lockholder exits while holding the lock
114  */
115 static void
116 shmif_lockbus(struct shmif_mem *busmem)
117 {
118 	int i = 0;
119 
120 	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
121 	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
122 		if (__predict_false(++i > LOCK_COOLDOWN)) {
123 			uint64_t sec, nsec;
124 			int error;
125 
126 			sec = 0;
127 			nsec = 1000*1000; /* 1ms */
128 			rumpuser_nanosleep(&sec, &nsec, &error);
129 			i = 0;
130 		}
131 		continue;
132 	}
133 	membar_enter();
134 }
135 
136 static void
137 shmif_unlockbus(struct shmif_mem *busmem)
138 {
139 	unsigned int old;
140 
141 	membar_exit();
142 	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
143 	KASSERT(old == LOCK_LOCKED);
144 }
145 
146 static int
147 allocif(int unit, struct shmif_sc **scp)
148 {
149 	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
150 	struct shmif_sc *sc;
151 	struct ifnet *ifp;
152 	uint32_t randnum;
153 	int error;
154 
155 	randnum = arc4random();
156 	memcpy(&enaddr[2], &randnum, sizeof(randnum));
157 
158 	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
159 	sc->sc_memfd = -1;
160 	sc->sc_unit = unit;
161 
162 	ifp = &sc->sc_ec.ec_if;
163 
164 	sprintf(ifp->if_xname, "shmif%d", unit);
165 	ifp->if_softc = sc;
166 	ifp->if_flags = IFF_BROADCAST | IFF_MULTICAST;
167 	ifp->if_init = shmif_init;
168 	ifp->if_ioctl = shmif_ioctl;
169 	ifp->if_start = shmif_start;
170 	ifp->if_stop = shmif_stop;
171 	ifp->if_mtu = ETHERMTU;
172 	ifp->if_dlt = DLT_EN10MB;
173 
174 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
175 	cv_init(&sc->sc_cv, "shmifcv");
176 
177 	if_attach(ifp);
178 	ether_ifattach(ifp, enaddr);
179 
180 	aprint_verbose("shmif%d: Ethernet address %s\n",
181 	    unit, ether_sprintf(enaddr));
182 
183 	if (scp)
184 		*scp = sc;
185 
186 	error = 0;
187 	if (rump_threads) {
188 		error = kthread_create(PRI_NONE,
189 		    KTHREAD_MPSAFE | KTHREAD_JOINABLE, NULL,
190 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
191 	} else {
192 		printf("WARNING: threads not enabled, shmif NOT working\n");
193 	}
194 
195 	if (error) {
196 		shmif_unclone(ifp);
197 	}
198 
199 	return error;
200 }
201 
202 static int
203 initbackend(struct shmif_sc *sc, int memfd)
204 {
205 	volatile uint8_t v;
206 	volatile uint8_t *p;
207 	int error;
208 
209 	sc->sc_busmem = rumpuser_filemmap(memfd, 0, BUSMEM_SIZE,
210 	    RUMPUSER_FILEMMAP_TRUNCATE | RUMPUSER_FILEMMAP_SHARED
211 	    | RUMPUSER_FILEMMAP_READ | RUMPUSER_FILEMMAP_WRITE, &error);
212 	if (error)
213 		return error;
214 
215 	if (sc->sc_busmem->shm_magic
216 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
217 		printf("bus is not magical");
218 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
219 		return ENOEXEC;
220 	}
221 
222 	/*
223 	 * Prefault in pages to minimize runtime penalty with buslock.
224 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
225 	 * rump kernel PAGE_SIZE > host page size.
226 	 */
227 	for (p = (uint8_t *)sc->sc_busmem;
228 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
229 	    p += 512)
230 		v = *p;
231 
232 	shmif_lockbus(sc->sc_busmem);
233 	/* we're first?  initialize bus */
234 	if (sc->sc_busmem->shm_magic == 0) {
235 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
236 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
237 	}
238 
239 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
240 	sc->sc_devgen = sc->sc_busmem->shm_gen;
241 
242 #ifdef PREFAULT_RW
243 	for (p = (uint8_t *)sc->sc_busmem;
244 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
245 	    p += PAGE_SIZE) {
246 		v = *p;
247 		*p = v;
248 	}
249 #endif
250 	shmif_unlockbus(sc->sc_busmem);
251 
252 	sc->sc_kq = rumpuser_writewatchfile_setup(-1, memfd, 0, &error);
253 	if (sc->sc_kq == -1) {
254 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
255 		return error;
256 	}
257 
258 	sc->sc_memfd = memfd;
259 
260 	return error;
261 }
262 
263 static void
264 finibackend(struct shmif_sc *sc)
265 {
266 
267 	if (sc->sc_backfile == NULL)
268 		return;
269 
270 	if (sc->sc_backfile) {
271 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
272 		sc->sc_backfile = NULL;
273 		sc->sc_backfilelen = 0;
274 	}
275 
276 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
277 	rumpuser_close(sc->sc_memfd, NULL);
278 	rumpuser_close(sc->sc_kq, NULL);
279 
280 	sc->sc_memfd = -1;
281 }
282 
283 int
284 rump_shmif_create(const char *path, int *ifnum)
285 {
286 	struct shmif_sc *sc;
287 	int unit, error;
288 	int memfd = -1; /* XXXgcc */
289 
290 	if (path) {
291 		memfd = rumpuser_open(path, O_RDWR | O_CREAT, &error);
292 		if (memfd == -1)
293 			return error;
294 	}
295 
296 	unit = vmem_xalloc(shmif_units, 1, 0, 0, 0, 0, 0,
297 	    VM_INSTANTFIT | VM_SLEEP) - 1;
298 
299 	if ((error = allocif(unit, &sc)) != 0) {
300 		if (path)
301 			rumpuser_close(memfd, NULL);
302 		return error;
303 	}
304 
305 	if (!path)
306 		goto out;
307 
308 	error = initbackend(sc, memfd);
309 	if (error) {
310 		shmif_unclone(&sc->sc_ec.ec_if);
311 		return error;
312 	}
313 
314 	sc->sc_backfilelen = strlen(path)+1;
315 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
316 	strcpy(sc->sc_backfile, path);
317 
318  out:
319 	if (ifnum)
320 		*ifnum = unit;
321 
322 	return 0;
323 }
324 
325 static int
326 shmif_clone(struct if_clone *ifc, int unit)
327 {
328 	int unit2;
329 
330 	/*
331 	 * Ok, we know the unit number, but we must still reserve it.
332 	 * Otherwise the wildcard-side of things might get the same one.
333 	 * This is slightly offset-happy due to vmem.  First, we offset
334 	 * the range of unit numbers by +1 since vmem cannot deal with
335 	 * ranges starting from 0.  Second, since vmem_xalloc() allocates
336 	 * from [min,max) (half-*open* interval), we need to add one extra
337 	 * to the one extra we add to maxaddr.  Talk about uuuh.
338 	 */
339 	unit2 = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+3,
340 	    VM_SLEEP | VM_INSTANTFIT);
341 	KASSERT(unit2-1 == unit);
342 
343 	return allocif(unit, NULL);
344 }
345 
346 static int
347 shmif_unclone(struct ifnet *ifp)
348 {
349 	struct shmif_sc *sc = ifp->if_softc;
350 
351 	shmif_stop(ifp, 1);
352 	if_down(ifp);
353 	finibackend(sc);
354 
355 	mutex_enter(&sc->sc_mtx);
356 	sc->sc_dying = true;
357 	cv_broadcast(&sc->sc_cv);
358 	mutex_exit(&sc->sc_mtx);
359 
360 	if (sc->sc_rcvl)
361 		kthread_join(sc->sc_rcvl);
362 	sc->sc_rcvl = NULL;
363 
364 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
365 
366 	ether_ifdetach(ifp);
367 	if_detach(ifp);
368 
369 	cv_destroy(&sc->sc_cv);
370 	mutex_destroy(&sc->sc_mtx);
371 
372 	kmem_free(sc, sizeof(*sc));
373 
374 	return 0;
375 }
376 
377 static int
378 shmif_init(struct ifnet *ifp)
379 {
380 	struct shmif_sc *sc = ifp->if_softc;
381 	int error = 0;
382 
383 	if (sc->sc_memfd == -1)
384 		return ENXIO;
385 	KASSERT(sc->sc_busmem);
386 
387 	ifp->if_flags |= IFF_RUNNING;
388 
389 	mutex_enter(&sc->sc_mtx);
390 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
391 	sc->sc_devgen = sc->sc_busmem->shm_gen;
392 
393 	cv_broadcast(&sc->sc_cv);
394 	mutex_exit(&sc->sc_mtx);
395 
396 	return error;
397 }
398 
399 static int
400 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
401 {
402 	struct shmif_sc *sc = ifp->if_softc;
403 	struct ifdrv *ifd;
404 	char *path;
405 	int s, rv, memfd;
406 
407 	s = splnet();
408 	switch (cmd) {
409 	case SIOCGLINKSTR:
410 		ifd = data;
411 
412 		if (sc->sc_backfilelen == 0) {
413 			rv = ENOENT;
414 			break;
415 		}
416 
417 		ifd->ifd_len = sc->sc_backfilelen;
418 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
419 			rv = 0;
420 			break;
421 		}
422 
423 		if (ifd->ifd_cmd != 0) {
424 			rv = EINVAL;
425 			break;
426 		}
427 
428 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
429 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
430 		break;
431 	case SIOCSLINKSTR:
432 		if (ifp->if_flags & IFF_UP) {
433 			rv = EBUSY;
434 			break;
435 		}
436 
437 		ifd = data;
438 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
439 			finibackend(sc);
440 			rv = 0;
441 			break;
442 		} else if (ifd->ifd_cmd != 0) {
443 			rv = EINVAL;
444 			break;
445 		} else if (sc->sc_backfile) {
446 			rv = EBUSY;
447 			break;
448 		}
449 
450 		if (ifd->ifd_len > MAXPATHLEN) {
451 			rv = E2BIG;
452 			break;
453 		} else if (ifd->ifd_len < 1) {
454 			rv = EINVAL;
455 			break;
456 		}
457 
458 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
459 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
460 		if (rv) {
461 			kmem_free(path, ifd->ifd_len);
462 			break;
463 		}
464 		memfd = rumpuser_open(path, O_RDWR | O_CREAT, &rv);
465 		if (memfd == -1) {
466 			kmem_free(path, ifd->ifd_len);
467 			break;
468 		}
469 		rv = initbackend(sc, memfd);
470 		if (rv) {
471 			kmem_free(path, ifd->ifd_len);
472 			rumpuser_close(memfd, NULL);
473 			break;
474 		}
475 		sc->sc_backfile = path;
476 		sc->sc_backfilelen = ifd->ifd_len;
477 
478 		break;
479 	default:
480 		rv = ether_ioctl(ifp, cmd, data);
481 		if (rv == ENETRESET)
482 			rv = 0;
483 		break;
484 	}
485 	splx(s);
486 
487 	return rv;
488 }
489 
490 /* send everything in-context since it's just a matter of mem-to-mem copy */
491 static void
492 shmif_start(struct ifnet *ifp)
493 {
494 	struct shmif_sc *sc = ifp->if_softc;
495 	struct shmif_mem *busmem = sc->sc_busmem;
496 	struct mbuf *m, *m0;
497 	uint32_t dataoff;
498 	uint32_t pktsize, pktwrote;
499 	bool wrote = false;
500 	bool wrap;
501 	int error;
502 
503 	ifp->if_flags |= IFF_OACTIVE;
504 
505 	for (;;) {
506 		struct shmif_pkthdr sp;
507 		struct timeval tv;
508 
509 		IF_DEQUEUE(&ifp->if_snd, m0);
510 		if (m0 == NULL) {
511 			break;
512 		}
513 
514 		pktsize = 0;
515 		for (m = m0; m != NULL; m = m->m_next) {
516 			pktsize += m->m_len;
517 		}
518 		KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
519 
520 		getmicrouptime(&tv);
521 		sp.sp_len = pktsize;
522 		sp.sp_sec = tv.tv_sec;
523 		sp.sp_usec = tv.tv_usec;
524 
525 		bpf_mtap(ifp, m0);
526 
527 		shmif_lockbus(busmem);
528 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
529 		busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
530 
531 		wrap = false;
532 		dataoff = shmif_buswrite(busmem,
533 		    busmem->shm_last, &sp, sizeof(sp), &wrap);
534 		pktwrote = 0;
535 		for (m = m0; m != NULL; m = m->m_next) {
536 			pktwrote += m->m_len;
537 			dataoff = shmif_buswrite(busmem, dataoff,
538 			    mtod(m, void *), m->m_len, &wrap);
539 		}
540 		KASSERT(pktwrote == pktsize);
541 		if (wrap) {
542 			busmem->shm_gen++;
543 			DPRINTF(("bus generation now %d\n", busmem->shm_gen));
544 		}
545 		shmif_unlockbus(busmem);
546 
547 		m_freem(m0);
548 		wrote = true;
549 
550 		DPRINTF(("shmif_start: send %d bytes at off %d\n",
551 		    pktsize, busmem->shm_last));
552 	}
553 
554 	ifp->if_flags &= ~IFF_OACTIVE;
555 
556 	/* wakeup? */
557 	if (wrote)
558 		rumpuser_pwrite(sc->sc_memfd,
559 		    &busversion, sizeof(busversion), IFMEM_WAKEUP, &error);
560 }
561 
562 static void
563 shmif_stop(struct ifnet *ifp, int disable)
564 {
565 	struct shmif_sc *sc = ifp->if_softc;
566 
567 	ifp->if_flags &= ~IFF_RUNNING;
568 	membar_producer();
569 
570 	/*
571 	 * wakeup thread.  this will of course wake up all bus
572 	 * listeners, but that's life.
573 	 */
574 	if (sc->sc_memfd != -1)
575 		rumpuser_pwrite(sc->sc_memfd,
576 		    &busversion, sizeof(busversion), IFMEM_WAKEUP, NULL);
577 }
578 
579 
580 /*
581  * Check if we have been sleeping too long.  Basically,
582  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
583  * We use the fact that first is guaranteed to never overlap
584  * with the last frame in the ring.
585  */
586 static __inline bool
587 stillvalid_p(struct shmif_sc *sc)
588 {
589 	struct shmif_mem *busmem = sc->sc_busmem;
590 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
591 	uint32_t lastoff, devoff;
592 
593 	KASSERT(busmem->shm_first != busmem->shm_last);
594 
595 	/* normalize onto a 2x busmem chunk */
596 	devoff = sc->sc_nextpacket;
597 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
598 
599 	/* trivial case */
600 	if (gendiff > 1)
601 		return false;
602 	KASSERT(gendiff <= 1);
603 
604 	/* Normalize onto 2x busmem chunk */
605 	if (busmem->shm_first >= lastoff) {
606 		lastoff += BUSMEM_DATASIZE;
607 		if (gendiff == 0)
608 			devoff += BUSMEM_DATASIZE;
609 	} else {
610 		if (gendiff)
611 			return false;
612 	}
613 
614 	return devoff >= busmem->shm_first && devoff <= lastoff;
615 }
616 
617 static void
618 shmif_rcv(void *arg)
619 {
620 	struct ifnet *ifp = arg;
621 	struct shmif_sc *sc = ifp->if_softc;
622 	struct shmif_mem *busmem;
623 	struct mbuf *m = NULL;
624 	struct ether_header *eth;
625 	uint32_t nextpkt;
626 	bool wrap, passup;
627 	int error;
628 
629  reup:
630 	mutex_enter(&sc->sc_mtx);
631 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
632 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
633 	mutex_exit(&sc->sc_mtx);
634 
635 	busmem = sc->sc_busmem;
636 
637 	while (ifp->if_flags & IFF_RUNNING) {
638 		struct shmif_pkthdr sp;
639 
640 		if (m == NULL) {
641 			m = m_gethdr(M_WAIT, MT_DATA);
642 			MCLGET(m, M_WAIT);
643 		}
644 
645 		DPRINTF(("waiting %d/%d\n", sc->sc_nextpacket, sc->sc_devgen));
646 		KASSERT(m->m_flags & M_EXT);
647 
648 		shmif_lockbus(busmem);
649 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
650 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
651 
652 		/* need more data? */
653 		if (sc->sc_devgen == busmem->shm_gen &&
654 		    shmif_nextpktoff(busmem, busmem->shm_last)
655 		     == sc->sc_nextpacket) {
656 			shmif_unlockbus(busmem);
657 			error = 0;
658 			rumpuser_writewatchfile_wait(sc->sc_kq, NULL, &error);
659 			if (__predict_false(error))
660 				printf("shmif_rcv: wait failed %d\n", error);
661 			membar_consumer();
662 			continue;
663 		}
664 
665 		if (stillvalid_p(sc)) {
666 			nextpkt = sc->sc_nextpacket;
667 		} else {
668 			KASSERT(busmem->shm_gen > 0);
669 			nextpkt = busmem->shm_first;
670 			if (busmem->shm_first > busmem->shm_last)
671 				sc->sc_devgen = busmem->shm_gen - 1;
672 			else
673 				sc->sc_devgen = busmem->shm_gen;
674 			DPRINTF(("dev %p overrun, new data: %d/%d\n",
675 			    sc, nextpkt, sc->sc_devgen));
676 		}
677 
678 		/*
679 		 * If our read pointer is ahead the bus last write, our
680 		 * generation must be one behind.
681 		 */
682 		KASSERT(!(nextpkt > busmem->shm_last
683 		    && sc->sc_devgen == busmem->shm_gen));
684 
685 		wrap = false;
686 		nextpkt = shmif_busread(busmem, &sp,
687 		    nextpkt, sizeof(sp), &wrap);
688 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
689 		nextpkt = shmif_busread(busmem, mtod(m, void *),
690 		    nextpkt, sp.sp_len, &wrap);
691 
692 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
693 		    sp.sp_len, nextpkt));
694 
695 		sc->sc_nextpacket = nextpkt;
696 		shmif_unlockbus(sc->sc_busmem);
697 
698 		if (wrap) {
699 			sc->sc_devgen++;
700 			DPRINTF(("dev %p generation now %d\n",
701 			    sc, sc->sc_devgen));
702 		}
703 
704 		m->m_len = m->m_pkthdr.len = sp.sp_len;
705 		m->m_pkthdr.rcvif = ifp;
706 
707 		/*
708 		 * Test if we want to pass the packet upwards
709 		 */
710 		eth = mtod(m, struct ether_header *);
711 		if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
712 		    ETHER_ADDR_LEN) == 0) {
713 			passup = true;
714 		} else if (memcmp(eth->ether_dhost, etherbroadcastaddr,
715 		    ETHER_ADDR_LEN) == 0) {
716 			passup = true;
717 		} else if (ifp->if_flags & IFF_PROMISC) {
718 			m->m_flags |= M_PROMISC;
719 			passup = true;
720 		} else {
721 			passup = false;
722 		}
723 
724 		if (passup) {
725 			KERNEL_LOCK(1, NULL);
726 			bpf_mtap(ifp, m);
727 			ifp->if_input(ifp, m);
728 			KERNEL_UNLOCK_ONE(NULL);
729 			m = NULL;
730 		}
731 		/* else: reuse mbuf for a future packet */
732 	}
733 	m_freem(m);
734 	m = NULL;
735 
736 	if (!sc->sc_dying)
737 		goto reup;
738 
739 	kthread_exit(0);
740 }
741