xref: /netbsd-src/sys/rump/net/lib/libshmif/if_shmem.c (revision 9aa0541bdf64142d9a27c2cf274394d60182818f)
1 /*	$NetBSD: if_shmem.c,v 1.43 2011/09/02 22:25:08 dyoung 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 1.43 2011/09/02 22:25:08 dyoung 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_MUSTJOIN, 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 	vmem_addr_t t;
288 	int unit, error;
289 	int memfd = -1; /* XXXgcc */
290 
291 	if (path) {
292 		memfd = rumpuser_open(path, O_RDWR | O_CREAT, &error);
293 		if (memfd == -1)
294 			return error;
295 	}
296 
297 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
298 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
299 
300 	if (error != 0) {
301 		if (path)
302 			rumpuser_close(memfd, NULL);
303 		return error;
304 	}
305 
306 	unit = t - 1;
307 
308 	if ((error = allocif(unit, &sc)) != 0) {
309 		if (path)
310 			rumpuser_close(memfd, NULL);
311 		return error;
312 	}
313 
314 	if (!path)
315 		goto out;
316 
317 	error = initbackend(sc, memfd);
318 	if (error) {
319 		shmif_unclone(&sc->sc_ec.ec_if);
320 		return error;
321 	}
322 
323 	sc->sc_backfilelen = strlen(path)+1;
324 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
325 	strcpy(sc->sc_backfile, path);
326 
327  out:
328 	if (ifnum)
329 		*ifnum = unit;
330 
331 	return 0;
332 }
333 
334 static int
335 shmif_clone(struct if_clone *ifc, int unit)
336 {
337 	int rc;
338 	vmem_addr_t unit2;
339 
340 	/*
341 	 * Ok, we know the unit number, but we must still reserve it.
342 	 * Otherwise the wildcard-side of things might get the same one.
343 	 * This is slightly offset-happy due to vmem.  First, we offset
344 	 * the range of unit numbers by +1 since vmem cannot deal with
345 	 * ranges starting from 0.  Talk about uuuh.
346 	 */
347 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
348 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
349 	KASSERT(rc == 0 && unit2-1 == unit);
350 
351 	return allocif(unit, NULL);
352 }
353 
354 static int
355 shmif_unclone(struct ifnet *ifp)
356 {
357 	struct shmif_sc *sc = ifp->if_softc;
358 
359 	shmif_stop(ifp, 1);
360 	if_down(ifp);
361 	finibackend(sc);
362 
363 	mutex_enter(&sc->sc_mtx);
364 	sc->sc_dying = true;
365 	cv_broadcast(&sc->sc_cv);
366 	mutex_exit(&sc->sc_mtx);
367 
368 	if (sc->sc_rcvl)
369 		kthread_join(sc->sc_rcvl);
370 	sc->sc_rcvl = NULL;
371 
372 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
373 
374 	ether_ifdetach(ifp);
375 	if_detach(ifp);
376 
377 	cv_destroy(&sc->sc_cv);
378 	mutex_destroy(&sc->sc_mtx);
379 
380 	kmem_free(sc, sizeof(*sc));
381 
382 	return 0;
383 }
384 
385 static int
386 shmif_init(struct ifnet *ifp)
387 {
388 	struct shmif_sc *sc = ifp->if_softc;
389 	int error = 0;
390 
391 	if (sc->sc_memfd == -1)
392 		return ENXIO;
393 	KASSERT(sc->sc_busmem);
394 
395 	ifp->if_flags |= IFF_RUNNING;
396 
397 	mutex_enter(&sc->sc_mtx);
398 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
399 	sc->sc_devgen = sc->sc_busmem->shm_gen;
400 
401 	cv_broadcast(&sc->sc_cv);
402 	mutex_exit(&sc->sc_mtx);
403 
404 	return error;
405 }
406 
407 static int
408 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
409 {
410 	struct shmif_sc *sc = ifp->if_softc;
411 	struct ifdrv *ifd;
412 	char *path;
413 	int s, rv, memfd;
414 
415 	s = splnet();
416 	switch (cmd) {
417 	case SIOCGLINKSTR:
418 		ifd = data;
419 
420 		if (sc->sc_backfilelen == 0) {
421 			rv = ENOENT;
422 			break;
423 		}
424 
425 		ifd->ifd_len = sc->sc_backfilelen;
426 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
427 			rv = 0;
428 			break;
429 		}
430 
431 		if (ifd->ifd_cmd != 0) {
432 			rv = EINVAL;
433 			break;
434 		}
435 
436 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
437 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
438 		break;
439 	case SIOCSLINKSTR:
440 		if (ifp->if_flags & IFF_UP) {
441 			rv = EBUSY;
442 			break;
443 		}
444 
445 		ifd = data;
446 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
447 			finibackend(sc);
448 			rv = 0;
449 			break;
450 		} else if (ifd->ifd_cmd != 0) {
451 			rv = EINVAL;
452 			break;
453 		} else if (sc->sc_backfile) {
454 			rv = EBUSY;
455 			break;
456 		}
457 
458 		if (ifd->ifd_len > MAXPATHLEN) {
459 			rv = E2BIG;
460 			break;
461 		} else if (ifd->ifd_len < 1) {
462 			rv = EINVAL;
463 			break;
464 		}
465 
466 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
467 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
468 		if (rv) {
469 			kmem_free(path, ifd->ifd_len);
470 			break;
471 		}
472 		memfd = rumpuser_open(path, O_RDWR | O_CREAT, &rv);
473 		if (memfd == -1) {
474 			kmem_free(path, ifd->ifd_len);
475 			break;
476 		}
477 		rv = initbackend(sc, memfd);
478 		if (rv) {
479 			kmem_free(path, ifd->ifd_len);
480 			rumpuser_close(memfd, NULL);
481 			break;
482 		}
483 		sc->sc_backfile = path;
484 		sc->sc_backfilelen = ifd->ifd_len;
485 
486 		break;
487 	default:
488 		rv = ether_ioctl(ifp, cmd, data);
489 		if (rv == ENETRESET)
490 			rv = 0;
491 		break;
492 	}
493 	splx(s);
494 
495 	return rv;
496 }
497 
498 /* send everything in-context since it's just a matter of mem-to-mem copy */
499 static void
500 shmif_start(struct ifnet *ifp)
501 {
502 	struct shmif_sc *sc = ifp->if_softc;
503 	struct shmif_mem *busmem = sc->sc_busmem;
504 	struct mbuf *m, *m0;
505 	uint32_t dataoff;
506 	uint32_t pktsize, pktwrote;
507 	bool wrote = false;
508 	bool wrap;
509 	int error;
510 
511 	ifp->if_flags |= IFF_OACTIVE;
512 
513 	for (;;) {
514 		struct shmif_pkthdr sp;
515 		struct timeval tv;
516 
517 		IF_DEQUEUE(&ifp->if_snd, m0);
518 		if (m0 == NULL) {
519 			break;
520 		}
521 
522 		pktsize = 0;
523 		for (m = m0; m != NULL; m = m->m_next) {
524 			pktsize += m->m_len;
525 		}
526 		KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
527 
528 		getmicrouptime(&tv);
529 		sp.sp_len = pktsize;
530 		sp.sp_sec = tv.tv_sec;
531 		sp.sp_usec = tv.tv_usec;
532 
533 		bpf_mtap(ifp, m0);
534 
535 		shmif_lockbus(busmem);
536 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
537 		busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
538 
539 		wrap = false;
540 		dataoff = shmif_buswrite(busmem,
541 		    busmem->shm_last, &sp, sizeof(sp), &wrap);
542 		pktwrote = 0;
543 		for (m = m0; m != NULL; m = m->m_next) {
544 			pktwrote += m->m_len;
545 			dataoff = shmif_buswrite(busmem, dataoff,
546 			    mtod(m, void *), m->m_len, &wrap);
547 		}
548 		KASSERT(pktwrote == pktsize);
549 		if (wrap) {
550 			busmem->shm_gen++;
551 			DPRINTF(("bus generation now %d\n", busmem->shm_gen));
552 		}
553 		shmif_unlockbus(busmem);
554 
555 		m_freem(m0);
556 		wrote = true;
557 
558 		DPRINTF(("shmif_start: send %d bytes at off %d\n",
559 		    pktsize, busmem->shm_last));
560 	}
561 
562 	ifp->if_flags &= ~IFF_OACTIVE;
563 
564 	/* wakeup? */
565 	if (wrote)
566 		rumpuser_pwrite(sc->sc_memfd,
567 		    &busversion, sizeof(busversion), IFMEM_WAKEUP, &error);
568 }
569 
570 static void
571 shmif_stop(struct ifnet *ifp, int disable)
572 {
573 	struct shmif_sc *sc = ifp->if_softc;
574 
575 	ifp->if_flags &= ~IFF_RUNNING;
576 	membar_producer();
577 
578 	/*
579 	 * wakeup thread.  this will of course wake up all bus
580 	 * listeners, but that's life.
581 	 */
582 	if (sc->sc_memfd != -1)
583 		rumpuser_pwrite(sc->sc_memfd,
584 		    &busversion, sizeof(busversion), IFMEM_WAKEUP, NULL);
585 }
586 
587 
588 /*
589  * Check if we have been sleeping too long.  Basically,
590  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
591  * We use the fact that first is guaranteed to never overlap
592  * with the last frame in the ring.
593  */
594 static __inline bool
595 stillvalid_p(struct shmif_sc *sc)
596 {
597 	struct shmif_mem *busmem = sc->sc_busmem;
598 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
599 	uint32_t lastoff, devoff;
600 
601 	KASSERT(busmem->shm_first != busmem->shm_last);
602 
603 	/* normalize onto a 2x busmem chunk */
604 	devoff = sc->sc_nextpacket;
605 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
606 
607 	/* trivial case */
608 	if (gendiff > 1)
609 		return false;
610 	KASSERT(gendiff <= 1);
611 
612 	/* Normalize onto 2x busmem chunk */
613 	if (busmem->shm_first >= lastoff) {
614 		lastoff += BUSMEM_DATASIZE;
615 		if (gendiff == 0)
616 			devoff += BUSMEM_DATASIZE;
617 	} else {
618 		if (gendiff)
619 			return false;
620 	}
621 
622 	return devoff >= busmem->shm_first && devoff <= lastoff;
623 }
624 
625 static void
626 shmif_rcv(void *arg)
627 {
628 	struct ifnet *ifp = arg;
629 	struct shmif_sc *sc = ifp->if_softc;
630 	struct shmif_mem *busmem;
631 	struct mbuf *m = NULL;
632 	struct ether_header *eth;
633 	uint32_t nextpkt;
634 	bool wrap, passup;
635 	int error;
636 
637  reup:
638 	mutex_enter(&sc->sc_mtx);
639 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
640 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
641 	mutex_exit(&sc->sc_mtx);
642 
643 	busmem = sc->sc_busmem;
644 
645 	while (ifp->if_flags & IFF_RUNNING) {
646 		struct shmif_pkthdr sp;
647 
648 		if (m == NULL) {
649 			m = m_gethdr(M_WAIT, MT_DATA);
650 			MCLGET(m, M_WAIT);
651 		}
652 
653 		DPRINTF(("waiting %d/%d\n", sc->sc_nextpacket, sc->sc_devgen));
654 		KASSERT(m->m_flags & M_EXT);
655 
656 		shmif_lockbus(busmem);
657 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
658 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
659 
660 		/* need more data? */
661 		if (sc->sc_devgen == busmem->shm_gen &&
662 		    shmif_nextpktoff(busmem, busmem->shm_last)
663 		     == sc->sc_nextpacket) {
664 			shmif_unlockbus(busmem);
665 			error = 0;
666 			rumpuser_writewatchfile_wait(sc->sc_kq, NULL, &error);
667 			if (__predict_false(error))
668 				printf("shmif_rcv: wait failed %d\n", error);
669 			membar_consumer();
670 			continue;
671 		}
672 
673 		if (stillvalid_p(sc)) {
674 			nextpkt = sc->sc_nextpacket;
675 		} else {
676 			KASSERT(busmem->shm_gen > 0);
677 			nextpkt = busmem->shm_first;
678 			if (busmem->shm_first > busmem->shm_last)
679 				sc->sc_devgen = busmem->shm_gen - 1;
680 			else
681 				sc->sc_devgen = busmem->shm_gen;
682 			DPRINTF(("dev %p overrun, new data: %d/%d\n",
683 			    sc, nextpkt, sc->sc_devgen));
684 		}
685 
686 		/*
687 		 * If our read pointer is ahead the bus last write, our
688 		 * generation must be one behind.
689 		 */
690 		KASSERT(!(nextpkt > busmem->shm_last
691 		    && sc->sc_devgen == busmem->shm_gen));
692 
693 		wrap = false;
694 		nextpkt = shmif_busread(busmem, &sp,
695 		    nextpkt, sizeof(sp), &wrap);
696 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
697 		nextpkt = shmif_busread(busmem, mtod(m, void *),
698 		    nextpkt, sp.sp_len, &wrap);
699 
700 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
701 		    sp.sp_len, nextpkt));
702 
703 		sc->sc_nextpacket = nextpkt;
704 		shmif_unlockbus(sc->sc_busmem);
705 
706 		if (wrap) {
707 			sc->sc_devgen++;
708 			DPRINTF(("dev %p generation now %d\n",
709 			    sc, sc->sc_devgen));
710 		}
711 
712 		m->m_len = m->m_pkthdr.len = sp.sp_len;
713 		m->m_pkthdr.rcvif = ifp;
714 
715 		/*
716 		 * Test if we want to pass the packet upwards
717 		 */
718 		eth = mtod(m, struct ether_header *);
719 		if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
720 		    ETHER_ADDR_LEN) == 0) {
721 			passup = true;
722 		} else if (memcmp(eth->ether_dhost, etherbroadcastaddr,
723 		    ETHER_ADDR_LEN) == 0) {
724 			passup = true;
725 		} else if (ifp->if_flags & IFF_PROMISC) {
726 			m->m_flags |= M_PROMISC;
727 			passup = true;
728 		} else {
729 			passup = false;
730 		}
731 
732 		if (passup) {
733 			KERNEL_LOCK(1, NULL);
734 			bpf_mtap(ifp, m);
735 			ifp->if_input(ifp, m);
736 			KERNEL_UNLOCK_ONE(NULL);
737 			m = NULL;
738 		}
739 		/* else: reuse mbuf for a future packet */
740 	}
741 	m_freem(m);
742 	m = NULL;
743 
744 	if (!sc->sc_dying)
745 		goto reup;
746 
747 	kthread_exit(0);
748 }
749