xref: /netbsd-src/sys/rump/net/lib/libshmif/if_shmem.c (revision 4724848cf0da353df257f730694b7882798e5daf)
1 /*	$NetBSD: if_shmem.c,v 1.84 2022/04/09 23:45:02 riastradh 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 1.84 2022/04/09 23:45:02 riastradh 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 #include <sys/cprng.h>
41 
42 #include <net/bpf.h>
43 #include <net/if.h>
44 #include <net/if_dl.h>
45 #include <net/if_ether.h>
46 #include <net/ether_sw_offload.h>
47 
48 #include <netinet/in.h>
49 #include <netinet/in_var.h>
50 
51 #include <rump-sys/kern.h>
52 #include <rump-sys/net.h>
53 
54 #include <rump/rump.h>
55 #include <rump/rumpuser.h>
56 
57 #include "shmif_user.h"
58 
59 static int shmif_clone(struct if_clone *, int);
60 static int shmif_unclone(struct ifnet *);
61 
62 struct if_clone shmif_cloner =
63     IF_CLONE_INITIALIZER("shmif", shmif_clone, shmif_unclone);
64 
65 /*
66  * Do r/w prefault for backend pages when attaching the interface.
67  * At least logically thinking improves performance (although no
68  * mlocking is done, so they might go away).
69  */
70 #define PREFAULT_RW
71 
72 /*
73  * A virtual ethernet interface which uses shared memory from a
74  * memory mapped file as the bus.
75  */
76 
77 static int	shmif_init(struct ifnet *);
78 static int	shmif_ioctl(struct ifnet *, u_long, void *);
79 static void	shmif_start(struct ifnet *);
80 static void	shmif_snd(struct ifnet *, struct mbuf *);
81 static void	shmif_stop(struct ifnet *, int);
82 
83 #include "shmifvar.h"
84 
85 struct shmif_sc {
86 	struct ethercom sc_ec;
87 	struct shmif_mem *sc_busmem;
88 	int sc_memfd;
89 	int sc_kq;
90 	int sc_unit;
91 
92 	char *sc_backfile;
93 	size_t sc_backfilelen;
94 
95 	uint64_t sc_devgen;
96 	uint32_t sc_nextpacket;
97 
98 	kmutex_t sc_mtx;
99 	kcondvar_t sc_cv;
100 
101 	struct lwp *sc_rcvl;
102 	bool sc_dying;
103 
104 	uint64_t sc_uid;
105 };
106 
107 static void shmif_rcv(void *);
108 
109 #define LOCK_UNLOCKED	0
110 #define LOCK_LOCKED	1
111 #define LOCK_COOLDOWN	1001
112 
113 vmem_t *shmif_units;
114 
115 static void
116 dowakeup(struct shmif_sc *sc)
117 {
118 	struct rumpuser_iovec iov;
119 	uint32_t ver = SHMIF_VERSION;
120 	size_t n;
121 
122 	iov.iov_base = &ver;
123 	iov.iov_len = sizeof(ver);
124 	rumpuser_iovwrite(sc->sc_memfd, &iov, 1, IFMEM_WAKEUP, &n);
125 }
126 
127 /*
128  * This locking needs work and will misbehave severely if:
129  * 1) the backing memory has to be paged in
130  * 2) some lockholder exits while holding the lock
131  */
132 static void
133 shmif_lockbus(struct shmif_mem *busmem)
134 {
135 	int i = 0;
136 
137 	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
138 	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
139 		if (__predict_false(++i > LOCK_COOLDOWN)) {
140 			/* wait 1ms */
141 			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL,
142 			    0, 1000*1000);
143 			i = 0;
144 		}
145 		continue;
146 	}
147 	membar_acquire();
148 }
149 
150 static void
151 shmif_unlockbus(struct shmif_mem *busmem)
152 {
153 	unsigned int old __diagused;
154 
155 	membar_release();
156 	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
157 	KASSERT(old == LOCK_LOCKED);
158 }
159 
160 static int
161 allocif(int unit, struct shmif_sc **scp)
162 {
163 	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
164 	struct shmif_sc *sc;
165 	struct ifnet *ifp;
166 	uint64_t randnum;
167 	int error = 0;
168 
169 	randnum = cprng_strong64();
170 	memcpy(&enaddr[2], &randnum, 4);
171 
172 	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
173 	sc->sc_memfd = -1;
174 	sc->sc_unit = unit;
175 	sc->sc_uid = randnum;
176 
177 	ifp = &sc->sc_ec.ec_if;
178 
179 	snprintf(ifp->if_xname, sizeof(ifp->if_xname), "shmif%d", unit);
180 	ifp->if_softc = sc;
181 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
182 	ifp->if_init = shmif_init;
183 	ifp->if_ioctl = shmif_ioctl;
184 	ifp->if_start = shmif_start;
185 	ifp->if_stop = shmif_stop;
186 	ifp->if_mtu = ETHERMTU;
187 	ifp->if_dlt = DLT_EN10MB;
188 	ifp->if_capabilities = IFCAP_TSOv4 | IFCAP_TSOv6 |
189 	    IFCAP_CSUM_IPv4_Rx	| IFCAP_CSUM_IPv4_Tx |
190 	    IFCAP_CSUM_TCPv4_Rx	| IFCAP_CSUM_TCPv4_Tx |
191 	    IFCAP_CSUM_UDPv4_Rx	| IFCAP_CSUM_UDPv4_Tx |
192 	    IFCAP_CSUM_TCPv6_Rx	| IFCAP_CSUM_TCPv6_Tx |
193 	    IFCAP_CSUM_UDPv6_Rx	| IFCAP_CSUM_UDPv6_Tx;
194 	IFQ_SET_READY(&ifp->if_snd);
195 
196 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
197 	cv_init(&sc->sc_cv, "shmifcv");
198 
199 	if_initialize(ifp);
200 #if 1
201 	char buf[256];
202 
203 	if (rumpuser_getparam("RUMP_SHMIF_CAPENABLE", buf, sizeof(buf)) == 0) {
204 		uint64_t capen = strtoul(buf, NULL, 0);
205 
206 		ifp->if_capenable = capen & ifp->if_capabilities;
207 	}
208 #endif
209 
210 	if_deferred_start_init(ifp, NULL);
211 	ether_ifattach(ifp, enaddr);
212 	if_register(ifp);
213 
214 	aprint_verbose("shmif%d: Ethernet address %s\n",
215 	    unit, ether_sprintf(enaddr));
216 
217 	if (scp)
218 		*scp = sc;
219 
220 	if (rump_threads) {
221 		error = kthread_create(PRI_NONE,
222 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
223 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
224 	} else {
225 		printf("WARNING: threads not enabled, shmif NOT working\n");
226 	}
227 
228 	if (error) {
229 		shmif_unclone(ifp);
230 	}
231 
232 	return 0;
233 }
234 
235 static int
236 initbackend(struct shmif_sc *sc, int memfd)
237 {
238 	volatile uint8_t v;
239 	volatile uint8_t *p;
240 	void *mem;
241 	int error;
242 
243 	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
244 	if (error)
245 		return error;
246 	sc->sc_busmem = mem;
247 
248 	if (sc->sc_busmem->shm_magic
249 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
250 		printf("bus is not magical");
251 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
252 		return ENOEXEC;
253 	}
254 
255 	/*
256 	 * Prefault in pages to minimize runtime penalty with buslock.
257 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
258 	 * rump kernel PAGE_SIZE > host page size.
259 	 */
260 	for (p = (uint8_t *)sc->sc_busmem;
261 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
262 	    p += 512)
263 		v = *p;
264 
265 	shmif_lockbus(sc->sc_busmem);
266 	/* we're first?  initialize bus */
267 	if (sc->sc_busmem->shm_magic == 0) {
268 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
269 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
270 	}
271 
272 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
273 	sc->sc_devgen = sc->sc_busmem->shm_gen;
274 
275 #ifdef PREFAULT_RW
276 	for (p = (uint8_t *)sc->sc_busmem;
277 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
278 	    p += PAGE_SIZE) {
279 		v = *p;
280 		*p = v;
281 	}
282 #endif
283 	shmif_unlockbus(sc->sc_busmem);
284 
285 	sc->sc_kq = -1;
286 	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
287 	if (error) {
288 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
289 		return error;
290 	}
291 
292 	sc->sc_memfd = memfd;
293 
294 	return error;
295 }
296 
297 static void
298 finibackend(struct shmif_sc *sc)
299 {
300 
301 	if (sc->sc_backfile == NULL)
302 		return;
303 
304 	if (sc->sc_backfile) {
305 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
306 		sc->sc_backfile = NULL;
307 		sc->sc_backfilelen = 0;
308 	}
309 
310 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
311 	rumpuser_close(sc->sc_memfd);
312 	rumpuser_close(sc->sc_kq);
313 
314 	sc->sc_memfd = -1;
315 }
316 
317 int
318 rump_shmif_create(const char *path, int *ifnum)
319 {
320 	struct shmif_sc *sc;
321 	vmem_addr_t t;
322 	int unit, error;
323 	int memfd = -1; /* XXXgcc */
324 
325 	if (path) {
326 		error = rumpuser_open(path,
327 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
328 		if (error)
329 			return error;
330 	}
331 
332 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
333 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
334 
335 	if (error != 0) {
336 		if (path)
337 			rumpuser_close(memfd);
338 		return error;
339 	}
340 
341 	unit = t - 1;
342 
343 	if ((error = allocif(unit, &sc)) != 0) {
344 		if (path)
345 			rumpuser_close(memfd);
346 		return error;
347 	}
348 
349 	if (!path)
350 		goto out;
351 
352 	error = initbackend(sc, memfd);
353 	if (error) {
354 		shmif_unclone(&sc->sc_ec.ec_if);
355 		return error;
356 	}
357 
358 	sc->sc_backfilelen = strlen(path)+1;
359 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
360 	strcpy(sc->sc_backfile, path);
361 
362  out:
363 	if (ifnum)
364 		*ifnum = unit;
365 
366 	return 0;
367 }
368 
369 static int
370 shmif_clone(struct if_clone *ifc, int unit)
371 {
372 	int rc __diagused;
373 	vmem_addr_t unit2;
374 
375 	/*
376 	 * Ok, we know the unit number, but we must still reserve it.
377 	 * Otherwise the wildcard-side of things might get the same one.
378 	 * This is slightly offset-happy due to vmem.  First, we offset
379 	 * the range of unit numbers by +1 since vmem cannot deal with
380 	 * ranges starting from 0.  Talk about uuuh.
381 	 */
382 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
383 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
384 	KASSERT(rc == 0 && unit2-1 == unit);
385 
386 	return allocif(unit, NULL);
387 }
388 
389 static int
390 shmif_unclone(struct ifnet *ifp)
391 {
392 	struct shmif_sc *sc = ifp->if_softc;
393 
394 	shmif_stop(ifp, 1);
395 	if_down(ifp);
396 
397 	mutex_enter(&sc->sc_mtx);
398 	sc->sc_dying = true;
399 	cv_broadcast(&sc->sc_cv);
400 	mutex_exit(&sc->sc_mtx);
401 
402 	if (sc->sc_rcvl)
403 		kthread_join(sc->sc_rcvl);
404 	sc->sc_rcvl = NULL;
405 
406 	/*
407 	 * Need to be called after the kthread left, otherwise closing kqueue
408 	 * (sc_kq) hangs sometimes perhaps because of a race condition between
409 	 * close and kevent in the kthread on the kqueue.
410 	 */
411 	finibackend(sc);
412 
413 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
414 
415 	ether_ifdetach(ifp);
416 	if_detach(ifp);
417 
418 	cv_destroy(&sc->sc_cv);
419 	mutex_destroy(&sc->sc_mtx);
420 
421 	kmem_free(sc, sizeof(*sc));
422 
423 	return 0;
424 }
425 
426 static int
427 shmif_init(struct ifnet *ifp)
428 {
429 	struct shmif_sc *sc = ifp->if_softc;
430 	int error = 0;
431 
432 	if (sc->sc_memfd == -1)
433 		return ENXIO;
434 	KASSERT(sc->sc_busmem);
435 
436 	ifp->if_flags |= IFF_RUNNING;
437 
438 	mutex_enter(&sc->sc_mtx);
439 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
440 	sc->sc_devgen = sc->sc_busmem->shm_gen;
441 
442 	cv_broadcast(&sc->sc_cv);
443 	mutex_exit(&sc->sc_mtx);
444 
445 	return error;
446 }
447 
448 static int
449 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
450 {
451 	struct shmif_sc *sc = ifp->if_softc;
452 	struct ifdrv *ifd;
453 	char *path;
454 	int s, rv, memfd;
455 
456 	s = splnet();
457 	switch (cmd) {
458 	case SIOCGLINKSTR:
459 		ifd = data;
460 
461 		if (sc->sc_backfilelen == 0) {
462 			rv = ENOENT;
463 			break;
464 		}
465 
466 		ifd->ifd_len = sc->sc_backfilelen;
467 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
468 			rv = 0;
469 			break;
470 		}
471 
472 		if (ifd->ifd_cmd != 0) {
473 			rv = EINVAL;
474 			break;
475 		}
476 
477 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
478 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
479 		break;
480 	case SIOCSLINKSTR:
481 		if (ifp->if_flags & IFF_UP) {
482 			rv = EBUSY;
483 			break;
484 		}
485 
486 		ifd = data;
487 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
488 			finibackend(sc);
489 			rv = 0;
490 			break;
491 		} else if (ifd->ifd_cmd != 0) {
492 			rv = EINVAL;
493 			break;
494 		} else if (sc->sc_backfile) {
495 			rv = EBUSY;
496 			break;
497 		}
498 
499 		if (ifd->ifd_len > MAXPATHLEN) {
500 			rv = E2BIG;
501 			break;
502 		} else if (ifd->ifd_len < 1) {
503 			rv = EINVAL;
504 			break;
505 		}
506 
507 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
508 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
509 		if (rv) {
510 			kmem_free(path, ifd->ifd_len);
511 			break;
512 		}
513 		rv = rumpuser_open(path,
514 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
515 		if (rv) {
516 			kmem_free(path, ifd->ifd_len);
517 			break;
518 		}
519 		rv = initbackend(sc, memfd);
520 		if (rv) {
521 			kmem_free(path, ifd->ifd_len);
522 			rumpuser_close(memfd);
523 			break;
524 		}
525 		sc->sc_backfile = path;
526 		sc->sc_backfilelen = ifd->ifd_len;
527 
528 		break;
529 	default:
530 		rv = ether_ioctl(ifp, cmd, data);
531 		if (rv == ENETRESET)
532 			rv = 0;
533 		break;
534 	}
535 	splx(s);
536 
537 	return rv;
538 }
539 
540 static void
541 shmif_start(struct ifnet *ifp)
542 {
543 	struct shmif_sc *sc = ifp->if_softc;
544 	struct mbuf *m, *n;
545 	bool wrote = false;
546 
547 	ifp->if_flags |= IFF_OACTIVE;
548 
549 	for (;;) {
550 		IFQ_DEQUEUE(&ifp->if_snd, m);
551 		if (m == NULL)
552 			break;
553 
554 		m = ether_sw_offload_tx(ifp, m);
555 		if (m == NULL) {
556 			if_statinc(ifp, if_oerrors);
557 			break;
558 		}
559 
560 		do {
561 			n = m->m_nextpkt;
562 			shmif_snd(ifp, m);
563 			m = n;
564 		} while (m != NULL);
565 
566 		wrote = true;
567 	}
568 
569 	ifp->if_flags &= ~IFF_OACTIVE;
570 
571 	/* wakeup? */
572 	if (wrote) {
573 		dowakeup(sc);
574 	}
575 }
576 
577 /* send everything in-context since it's just a matter of mem-to-mem copy */
578 static void
579 shmif_snd(struct ifnet *ifp, struct mbuf *m0)
580 {
581 	struct shmif_sc *sc = ifp->if_softc;
582 	struct shmif_mem *busmem = sc->sc_busmem;
583 	struct shmif_pkthdr sp;
584 	struct timeval tv;
585 	struct mbuf *m;
586 	uint32_t dataoff;
587 	uint32_t pktsize, pktwrote;
588 	bool wrap;
589 
590 	pktsize = 0;
591 	for (m = m0; m != NULL; m = m->m_next) {
592 		pktsize += m->m_len;
593 	}
594 	KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
595 
596 	getmicrouptime(&tv);
597 	sp.sp_len = pktsize;
598 	sp.sp_sec = tv.tv_sec;
599 	sp.sp_usec = tv.tv_usec;
600 	sp.sp_sender = sc->sc_uid;
601 
602 	bpf_mtap(ifp, m0, BPF_D_OUT);
603 
604 	shmif_lockbus(busmem);
605 	KASSERT(busmem->shm_magic == SHMIF_MAGIC);
606 	busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
607 
608 	wrap = false;
609 	dataoff =
610 	    shmif_buswrite(busmem, busmem->shm_last, &sp, sizeof(sp), &wrap);
611 	pktwrote = 0;
612 	for (m = m0; m != NULL; m = m->m_next) {
613 		pktwrote += m->m_len;
614 		dataoff = shmif_buswrite(busmem, dataoff, mtod(m, void *),
615 		    m->m_len, &wrap);
616 	}
617 	KASSERT(pktwrote == pktsize);
618 	if (wrap) {
619 		busmem->shm_gen++;
620 		DPRINTF(("bus generation now %" PRIu64 "\n", busmem->shm_gen));
621 	}
622 	shmif_unlockbus(busmem);
623 
624 	m_freem(m0);
625 	if_statinc(ifp, if_opackets);
626 
627 	DPRINTF(("shmif_start: send %d bytes at off %d\n", pktsize,
628 	    busmem->shm_last));
629 }
630 
631 static void
632 shmif_stop(struct ifnet *ifp, int disable)
633 {
634 	struct shmif_sc *sc = ifp->if_softc;
635 
636 	ifp->if_flags &= ~IFF_RUNNING;
637 	membar_producer();
638 
639 	/*
640 	 * wakeup thread.  this will of course wake up all bus
641 	 * listeners, but that's life.
642 	 */
643 	if (sc->sc_memfd != -1) {
644 		dowakeup(sc);
645 	}
646 }
647 
648 
649 /*
650  * Check if we have been sleeping too long.  Basically,
651  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
652  * We use the fact that first is guaranteed to never overlap
653  * with the last frame in the ring.
654  */
655 static __inline bool
656 stillvalid_p(struct shmif_sc *sc)
657 {
658 	struct shmif_mem *busmem = sc->sc_busmem;
659 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
660 	uint32_t lastoff, devoff;
661 
662 	KASSERT(busmem->shm_first != busmem->shm_last);
663 
664 	/* normalize onto a 2x busmem chunk */
665 	devoff = sc->sc_nextpacket;
666 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
667 
668 	/* trivial case */
669 	if (gendiff > 1)
670 		return false;
671 	KASSERT(gendiff <= 1);
672 
673 	/* Normalize onto 2x busmem chunk */
674 	if (busmem->shm_first >= lastoff) {
675 		lastoff += BUSMEM_DATASIZE;
676 		if (gendiff == 0)
677 			devoff += BUSMEM_DATASIZE;
678 	} else {
679 		if (gendiff)
680 			return false;
681 	}
682 
683 	return devoff >= busmem->shm_first && devoff <= lastoff;
684 }
685 
686 static void
687 shmif_rcv(void *arg)
688 {
689 	struct ifnet *ifp = arg;
690 	struct shmif_sc *sc = ifp->if_softc;
691 	struct shmif_mem *busmem;
692 	struct mbuf *m = NULL;
693 	struct ether_header *eth;
694 	uint32_t nextpkt;
695 	bool wrap, passup;
696 	int error;
697 	const int align
698 	    = ALIGN(sizeof(struct ether_header)) - sizeof(struct ether_header);
699 
700  reup:
701 	mutex_enter(&sc->sc_mtx);
702 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
703 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
704 	mutex_exit(&sc->sc_mtx);
705 
706 	busmem = sc->sc_busmem;
707 
708 	while (ifp->if_flags & IFF_RUNNING) {
709 		struct shmif_pkthdr sp;
710 
711 		if (m == NULL) {
712 			m = m_gethdr(M_WAIT, MT_DATA);
713 			MCLGET(m, M_WAIT);
714 			m->m_data += align;
715 		}
716 
717 		DPRINTF(("waiting %d/%" PRIu64 "\n",
718 		    sc->sc_nextpacket, sc->sc_devgen));
719 		KASSERT(m->m_flags & M_EXT);
720 
721 		shmif_lockbus(busmem);
722 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
723 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
724 
725 		/* need more data? */
726 		if (sc->sc_devgen == busmem->shm_gen &&
727 		    shmif_nextpktoff(busmem, busmem->shm_last)
728 		     == sc->sc_nextpacket) {
729 			shmif_unlockbus(busmem);
730 			error = rumpcomp_shmif_watchwait(sc->sc_kq);
731 			if (__predict_false(error))
732 				printf("shmif_rcv: wait failed %d\n", error);
733 			membar_consumer();
734 			continue;
735 		}
736 
737 		if (stillvalid_p(sc)) {
738 			nextpkt = sc->sc_nextpacket;
739 		} else {
740 			KASSERT(busmem->shm_gen > 0);
741 			nextpkt = busmem->shm_first;
742 			if (busmem->shm_first > busmem->shm_last)
743 				sc->sc_devgen = busmem->shm_gen - 1;
744 			else
745 				sc->sc_devgen = busmem->shm_gen;
746 			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
747 			    sc, nextpkt, sc->sc_devgen));
748 		}
749 
750 		/*
751 		 * If our read pointer is ahead the bus last write, our
752 		 * generation must be one behind.
753 		 */
754 		KASSERT(!(nextpkt > busmem->shm_last
755 		    && sc->sc_devgen == busmem->shm_gen));
756 
757 		wrap = false;
758 		nextpkt = shmif_busread(busmem, &sp,
759 		    nextpkt, sizeof(sp), &wrap);
760 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
761 		nextpkt = shmif_busread(busmem, mtod(m, void *),
762 		    nextpkt, sp.sp_len, &wrap);
763 
764 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
765 		    sp.sp_len, nextpkt));
766 
767 		sc->sc_nextpacket = nextpkt;
768 		shmif_unlockbus(sc->sc_busmem);
769 
770 		if (wrap) {
771 			sc->sc_devgen++;
772 			DPRINTF(("dev %p generation now %" PRIu64 "\n",
773 			    sc, sc->sc_devgen));
774 		}
775 
776 		/*
777 		 * Ignore packets too short to possibly be valid.
778 		 * This is hit at least for the first frame on a new bus.
779 		 */
780 		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
781 			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
782 			    sp.sp_len));
783 			continue;
784 		}
785 
786 		m->m_len = m->m_pkthdr.len = sp.sp_len;
787 		m_set_rcvif(m, ifp);
788 
789 		/*
790 		 * Test if we want to pass the packet upwards
791 		 */
792 		eth = mtod(m, struct ether_header *);
793 		if (sp.sp_sender == sc->sc_uid) {
794 			passup = false;
795 		} else if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
796 		    ETHER_ADDR_LEN) == 0) {
797 			passup = true;
798 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
799 			passup = true;
800 		} else if (ifp->if_flags & IFF_PROMISC) {
801 			m->m_flags |= M_PROMISC;
802 			passup = true;
803 		} else {
804 			passup = false;
805 		}
806 
807 		if (passup) {
808 			int bound;
809 
810 			m = ether_sw_offload_rx(ifp, m);
811 
812 			KERNEL_LOCK(1, NULL);
813 			/* Prevent LWP migrations between CPUs for psref(9) */
814 			bound = curlwp_bind();
815 			if_input(ifp, m);
816 			curlwp_bindx(bound);
817 			KERNEL_UNLOCK_ONE(NULL);
818 
819 			m = NULL;
820 		}
821 		/* else: reuse mbuf for a future packet */
822 	}
823 	m_freem(m);
824 	m = NULL;
825 
826 	if (!sc->sc_dying)
827 		goto reup;
828 
829 	kthread_exit(0);
830 }
831