xref: /netbsd-src/sys/rump/net/lib/libshmif/if_shmem.c (revision 181254a7b1bdde6873432bffef2d2decc4b5c22f)
1 /*	$NetBSD: if_shmem.c,v 1.81 2020/02/25 03:26:18 ozaki-r 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 1.81 2020/02/25 03:26:18 ozaki-r 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_enter();
148 }
149 
150 static void
151 shmif_unlockbus(struct shmif_mem *busmem)
152 {
153 	unsigned int old __diagused;
154 
155 	membar_exit();
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;
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 
195 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
196 	cv_init(&sc->sc_cv, "shmifcv");
197 
198 	error = if_initialize(ifp);
199 	if (error != 0) {
200 		aprint_error("shmif%d: if_initialize failed(%d)\n", unit,
201 		    error);
202 		cv_destroy(&sc->sc_cv);
203 		mutex_destroy(&sc->sc_mtx);
204 		kmem_free(sc, sizeof(*sc));
205 
206 		return error;
207 	}
208 #if 1
209 	char buf[256];
210 
211 	if (rumpuser_getparam("RUMP_SHMIF_CAPENABLE", buf, sizeof(buf)) == 0) {
212 		uint64_t capen = strtoul(buf, NULL, 0);
213 
214 		ifp->if_capenable = capen & ifp->if_capabilities;
215 	}
216 #endif
217 
218 	ether_ifattach(ifp, enaddr);
219 	if_register(ifp);
220 
221 	aprint_verbose("shmif%d: Ethernet address %s\n",
222 	    unit, ether_sprintf(enaddr));
223 
224 	if (scp)
225 		*scp = sc;
226 
227 	error = 0;
228 	if (rump_threads) {
229 		error = kthread_create(PRI_NONE,
230 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
231 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
232 	} else {
233 		printf("WARNING: threads not enabled, shmif NOT working\n");
234 	}
235 
236 	if (error) {
237 		shmif_unclone(ifp);
238 	}
239 
240 	return error;
241 }
242 
243 static int
244 initbackend(struct shmif_sc *sc, int memfd)
245 {
246 	volatile uint8_t v;
247 	volatile uint8_t *p;
248 	void *mem;
249 	int error;
250 
251 	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
252 	if (error)
253 		return error;
254 	sc->sc_busmem = mem;
255 
256 	if (sc->sc_busmem->shm_magic
257 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
258 		printf("bus is not magical");
259 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
260 		return ENOEXEC;
261 	}
262 
263 	/*
264 	 * Prefault in pages to minimize runtime penalty with buslock.
265 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
266 	 * rump kernel PAGE_SIZE > host page size.
267 	 */
268 	for (p = (uint8_t *)sc->sc_busmem;
269 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
270 	    p += 512)
271 		v = *p;
272 
273 	shmif_lockbus(sc->sc_busmem);
274 	/* we're first?  initialize bus */
275 	if (sc->sc_busmem->shm_magic == 0) {
276 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
277 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
278 	}
279 
280 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
281 	sc->sc_devgen = sc->sc_busmem->shm_gen;
282 
283 #ifdef PREFAULT_RW
284 	for (p = (uint8_t *)sc->sc_busmem;
285 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
286 	    p += PAGE_SIZE) {
287 		v = *p;
288 		*p = v;
289 	}
290 #endif
291 	shmif_unlockbus(sc->sc_busmem);
292 
293 	sc->sc_kq = -1;
294 	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
295 	if (error) {
296 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
297 		return error;
298 	}
299 
300 	sc->sc_memfd = memfd;
301 
302 	return error;
303 }
304 
305 static void
306 finibackend(struct shmif_sc *sc)
307 {
308 
309 	if (sc->sc_backfile == NULL)
310 		return;
311 
312 	if (sc->sc_backfile) {
313 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
314 		sc->sc_backfile = NULL;
315 		sc->sc_backfilelen = 0;
316 	}
317 
318 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
319 	rumpuser_close(sc->sc_memfd);
320 	rumpuser_close(sc->sc_kq);
321 
322 	sc->sc_memfd = -1;
323 }
324 
325 int
326 rump_shmif_create(const char *path, int *ifnum)
327 {
328 	struct shmif_sc *sc;
329 	vmem_addr_t t;
330 	int unit, error;
331 	int memfd = -1; /* XXXgcc */
332 
333 	if (path) {
334 		error = rumpuser_open(path,
335 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
336 		if (error)
337 			return error;
338 	}
339 
340 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
341 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
342 
343 	if (error != 0) {
344 		if (path)
345 			rumpuser_close(memfd);
346 		return error;
347 	}
348 
349 	unit = t - 1;
350 
351 	if ((error = allocif(unit, &sc)) != 0) {
352 		if (path)
353 			rumpuser_close(memfd);
354 		return error;
355 	}
356 
357 	if (!path)
358 		goto out;
359 
360 	error = initbackend(sc, memfd);
361 	if (error) {
362 		shmif_unclone(&sc->sc_ec.ec_if);
363 		return error;
364 	}
365 
366 	sc->sc_backfilelen = strlen(path)+1;
367 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
368 	strcpy(sc->sc_backfile, path);
369 
370  out:
371 	if (ifnum)
372 		*ifnum = unit;
373 
374 	return 0;
375 }
376 
377 static int
378 shmif_clone(struct if_clone *ifc, int unit)
379 {
380 	int rc __diagused;
381 	vmem_addr_t unit2;
382 
383 	/*
384 	 * Ok, we know the unit number, but we must still reserve it.
385 	 * Otherwise the wildcard-side of things might get the same one.
386 	 * This is slightly offset-happy due to vmem.  First, we offset
387 	 * the range of unit numbers by +1 since vmem cannot deal with
388 	 * ranges starting from 0.  Talk about uuuh.
389 	 */
390 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
391 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
392 	KASSERT(rc == 0 && unit2-1 == unit);
393 
394 	return allocif(unit, NULL);
395 }
396 
397 static int
398 shmif_unclone(struct ifnet *ifp)
399 {
400 	struct shmif_sc *sc = ifp->if_softc;
401 
402 	shmif_stop(ifp, 1);
403 	if_down(ifp);
404 
405 	mutex_enter(&sc->sc_mtx);
406 	sc->sc_dying = true;
407 	cv_broadcast(&sc->sc_cv);
408 	mutex_exit(&sc->sc_mtx);
409 
410 	if (sc->sc_rcvl)
411 		kthread_join(sc->sc_rcvl);
412 	sc->sc_rcvl = NULL;
413 
414 	/*
415 	 * Need to be called after the kthread left, otherwise closing kqueue
416 	 * (sc_kq) hangs sometimes perhaps because of a race condition between
417 	 * close and kevent in the kthread on the kqueue.
418 	 */
419 	finibackend(sc);
420 
421 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
422 
423 	ether_ifdetach(ifp);
424 	if_detach(ifp);
425 
426 	cv_destroy(&sc->sc_cv);
427 	mutex_destroy(&sc->sc_mtx);
428 
429 	kmem_free(sc, sizeof(*sc));
430 
431 	return 0;
432 }
433 
434 static int
435 shmif_init(struct ifnet *ifp)
436 {
437 	struct shmif_sc *sc = ifp->if_softc;
438 	int error = 0;
439 
440 	if (sc->sc_memfd == -1)
441 		return ENXIO;
442 	KASSERT(sc->sc_busmem);
443 
444 	ifp->if_flags |= IFF_RUNNING;
445 
446 	mutex_enter(&sc->sc_mtx);
447 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
448 	sc->sc_devgen = sc->sc_busmem->shm_gen;
449 
450 	cv_broadcast(&sc->sc_cv);
451 	mutex_exit(&sc->sc_mtx);
452 
453 	return error;
454 }
455 
456 static int
457 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
458 {
459 	struct shmif_sc *sc = ifp->if_softc;
460 	struct ifdrv *ifd;
461 	char *path;
462 	int s, rv, memfd;
463 
464 	s = splnet();
465 	switch (cmd) {
466 	case SIOCGLINKSTR:
467 		ifd = data;
468 
469 		if (sc->sc_backfilelen == 0) {
470 			rv = ENOENT;
471 			break;
472 		}
473 
474 		ifd->ifd_len = sc->sc_backfilelen;
475 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
476 			rv = 0;
477 			break;
478 		}
479 
480 		if (ifd->ifd_cmd != 0) {
481 			rv = EINVAL;
482 			break;
483 		}
484 
485 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
486 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
487 		break;
488 	case SIOCSLINKSTR:
489 		if (ifp->if_flags & IFF_UP) {
490 			rv = EBUSY;
491 			break;
492 		}
493 
494 		ifd = data;
495 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
496 			finibackend(sc);
497 			rv = 0;
498 			break;
499 		} else if (ifd->ifd_cmd != 0) {
500 			rv = EINVAL;
501 			break;
502 		} else if (sc->sc_backfile) {
503 			rv = EBUSY;
504 			break;
505 		}
506 
507 		if (ifd->ifd_len > MAXPATHLEN) {
508 			rv = E2BIG;
509 			break;
510 		} else if (ifd->ifd_len < 1) {
511 			rv = EINVAL;
512 			break;
513 		}
514 
515 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
516 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
517 		if (rv) {
518 			kmem_free(path, ifd->ifd_len);
519 			break;
520 		}
521 		rv = rumpuser_open(path,
522 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
523 		if (rv) {
524 			kmem_free(path, ifd->ifd_len);
525 			break;
526 		}
527 		rv = initbackend(sc, memfd);
528 		if (rv) {
529 			kmem_free(path, ifd->ifd_len);
530 			rumpuser_close(memfd);
531 			break;
532 		}
533 		sc->sc_backfile = path;
534 		sc->sc_backfilelen = ifd->ifd_len;
535 
536 		break;
537 	default:
538 		rv = ether_ioctl(ifp, cmd, data);
539 		if (rv == ENETRESET)
540 			rv = 0;
541 		break;
542 	}
543 	splx(s);
544 
545 	return rv;
546 }
547 
548 static void
549 shmif_start(struct ifnet *ifp)
550 {
551 	struct shmif_sc *sc = ifp->if_softc;
552 	struct mbuf *m, *n;
553 	bool wrote = false;
554 
555 	ifp->if_flags |= IFF_OACTIVE;
556 
557 	for (;;) {
558 		IF_DEQUEUE(&ifp->if_snd, m);
559 		if (m == NULL)
560 			break;
561 
562 		m = ether_sw_offload_tx(ifp, m);
563 		if (m == NULL) {
564 			if_statinc(ifp, if_oerrors);
565 			break;
566 		}
567 
568 		do {
569 			n = m->m_nextpkt;
570 			shmif_snd(ifp, m);
571 			m = n;
572 		} while (m != NULL);
573 
574 		wrote = true;
575 	}
576 
577 	ifp->if_flags &= ~IFF_OACTIVE;
578 
579 	/* wakeup? */
580 	if (wrote) {
581 		dowakeup(sc);
582 	}
583 }
584 
585 /* send everything in-context since it's just a matter of mem-to-mem copy */
586 static void
587 shmif_snd(struct ifnet *ifp, struct mbuf *m0)
588 {
589 	struct shmif_sc *sc = ifp->if_softc;
590 	struct shmif_mem *busmem = sc->sc_busmem;
591 	struct shmif_pkthdr sp;
592 	struct timeval tv;
593 	struct mbuf *m;
594 	uint32_t dataoff;
595 	uint32_t pktsize, pktwrote;
596 	bool wrap;
597 
598 	pktsize = 0;
599 	for (m = m0; m != NULL; m = m->m_next) {
600 		pktsize += m->m_len;
601 	}
602 	KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
603 
604 	getmicrouptime(&tv);
605 	sp.sp_len = pktsize;
606 	sp.sp_sec = tv.tv_sec;
607 	sp.sp_usec = tv.tv_usec;
608 	sp.sp_sender = sc->sc_uid;
609 
610 	bpf_mtap(ifp, m0, BPF_D_OUT);
611 
612 	shmif_lockbus(busmem);
613 	KASSERT(busmem->shm_magic == SHMIF_MAGIC);
614 	busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
615 
616 	wrap = false;
617 	dataoff =
618 	    shmif_buswrite(busmem, busmem->shm_last, &sp, sizeof(sp), &wrap);
619 	pktwrote = 0;
620 	for (m = m0; m != NULL; m = m->m_next) {
621 		pktwrote += m->m_len;
622 		dataoff = shmif_buswrite(busmem, dataoff, mtod(m, void *),
623 		    m->m_len, &wrap);
624 	}
625 	KASSERT(pktwrote == pktsize);
626 	if (wrap) {
627 		busmem->shm_gen++;
628 		DPRINTF(("bus generation now %" PRIu64 "\n", busmem->shm_gen));
629 	}
630 	shmif_unlockbus(busmem);
631 
632 	m_freem(m0);
633 	if_statinc(ifp, if_opackets);
634 
635 	DPRINTF(("shmif_start: send %d bytes at off %d\n", pktsize,
636 	    busmem->shm_last));
637 }
638 
639 static void
640 shmif_stop(struct ifnet *ifp, int disable)
641 {
642 	struct shmif_sc *sc = ifp->if_softc;
643 
644 	ifp->if_flags &= ~IFF_RUNNING;
645 	membar_producer();
646 
647 	/*
648 	 * wakeup thread.  this will of course wake up all bus
649 	 * listeners, but that's life.
650 	 */
651 	if (sc->sc_memfd != -1) {
652 		dowakeup(sc);
653 	}
654 }
655 
656 
657 /*
658  * Check if we have been sleeping too long.  Basically,
659  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
660  * We use the fact that first is guaranteed to never overlap
661  * with the last frame in the ring.
662  */
663 static __inline bool
664 stillvalid_p(struct shmif_sc *sc)
665 {
666 	struct shmif_mem *busmem = sc->sc_busmem;
667 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
668 	uint32_t lastoff, devoff;
669 
670 	KASSERT(busmem->shm_first != busmem->shm_last);
671 
672 	/* normalize onto a 2x busmem chunk */
673 	devoff = sc->sc_nextpacket;
674 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
675 
676 	/* trivial case */
677 	if (gendiff > 1)
678 		return false;
679 	KASSERT(gendiff <= 1);
680 
681 	/* Normalize onto 2x busmem chunk */
682 	if (busmem->shm_first >= lastoff) {
683 		lastoff += BUSMEM_DATASIZE;
684 		if (gendiff == 0)
685 			devoff += BUSMEM_DATASIZE;
686 	} else {
687 		if (gendiff)
688 			return false;
689 	}
690 
691 	return devoff >= busmem->shm_first && devoff <= lastoff;
692 }
693 
694 static void
695 shmif_rcv(void *arg)
696 {
697 	struct ifnet *ifp = arg;
698 	struct shmif_sc *sc = ifp->if_softc;
699 	struct shmif_mem *busmem;
700 	struct mbuf *m = NULL;
701 	struct ether_header *eth;
702 	uint32_t nextpkt;
703 	bool wrap, passup;
704 	int error;
705 	const int align
706 	    = ALIGN(sizeof(struct ether_header)) - sizeof(struct ether_header);
707 
708  reup:
709 	mutex_enter(&sc->sc_mtx);
710 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
711 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
712 	mutex_exit(&sc->sc_mtx);
713 
714 	busmem = sc->sc_busmem;
715 
716 	while (ifp->if_flags & IFF_RUNNING) {
717 		struct shmif_pkthdr sp;
718 
719 		if (m == NULL) {
720 			m = m_gethdr(M_WAIT, MT_DATA);
721 			MCLGET(m, M_WAIT);
722 			m->m_data += align;
723 		}
724 
725 		DPRINTF(("waiting %d/%" PRIu64 "\n",
726 		    sc->sc_nextpacket, sc->sc_devgen));
727 		KASSERT(m->m_flags & M_EXT);
728 
729 		shmif_lockbus(busmem);
730 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
731 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
732 
733 		/* need more data? */
734 		if (sc->sc_devgen == busmem->shm_gen &&
735 		    shmif_nextpktoff(busmem, busmem->shm_last)
736 		     == sc->sc_nextpacket) {
737 			shmif_unlockbus(busmem);
738 			error = rumpcomp_shmif_watchwait(sc->sc_kq);
739 			if (__predict_false(error))
740 				printf("shmif_rcv: wait failed %d\n", error);
741 			membar_consumer();
742 			continue;
743 		}
744 
745 		if (stillvalid_p(sc)) {
746 			nextpkt = sc->sc_nextpacket;
747 		} else {
748 			KASSERT(busmem->shm_gen > 0);
749 			nextpkt = busmem->shm_first;
750 			if (busmem->shm_first > busmem->shm_last)
751 				sc->sc_devgen = busmem->shm_gen - 1;
752 			else
753 				sc->sc_devgen = busmem->shm_gen;
754 			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
755 			    sc, nextpkt, sc->sc_devgen));
756 		}
757 
758 		/*
759 		 * If our read pointer is ahead the bus last write, our
760 		 * generation must be one behind.
761 		 */
762 		KASSERT(!(nextpkt > busmem->shm_last
763 		    && sc->sc_devgen == busmem->shm_gen));
764 
765 		wrap = false;
766 		nextpkt = shmif_busread(busmem, &sp,
767 		    nextpkt, sizeof(sp), &wrap);
768 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
769 		nextpkt = shmif_busread(busmem, mtod(m, void *),
770 		    nextpkt, sp.sp_len, &wrap);
771 
772 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
773 		    sp.sp_len, nextpkt));
774 
775 		sc->sc_nextpacket = nextpkt;
776 		shmif_unlockbus(sc->sc_busmem);
777 
778 		if (wrap) {
779 			sc->sc_devgen++;
780 			DPRINTF(("dev %p generation now %" PRIu64 "\n",
781 			    sc, sc->sc_devgen));
782 		}
783 
784 		/*
785 		 * Ignore packets too short to possibly be valid.
786 		 * This is hit at least for the first frame on a new bus.
787 		 */
788 		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
789 			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
790 			    sp.sp_len));
791 			continue;
792 		}
793 
794 		m->m_len = m->m_pkthdr.len = sp.sp_len;
795 		m_set_rcvif(m, ifp);
796 
797 		/*
798 		 * Test if we want to pass the packet upwards
799 		 */
800 		eth = mtod(m, struct ether_header *);
801 		if (sp.sp_sender == sc->sc_uid) {
802 			passup = false;
803 		} else if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
804 		    ETHER_ADDR_LEN) == 0) {
805 			passup = true;
806 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
807 			passup = true;
808 		} else if (ifp->if_flags & IFF_PROMISC) {
809 			m->m_flags |= M_PROMISC;
810 			passup = true;
811 		} else {
812 			passup = false;
813 		}
814 
815 		if (passup) {
816 			int bound;
817 
818 			m = ether_sw_offload_rx(ifp, m);
819 
820 			KERNEL_LOCK(1, NULL);
821 			/* Prevent LWP migrations between CPUs for psref(9) */
822 			bound = curlwp_bind();
823 			if_input(ifp, m);
824 			curlwp_bindx(bound);
825 			KERNEL_UNLOCK_ONE(NULL);
826 
827 			m = NULL;
828 		}
829 		/* else: reuse mbuf for a future packet */
830 	}
831 	m_freem(m);
832 	m = NULL;
833 
834 	if (!sc->sc_dying)
835 		goto reup;
836 
837 	kthread_exit(0);
838 }
839