Home | History | Annotate | Line # | Download | only in libshmif
if_shmem.c revision 1.62.2.1
      1 /*	$NetBSD: if_shmem.c,v 1.62.2.1 2014/08/17 03:26:51 riz 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 1.62.2.1 2014/08/17 03:26:51 riz 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 
     47 #include <netinet/in.h>
     48 #include <netinet/in_var.h>
     49 
     50 #include <rump/rump.h>
     51 #include <rump/rumpuser.h>
     52 
     53 #include "rump_private.h"
     54 #include "rump_net_private.h"
     55 #include "shmif_user.h"
     56 
     57 static int shmif_clone(struct if_clone *, int);
     58 static int shmif_unclone(struct ifnet *);
     59 
     60 struct if_clone shmif_cloner =
     61     IF_CLONE_INITIALIZER("shmif", shmif_clone, shmif_unclone);
     62 
     63 /*
     64  * Do r/w prefault for backend pages when attaching the interface.
     65  * At least logically thinking improves performance (although no
     66  * mlocking is done, so they might go away).
     67  */
     68 #define PREFAULT_RW
     69 
     70 /*
     71  * A virtual ethernet interface which uses shared memory from a
     72  * memory mapped file as the bus.
     73  */
     74 
     75 static int	shmif_init(struct ifnet *);
     76 static int	shmif_ioctl(struct ifnet *, u_long, void *);
     77 static void	shmif_start(struct ifnet *);
     78 static void	shmif_stop(struct ifnet *, int);
     79 
     80 #include "shmifvar.h"
     81 
     82 struct shmif_sc {
     83 	struct ethercom sc_ec;
     84 	struct shmif_mem *sc_busmem;
     85 	int sc_memfd;
     86 	int sc_kq;
     87 	int sc_unit;
     88 
     89 	char *sc_backfile;
     90 	size_t sc_backfilelen;
     91 
     92 	uint64_t sc_devgen;
     93 	uint32_t sc_nextpacket;
     94 
     95 	kmutex_t sc_mtx;
     96 	kcondvar_t sc_cv;
     97 
     98 	struct lwp *sc_rcvl;
     99 	bool sc_dying;
    100 
    101 	uint64_t sc_uuid;
    102 };
    103 
    104 static void shmif_rcv(void *);
    105 
    106 #define LOCK_UNLOCKED	0
    107 #define LOCK_LOCKED	1
    108 #define LOCK_COOLDOWN	1001
    109 
    110 vmem_t *shmif_units;
    111 
    112 static void
    113 dowakeup(struct shmif_sc *sc)
    114 {
    115 	struct rumpuser_iovec iov;
    116 	uint32_t ver = SHMIF_VERSION;
    117 	size_t n;
    118 
    119 	iov.iov_base = &ver;
    120 	iov.iov_len = sizeof(ver);
    121 	rumpuser_iovwrite(sc->sc_memfd, &iov, 1, IFMEM_WAKEUP, &n);
    122 }
    123 
    124 /*
    125  * This locking needs work and will misbehave severely if:
    126  * 1) the backing memory has to be paged in
    127  * 2) some lockholder exits while holding the lock
    128  */
    129 static void
    130 shmif_lockbus(struct shmif_mem *busmem)
    131 {
    132 	int i = 0;
    133 
    134 	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
    135 	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
    136 		if (__predict_false(++i > LOCK_COOLDOWN)) {
    137 			/* wait 1ms */
    138 			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL,
    139 			    0, 1000*1000);
    140 			i = 0;
    141 		}
    142 		continue;
    143 	}
    144 	membar_enter();
    145 }
    146 
    147 static void
    148 shmif_unlockbus(struct shmif_mem *busmem)
    149 {
    150 	unsigned int old __diagused;
    151 
    152 	membar_exit();
    153 	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
    154 	KASSERT(old == LOCK_LOCKED);
    155 }
    156 
    157 static int
    158 allocif(int unit, struct shmif_sc **scp)
    159 {
    160 	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
    161 	struct shmif_sc *sc;
    162 	struct ifnet *ifp;
    163 	uint32_t randnum;
    164 	int error;
    165 
    166 	randnum = cprng_fast32();
    167 	memcpy(&enaddr[2], &randnum, sizeof(randnum));
    168 
    169 	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
    170 	sc->sc_memfd = -1;
    171 	sc->sc_unit = unit;
    172 	sc->sc_uuid = cprng_fast64();
    173 
    174 	ifp = &sc->sc_ec.ec_if;
    175 
    176 	snprintf(ifp->if_xname, sizeof(ifp->if_xname), "shmif%d", unit);
    177 	ifp->if_softc = sc;
    178 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
    179 	ifp->if_init = shmif_init;
    180 	ifp->if_ioctl = shmif_ioctl;
    181 	ifp->if_start = shmif_start;
    182 	ifp->if_stop = shmif_stop;
    183 	ifp->if_mtu = ETHERMTU;
    184 	ifp->if_dlt = DLT_EN10MB;
    185 
    186 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
    187 	cv_init(&sc->sc_cv, "shmifcv");
    188 
    189 	if_attach(ifp);
    190 	ether_ifattach(ifp, enaddr);
    191 
    192 	aprint_verbose("shmif%d: Ethernet address %s\n",
    193 	    unit, ether_sprintf(enaddr));
    194 
    195 	if (scp)
    196 		*scp = sc;
    197 
    198 	error = 0;
    199 	if (rump_threads) {
    200 		error = kthread_create(PRI_NONE,
    201 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
    202 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
    203 	} else {
    204 		printf("WARNING: threads not enabled, shmif NOT working\n");
    205 	}
    206 
    207 	if (error) {
    208 		shmif_unclone(ifp);
    209 	}
    210 
    211 	return error;
    212 }
    213 
    214 static int
    215 initbackend(struct shmif_sc *sc, int memfd)
    216 {
    217 	volatile uint8_t v;
    218 	volatile uint8_t *p;
    219 	void *mem;
    220 	int error;
    221 
    222 	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
    223 	if (error)
    224 		return error;
    225 	sc->sc_busmem = mem;
    226 
    227 	if (sc->sc_busmem->shm_magic
    228 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
    229 		printf("bus is not magical");
    230 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    231 		return ENOEXEC;
    232 	}
    233 
    234 	/*
    235 	 * Prefault in pages to minimize runtime penalty with buslock.
    236 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
    237 	 * rump kernel PAGE_SIZE > host page size.
    238 	 */
    239 	for (p = (uint8_t *)sc->sc_busmem;
    240 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    241 	    p += 512)
    242 		v = *p;
    243 
    244 	shmif_lockbus(sc->sc_busmem);
    245 	/* we're first?  initialize bus */
    246 	if (sc->sc_busmem->shm_magic == 0) {
    247 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
    248 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
    249 	}
    250 
    251 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    252 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    253 
    254 #ifdef PREFAULT_RW
    255 	for (p = (uint8_t *)sc->sc_busmem;
    256 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    257 	    p += PAGE_SIZE) {
    258 		v = *p;
    259 		*p = v;
    260 	}
    261 #endif
    262 	shmif_unlockbus(sc->sc_busmem);
    263 
    264 	sc->sc_kq = -1;
    265 	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
    266 	if (error) {
    267 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    268 		return error;
    269 	}
    270 
    271 	sc->sc_memfd = memfd;
    272 
    273 	return error;
    274 }
    275 
    276 static void
    277 finibackend(struct shmif_sc *sc)
    278 {
    279 
    280 	if (sc->sc_backfile == NULL)
    281 		return;
    282 
    283 	if (sc->sc_backfile) {
    284 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
    285 		sc->sc_backfile = NULL;
    286 		sc->sc_backfilelen = 0;
    287 	}
    288 
    289 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    290 	rumpuser_close(sc->sc_memfd);
    291 	rumpuser_close(sc->sc_kq);
    292 
    293 	sc->sc_memfd = -1;
    294 }
    295 
    296 int
    297 rump_shmif_create(const char *path, int *ifnum)
    298 {
    299 	struct shmif_sc *sc;
    300 	vmem_addr_t t;
    301 	int unit, error;
    302 	int memfd = -1; /* XXXgcc */
    303 
    304 	if (path) {
    305 		error = rumpuser_open(path,
    306 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    307 		if (error)
    308 			return error;
    309 	}
    310 
    311 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
    312 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
    313 
    314 	if (error != 0) {
    315 		if (path)
    316 			rumpuser_close(memfd);
    317 		return error;
    318 	}
    319 
    320 	unit = t - 1;
    321 
    322 	if ((error = allocif(unit, &sc)) != 0) {
    323 		if (path)
    324 			rumpuser_close(memfd);
    325 		return error;
    326 	}
    327 
    328 	if (!path)
    329 		goto out;
    330 
    331 	error = initbackend(sc, memfd);
    332 	if (error) {
    333 		shmif_unclone(&sc->sc_ec.ec_if);
    334 		return error;
    335 	}
    336 
    337 	sc->sc_backfilelen = strlen(path)+1;
    338 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
    339 	strcpy(sc->sc_backfile, path);
    340 
    341  out:
    342 	if (ifnum)
    343 		*ifnum = unit;
    344 
    345 	return 0;
    346 }
    347 
    348 static int
    349 shmif_clone(struct if_clone *ifc, int unit)
    350 {
    351 	int rc __diagused;
    352 	vmem_addr_t unit2;
    353 
    354 	/*
    355 	 * Ok, we know the unit number, but we must still reserve it.
    356 	 * Otherwise the wildcard-side of things might get the same one.
    357 	 * This is slightly offset-happy due to vmem.  First, we offset
    358 	 * the range of unit numbers by +1 since vmem cannot deal with
    359 	 * ranges starting from 0.  Talk about uuuh.
    360 	 */
    361 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
    362 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
    363 	KASSERT(rc == 0 && unit2-1 == unit);
    364 
    365 	return allocif(unit, NULL);
    366 }
    367 
    368 static int
    369 shmif_unclone(struct ifnet *ifp)
    370 {
    371 	struct shmif_sc *sc = ifp->if_softc;
    372 
    373 	shmif_stop(ifp, 1);
    374 	if_down(ifp);
    375 	finibackend(sc);
    376 
    377 	mutex_enter(&sc->sc_mtx);
    378 	sc->sc_dying = true;
    379 	cv_broadcast(&sc->sc_cv);
    380 	mutex_exit(&sc->sc_mtx);
    381 
    382 	if (sc->sc_rcvl)
    383 		kthread_join(sc->sc_rcvl);
    384 	sc->sc_rcvl = NULL;
    385 
    386 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
    387 
    388 	ether_ifdetach(ifp);
    389 	if_detach(ifp);
    390 
    391 	cv_destroy(&sc->sc_cv);
    392 	mutex_destroy(&sc->sc_mtx);
    393 
    394 	kmem_free(sc, sizeof(*sc));
    395 
    396 	return 0;
    397 }
    398 
    399 static int
    400 shmif_init(struct ifnet *ifp)
    401 {
    402 	struct shmif_sc *sc = ifp->if_softc;
    403 	int error = 0;
    404 
    405 	if (sc->sc_memfd == -1)
    406 		return ENXIO;
    407 	KASSERT(sc->sc_busmem);
    408 
    409 	ifp->if_flags |= IFF_RUNNING;
    410 
    411 	mutex_enter(&sc->sc_mtx);
    412 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    413 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    414 
    415 	cv_broadcast(&sc->sc_cv);
    416 	mutex_exit(&sc->sc_mtx);
    417 
    418 	return error;
    419 }
    420 
    421 static int
    422 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
    423 {
    424 	struct shmif_sc *sc = ifp->if_softc;
    425 	struct ifdrv *ifd;
    426 	char *path;
    427 	int s, rv, memfd;
    428 
    429 	s = splnet();
    430 	switch (cmd) {
    431 	case SIOCGLINKSTR:
    432 		ifd = data;
    433 
    434 		if (sc->sc_backfilelen == 0) {
    435 			rv = ENOENT;
    436 			break;
    437 		}
    438 
    439 		ifd->ifd_len = sc->sc_backfilelen;
    440 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
    441 			rv = 0;
    442 			break;
    443 		}
    444 
    445 		if (ifd->ifd_cmd != 0) {
    446 			rv = EINVAL;
    447 			break;
    448 		}
    449 
    450 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
    451 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
    452 		break;
    453 	case SIOCSLINKSTR:
    454 		if (ifp->if_flags & IFF_UP) {
    455 			rv = EBUSY;
    456 			break;
    457 		}
    458 
    459 		ifd = data;
    460 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
    461 			finibackend(sc);
    462 			rv = 0;
    463 			break;
    464 		} else if (ifd->ifd_cmd != 0) {
    465 			rv = EINVAL;
    466 			break;
    467 		} else if (sc->sc_backfile) {
    468 			rv = EBUSY;
    469 			break;
    470 		}
    471 
    472 		if (ifd->ifd_len > MAXPATHLEN) {
    473 			rv = E2BIG;
    474 			break;
    475 		} else if (ifd->ifd_len < 1) {
    476 			rv = EINVAL;
    477 			break;
    478 		}
    479 
    480 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
    481 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
    482 		if (rv) {
    483 			kmem_free(path, ifd->ifd_len);
    484 			break;
    485 		}
    486 		rv = rumpuser_open(path,
    487 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    488 		if (rv) {
    489 			kmem_free(path, ifd->ifd_len);
    490 			break;
    491 		}
    492 		rv = initbackend(sc, memfd);
    493 		if (rv) {
    494 			kmem_free(path, ifd->ifd_len);
    495 			rumpuser_close(memfd);
    496 			break;
    497 		}
    498 		sc->sc_backfile = path;
    499 		sc->sc_backfilelen = ifd->ifd_len;
    500 
    501 		break;
    502 	default:
    503 		rv = ether_ioctl(ifp, cmd, data);
    504 		if (rv == ENETRESET)
    505 			rv = 0;
    506 		break;
    507 	}
    508 	splx(s);
    509 
    510 	return rv;
    511 }
    512 
    513 /* send everything in-context since it's just a matter of mem-to-mem copy */
    514 static void
    515 shmif_start(struct ifnet *ifp)
    516 {
    517 	struct shmif_sc *sc = ifp->if_softc;
    518 	struct shmif_mem *busmem = sc->sc_busmem;
    519 	struct mbuf *m, *m0;
    520 	uint32_t dataoff;
    521 	uint32_t pktsize, pktwrote;
    522 	bool wrote = false;
    523 	bool wrap;
    524 
    525 	ifp->if_flags |= IFF_OACTIVE;
    526 
    527 	for (;;) {
    528 		struct shmif_pkthdr sp;
    529 		struct timeval tv;
    530 
    531 		IF_DEQUEUE(&ifp->if_snd, m0);
    532 		if (m0 == NULL) {
    533 			break;
    534 		}
    535 
    536 		pktsize = 0;
    537 		for (m = m0; m != NULL; m = m->m_next) {
    538 			pktsize += m->m_len;
    539 		}
    540 		KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
    541 
    542 		getmicrouptime(&tv);
    543 		sp.sp_len = pktsize;
    544 		sp.sp_sec = tv.tv_sec;
    545 		sp.sp_usec = tv.tv_usec;
    546 		sp.sp_sender = sc->sc_uuid;
    547 
    548 		bpf_mtap(ifp, m0);
    549 
    550 		shmif_lockbus(busmem);
    551 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    552 		busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
    553 
    554 		wrap = false;
    555 		dataoff = shmif_buswrite(busmem,
    556 		    busmem->shm_last, &sp, sizeof(sp), &wrap);
    557 		pktwrote = 0;
    558 		for (m = m0; m != NULL; m = m->m_next) {
    559 			pktwrote += m->m_len;
    560 			dataoff = shmif_buswrite(busmem, dataoff,
    561 			    mtod(m, void *), m->m_len, &wrap);
    562 		}
    563 		KASSERT(pktwrote == pktsize);
    564 		if (wrap) {
    565 			busmem->shm_gen++;
    566 			DPRINTF(("bus generation now %" PRIu64 "\n",
    567 			    busmem->shm_gen));
    568 		}
    569 		shmif_unlockbus(busmem);
    570 
    571 		m_freem(m0);
    572 		wrote = true;
    573 		ifp->if_opackets++;
    574 
    575 		DPRINTF(("shmif_start: send %d bytes at off %d\n",
    576 		    pktsize, busmem->shm_last));
    577 	}
    578 
    579 	ifp->if_flags &= ~IFF_OACTIVE;
    580 
    581 	/* wakeup? */
    582 	if (wrote) {
    583 		dowakeup(sc);
    584 	}
    585 }
    586 
    587 static void
    588 shmif_stop(struct ifnet *ifp, int disable)
    589 {
    590 	struct shmif_sc *sc = ifp->if_softc;
    591 
    592 	ifp->if_flags &= ~IFF_RUNNING;
    593 	membar_producer();
    594 
    595 	/*
    596 	 * wakeup thread.  this will of course wake up all bus
    597 	 * listeners, but that's life.
    598 	 */
    599 	if (sc->sc_memfd != -1) {
    600 		dowakeup(sc);
    601 	}
    602 }
    603 
    604 
    605 /*
    606  * Check if we have been sleeping too long.  Basically,
    607  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
    608  * We use the fact that first is guaranteed to never overlap
    609  * with the last frame in the ring.
    610  */
    611 static __inline bool
    612 stillvalid_p(struct shmif_sc *sc)
    613 {
    614 	struct shmif_mem *busmem = sc->sc_busmem;
    615 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
    616 	uint32_t lastoff, devoff;
    617 
    618 	KASSERT(busmem->shm_first != busmem->shm_last);
    619 
    620 	/* normalize onto a 2x busmem chunk */
    621 	devoff = sc->sc_nextpacket;
    622 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
    623 
    624 	/* trivial case */
    625 	if (gendiff > 1)
    626 		return false;
    627 	KASSERT(gendiff <= 1);
    628 
    629 	/* Normalize onto 2x busmem chunk */
    630 	if (busmem->shm_first >= lastoff) {
    631 		lastoff += BUSMEM_DATASIZE;
    632 		if (gendiff == 0)
    633 			devoff += BUSMEM_DATASIZE;
    634 	} else {
    635 		if (gendiff)
    636 			return false;
    637 	}
    638 
    639 	return devoff >= busmem->shm_first && devoff <= lastoff;
    640 }
    641 
    642 static void
    643 shmif_rcv(void *arg)
    644 {
    645 	struct ifnet *ifp = arg;
    646 	struct shmif_sc *sc = ifp->if_softc;
    647 	struct shmif_mem *busmem;
    648 	struct mbuf *m = NULL;
    649 	struct ether_header *eth;
    650 	uint32_t nextpkt;
    651 	bool wrap, passup;
    652 	int error;
    653 	const int align
    654 	    = ALIGN(sizeof(struct ether_header)) - sizeof(struct ether_header);
    655 
    656  reup:
    657 	mutex_enter(&sc->sc_mtx);
    658 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
    659 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
    660 	mutex_exit(&sc->sc_mtx);
    661 
    662 	busmem = sc->sc_busmem;
    663 
    664 	while (ifp->if_flags & IFF_RUNNING) {
    665 		struct shmif_pkthdr sp;
    666 
    667 		if (m == NULL) {
    668 			m = m_gethdr(M_WAIT, MT_DATA);
    669 			MCLGET(m, M_WAIT);
    670 			m->m_data += align;
    671 		}
    672 
    673 		DPRINTF(("waiting %d/%" PRIu64 "\n",
    674 		    sc->sc_nextpacket, sc->sc_devgen));
    675 		KASSERT(m->m_flags & M_EXT);
    676 
    677 		shmif_lockbus(busmem);
    678 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    679 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
    680 
    681 		/* need more data? */
    682 		if (sc->sc_devgen == busmem->shm_gen &&
    683 		    shmif_nextpktoff(busmem, busmem->shm_last)
    684 		     == sc->sc_nextpacket) {
    685 			shmif_unlockbus(busmem);
    686 			error = 0;
    687 			rumpcomp_shmif_watchwait(sc->sc_kq);
    688 			if (__predict_false(error))
    689 				printf("shmif_rcv: wait failed %d\n", error);
    690 			membar_consumer();
    691 			continue;
    692 		}
    693 
    694 		if (stillvalid_p(sc)) {
    695 			nextpkt = sc->sc_nextpacket;
    696 		} else {
    697 			KASSERT(busmem->shm_gen > 0);
    698 			nextpkt = busmem->shm_first;
    699 			if (busmem->shm_first > busmem->shm_last)
    700 				sc->sc_devgen = busmem->shm_gen - 1;
    701 			else
    702 				sc->sc_devgen = busmem->shm_gen;
    703 			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
    704 			    sc, nextpkt, sc->sc_devgen));
    705 		}
    706 
    707 		/*
    708 		 * If our read pointer is ahead the bus last write, our
    709 		 * generation must be one behind.
    710 		 */
    711 		KASSERT(!(nextpkt > busmem->shm_last
    712 		    && sc->sc_devgen == busmem->shm_gen));
    713 
    714 		wrap = false;
    715 		nextpkt = shmif_busread(busmem, &sp,
    716 		    nextpkt, sizeof(sp), &wrap);
    717 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
    718 		nextpkt = shmif_busread(busmem, mtod(m, void *),
    719 		    nextpkt, sp.sp_len, &wrap);
    720 
    721 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
    722 		    sp.sp_len, nextpkt));
    723 
    724 		sc->sc_nextpacket = nextpkt;
    725 		shmif_unlockbus(sc->sc_busmem);
    726 
    727 		if (wrap) {
    728 			sc->sc_devgen++;
    729 			DPRINTF(("dev %p generation now %" PRIu64 "\n",
    730 			    sc, sc->sc_devgen));
    731 		}
    732 
    733 		/*
    734 		 * Ignore packets too short to possibly be valid.
    735 		 * This is hit at least for the first frame on a new bus.
    736 		 */
    737 		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
    738 			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
    739 			    sp.sp_len));
    740 			continue;
    741 		}
    742 
    743 		m->m_len = m->m_pkthdr.len = sp.sp_len;
    744 		m->m_pkthdr.rcvif = ifp;
    745 
    746 		/*
    747 		 * Test if we want to pass the packet upwards
    748 		 */
    749 		eth = mtod(m, struct ether_header *);
    750 		if (sp.sp_sender == sc->sc_uuid) {
    751 			passup = false;
    752 		} else if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
    753 		    ETHER_ADDR_LEN) == 0) {
    754 			passup = true;
    755 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
    756 			passup = true;
    757 		} else if (ifp->if_flags & IFF_PROMISC) {
    758 			m->m_flags |= M_PROMISC;
    759 			passup = true;
    760 		} else {
    761 			passup = false;
    762 		}
    763 
    764 		if (passup) {
    765 			ifp->if_ipackets++;
    766 			KERNEL_LOCK(1, NULL);
    767 			bpf_mtap(ifp, m);
    768 			ifp->if_input(ifp, m);
    769 			KERNEL_UNLOCK_ONE(NULL);
    770 			m = NULL;
    771 		}
    772 		/* else: reuse mbuf for a future packet */
    773 	}
    774 	m_freem(m);
    775 	m = NULL;
    776 
    777 	if (!sc->sc_dying)
    778 		goto reup;
    779 
    780 	kthread_exit(0);
    781 }
    782