Home | History | Annotate | Line # | Download | only in libshmif
if_shmem.c revision 1.60
      1 /*	$NetBSD: if_shmem.c,v 1.60 2014/03/20 20:42:08 christos 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 1.60 2014/03/20 20:42:08 christos 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 
    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 static void
    111 dowakeup(struct shmif_sc *sc)
    112 {
    113 	struct rumpuser_iovec iov;
    114 	uint32_t ver = SHMIF_VERSION;
    115 	size_t n;
    116 
    117 	iov.iov_base = &ver;
    118 	iov.iov_len = sizeof(ver);
    119 	rumpuser_iovwrite(sc->sc_memfd, &iov, 1, IFMEM_WAKEUP, &n);
    120 }
    121 
    122 /*
    123  * This locking needs work and will misbehave severely if:
    124  * 1) the backing memory has to be paged in
    125  * 2) some lockholder exits while holding the lock
    126  */
    127 static void
    128 shmif_lockbus(struct shmif_mem *busmem)
    129 {
    130 	int i = 0;
    131 
    132 	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
    133 	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
    134 		if (__predict_false(++i > LOCK_COOLDOWN)) {
    135 			/* wait 1ms */
    136 			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL,
    137 			    0, 1000*1000);
    138 			i = 0;
    139 		}
    140 		continue;
    141 	}
    142 	membar_enter();
    143 }
    144 
    145 static void
    146 shmif_unlockbus(struct shmif_mem *busmem)
    147 {
    148 	unsigned int old;
    149 
    150 	membar_exit();
    151 	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
    152 	KASSERT(old == LOCK_LOCKED);
    153 }
    154 
    155 static int
    156 allocif(int unit, struct shmif_sc **scp)
    157 {
    158 	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
    159 	struct shmif_sc *sc;
    160 	struct ifnet *ifp;
    161 	uint32_t randnum;
    162 	int error;
    163 
    164 	randnum = cprng_fast32();
    165 	memcpy(&enaddr[2], &randnum, sizeof(randnum));
    166 
    167 	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
    168 	sc->sc_memfd = -1;
    169 	sc->sc_unit = unit;
    170 
    171 	ifp = &sc->sc_ec.ec_if;
    172 
    173 	snprintf(ifp->if_xname, sizeof(ifp->if_xname), "shmif%d", unit);
    174 	ifp->if_softc = sc;
    175 	ifp->if_flags = IFF_BROADCAST | IFF_MULTICAST;
    176 	ifp->if_init = shmif_init;
    177 	ifp->if_ioctl = shmif_ioctl;
    178 	ifp->if_start = shmif_start;
    179 	ifp->if_stop = shmif_stop;
    180 	ifp->if_mtu = ETHERMTU;
    181 	ifp->if_dlt = DLT_EN10MB;
    182 
    183 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
    184 	cv_init(&sc->sc_cv, "shmifcv");
    185 
    186 	if_attach(ifp);
    187 	ether_ifattach(ifp, enaddr);
    188 
    189 	aprint_verbose("shmif%d: Ethernet address %s\n",
    190 	    unit, ether_sprintf(enaddr));
    191 
    192 	if (scp)
    193 		*scp = sc;
    194 
    195 	error = 0;
    196 	if (rump_threads) {
    197 		error = kthread_create(PRI_NONE,
    198 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
    199 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
    200 	} else {
    201 		printf("WARNING: threads not enabled, shmif NOT working\n");
    202 	}
    203 
    204 	if (error) {
    205 		shmif_unclone(ifp);
    206 	}
    207 
    208 	return error;
    209 }
    210 
    211 static int
    212 initbackend(struct shmif_sc *sc, int memfd)
    213 {
    214 	volatile uint8_t v;
    215 	volatile uint8_t *p;
    216 	void *mem;
    217 	int error;
    218 
    219 	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
    220 	if (error)
    221 		return error;
    222 	sc->sc_busmem = mem;
    223 
    224 	if (sc->sc_busmem->shm_magic
    225 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
    226 		printf("bus is not magical");
    227 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    228 		return ENOEXEC;
    229 	}
    230 
    231 	/*
    232 	 * Prefault in pages to minimize runtime penalty with buslock.
    233 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
    234 	 * rump kernel PAGE_SIZE > host page size.
    235 	 */
    236 	for (p = (uint8_t *)sc->sc_busmem;
    237 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    238 	    p += 512)
    239 		v = *p;
    240 
    241 	shmif_lockbus(sc->sc_busmem);
    242 	/* we're first?  initialize bus */
    243 	if (sc->sc_busmem->shm_magic == 0) {
    244 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
    245 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
    246 	}
    247 
    248 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    249 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    250 
    251 #ifdef PREFAULT_RW
    252 	for (p = (uint8_t *)sc->sc_busmem;
    253 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    254 	    p += PAGE_SIZE) {
    255 		v = *p;
    256 		*p = v;
    257 	}
    258 #endif
    259 	shmif_unlockbus(sc->sc_busmem);
    260 
    261 	sc->sc_kq = -1;
    262 	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
    263 	if (error) {
    264 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    265 		return error;
    266 	}
    267 
    268 	sc->sc_memfd = memfd;
    269 
    270 	return error;
    271 }
    272 
    273 static void
    274 finibackend(struct shmif_sc *sc)
    275 {
    276 
    277 	if (sc->sc_backfile == NULL)
    278 		return;
    279 
    280 	if (sc->sc_backfile) {
    281 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
    282 		sc->sc_backfile = NULL;
    283 		sc->sc_backfilelen = 0;
    284 	}
    285 
    286 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    287 	rumpuser_close(sc->sc_memfd);
    288 	rumpuser_close(sc->sc_kq);
    289 
    290 	sc->sc_memfd = -1;
    291 }
    292 
    293 int
    294 rump_shmif_create(const char *path, int *ifnum)
    295 {
    296 	struct shmif_sc *sc;
    297 	vmem_addr_t t;
    298 	int unit, error;
    299 	int memfd = -1; /* XXXgcc */
    300 
    301 	if (path) {
    302 		error = rumpuser_open(path,
    303 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    304 		if (error)
    305 			return error;
    306 	}
    307 
    308 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
    309 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
    310 
    311 	if (error != 0) {
    312 		if (path)
    313 			rumpuser_close(memfd);
    314 		return error;
    315 	}
    316 
    317 	unit = t - 1;
    318 
    319 	if ((error = allocif(unit, &sc)) != 0) {
    320 		if (path)
    321 			rumpuser_close(memfd);
    322 		return error;
    323 	}
    324 
    325 	if (!path)
    326 		goto out;
    327 
    328 	error = initbackend(sc, memfd);
    329 	if (error) {
    330 		shmif_unclone(&sc->sc_ec.ec_if);
    331 		return error;
    332 	}
    333 
    334 	sc->sc_backfilelen = strlen(path)+1;
    335 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
    336 	strcpy(sc->sc_backfile, path);
    337 
    338  out:
    339 	if (ifnum)
    340 		*ifnum = unit;
    341 
    342 	return 0;
    343 }
    344 
    345 static int
    346 shmif_clone(struct if_clone *ifc, int unit)
    347 {
    348 	int rc;
    349 	vmem_addr_t unit2;
    350 
    351 	/*
    352 	 * Ok, we know the unit number, but we must still reserve it.
    353 	 * Otherwise the wildcard-side of things might get the same one.
    354 	 * This is slightly offset-happy due to vmem.  First, we offset
    355 	 * the range of unit numbers by +1 since vmem cannot deal with
    356 	 * ranges starting from 0.  Talk about uuuh.
    357 	 */
    358 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
    359 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
    360 	KASSERT(rc == 0 && unit2-1 == unit);
    361 
    362 	return allocif(unit, NULL);
    363 }
    364 
    365 static int
    366 shmif_unclone(struct ifnet *ifp)
    367 {
    368 	struct shmif_sc *sc = ifp->if_softc;
    369 
    370 	shmif_stop(ifp, 1);
    371 	if_down(ifp);
    372 	finibackend(sc);
    373 
    374 	mutex_enter(&sc->sc_mtx);
    375 	sc->sc_dying = true;
    376 	cv_broadcast(&sc->sc_cv);
    377 	mutex_exit(&sc->sc_mtx);
    378 
    379 	if (sc->sc_rcvl)
    380 		kthread_join(sc->sc_rcvl);
    381 	sc->sc_rcvl = NULL;
    382 
    383 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
    384 
    385 	ether_ifdetach(ifp);
    386 	if_detach(ifp);
    387 
    388 	cv_destroy(&sc->sc_cv);
    389 	mutex_destroy(&sc->sc_mtx);
    390 
    391 	kmem_free(sc, sizeof(*sc));
    392 
    393 	return 0;
    394 }
    395 
    396 static int
    397 shmif_init(struct ifnet *ifp)
    398 {
    399 	struct shmif_sc *sc = ifp->if_softc;
    400 	int error = 0;
    401 
    402 	if (sc->sc_memfd == -1)
    403 		return ENXIO;
    404 	KASSERT(sc->sc_busmem);
    405 
    406 	ifp->if_flags |= IFF_RUNNING;
    407 
    408 	mutex_enter(&sc->sc_mtx);
    409 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    410 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    411 
    412 	cv_broadcast(&sc->sc_cv);
    413 	mutex_exit(&sc->sc_mtx);
    414 
    415 	return error;
    416 }
    417 
    418 static int
    419 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
    420 {
    421 	struct shmif_sc *sc = ifp->if_softc;
    422 	struct ifdrv *ifd;
    423 	char *path;
    424 	int s, rv, memfd;
    425 
    426 	s = splnet();
    427 	switch (cmd) {
    428 	case SIOCGLINKSTR:
    429 		ifd = data;
    430 
    431 		if (sc->sc_backfilelen == 0) {
    432 			rv = ENOENT;
    433 			break;
    434 		}
    435 
    436 		ifd->ifd_len = sc->sc_backfilelen;
    437 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
    438 			rv = 0;
    439 			break;
    440 		}
    441 
    442 		if (ifd->ifd_cmd != 0) {
    443 			rv = EINVAL;
    444 			break;
    445 		}
    446 
    447 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
    448 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
    449 		break;
    450 	case SIOCSLINKSTR:
    451 		if (ifp->if_flags & IFF_UP) {
    452 			rv = EBUSY;
    453 			break;
    454 		}
    455 
    456 		ifd = data;
    457 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
    458 			finibackend(sc);
    459 			rv = 0;
    460 			break;
    461 		} else if (ifd->ifd_cmd != 0) {
    462 			rv = EINVAL;
    463 			break;
    464 		} else if (sc->sc_backfile) {
    465 			rv = EBUSY;
    466 			break;
    467 		}
    468 
    469 		if (ifd->ifd_len > MAXPATHLEN) {
    470 			rv = E2BIG;
    471 			break;
    472 		} else if (ifd->ifd_len < 1) {
    473 			rv = EINVAL;
    474 			break;
    475 		}
    476 
    477 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
    478 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
    479 		if (rv) {
    480 			kmem_free(path, ifd->ifd_len);
    481 			break;
    482 		}
    483 		rv = rumpuser_open(path,
    484 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    485 		if (rv) {
    486 			kmem_free(path, ifd->ifd_len);
    487 			break;
    488 		}
    489 		rv = initbackend(sc, memfd);
    490 		if (rv) {
    491 			kmem_free(path, ifd->ifd_len);
    492 			rumpuser_close(memfd);
    493 			break;
    494 		}
    495 		sc->sc_backfile = path;
    496 		sc->sc_backfilelen = ifd->ifd_len;
    497 
    498 		break;
    499 	default:
    500 		rv = ether_ioctl(ifp, cmd, data);
    501 		if (rv == ENETRESET)
    502 			rv = 0;
    503 		break;
    504 	}
    505 	splx(s);
    506 
    507 	return rv;
    508 }
    509 
    510 /* send everything in-context since it's just a matter of mem-to-mem copy */
    511 static void
    512 shmif_start(struct ifnet *ifp)
    513 {
    514 	struct shmif_sc *sc = ifp->if_softc;
    515 	struct shmif_mem *busmem = sc->sc_busmem;
    516 	struct mbuf *m, *m0;
    517 	uint32_t dataoff;
    518 	uint32_t pktsize, pktwrote;
    519 	bool wrote = false;
    520 	bool wrap;
    521 
    522 	ifp->if_flags |= IFF_OACTIVE;
    523 
    524 	for (;;) {
    525 		struct shmif_pkthdr sp;
    526 		struct timeval tv;
    527 
    528 		IF_DEQUEUE(&ifp->if_snd, m0);
    529 		if (m0 == NULL) {
    530 			break;
    531 		}
    532 
    533 		pktsize = 0;
    534 		for (m = m0; m != NULL; m = m->m_next) {
    535 			pktsize += m->m_len;
    536 		}
    537 		KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
    538 
    539 		getmicrouptime(&tv);
    540 		sp.sp_len = pktsize;
    541 		sp.sp_sec = tv.tv_sec;
    542 		sp.sp_usec = tv.tv_usec;
    543 
    544 		bpf_mtap(ifp, m0);
    545 
    546 		shmif_lockbus(busmem);
    547 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    548 		busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
    549 
    550 		wrap = false;
    551 		dataoff = shmif_buswrite(busmem,
    552 		    busmem->shm_last, &sp, sizeof(sp), &wrap);
    553 		pktwrote = 0;
    554 		for (m = m0; m != NULL; m = m->m_next) {
    555 			pktwrote += m->m_len;
    556 			dataoff = shmif_buswrite(busmem, dataoff,
    557 			    mtod(m, void *), m->m_len, &wrap);
    558 		}
    559 		KASSERT(pktwrote == pktsize);
    560 		if (wrap) {
    561 			busmem->shm_gen++;
    562 			DPRINTF(("bus generation now %" PRIu64 "\n",
    563 			    busmem->shm_gen));
    564 		}
    565 		shmif_unlockbus(busmem);
    566 
    567 		m_freem(m0);
    568 		wrote = true;
    569 
    570 		DPRINTF(("shmif_start: send %d bytes at off %d\n",
    571 		    pktsize, busmem->shm_last));
    572 	}
    573 
    574 	ifp->if_flags &= ~IFF_OACTIVE;
    575 
    576 	/* wakeup? */
    577 	if (wrote) {
    578 		dowakeup(sc);
    579 	}
    580 }
    581 
    582 static void
    583 shmif_stop(struct ifnet *ifp, int disable)
    584 {
    585 	struct shmif_sc *sc = ifp->if_softc;
    586 
    587 	ifp->if_flags &= ~IFF_RUNNING;
    588 	membar_producer();
    589 
    590 	/*
    591 	 * wakeup thread.  this will of course wake up all bus
    592 	 * listeners, but that's life.
    593 	 */
    594 	if (sc->sc_memfd != -1) {
    595 		dowakeup(sc);
    596 	}
    597 }
    598 
    599 
    600 /*
    601  * Check if we have been sleeping too long.  Basically,
    602  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
    603  * We use the fact that first is guaranteed to never overlap
    604  * with the last frame in the ring.
    605  */
    606 static __inline bool
    607 stillvalid_p(struct shmif_sc *sc)
    608 {
    609 	struct shmif_mem *busmem = sc->sc_busmem;
    610 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
    611 	uint32_t lastoff, devoff;
    612 
    613 	KASSERT(busmem->shm_first != busmem->shm_last);
    614 
    615 	/* normalize onto a 2x busmem chunk */
    616 	devoff = sc->sc_nextpacket;
    617 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
    618 
    619 	/* trivial case */
    620 	if (gendiff > 1)
    621 		return false;
    622 	KASSERT(gendiff <= 1);
    623 
    624 	/* Normalize onto 2x busmem chunk */
    625 	if (busmem->shm_first >= lastoff) {
    626 		lastoff += BUSMEM_DATASIZE;
    627 		if (gendiff == 0)
    628 			devoff += BUSMEM_DATASIZE;
    629 	} else {
    630 		if (gendiff)
    631 			return false;
    632 	}
    633 
    634 	return devoff >= busmem->shm_first && devoff <= lastoff;
    635 }
    636 
    637 static void
    638 shmif_rcv(void *arg)
    639 {
    640 	struct ifnet *ifp = arg;
    641 	struct shmif_sc *sc = ifp->if_softc;
    642 	struct shmif_mem *busmem;
    643 	struct mbuf *m = NULL;
    644 	struct ether_header *eth;
    645 	uint32_t nextpkt;
    646 	bool wrap, passup;
    647 	int error;
    648 	const int align
    649 	    = ALIGN(sizeof(struct ether_header)) - sizeof(struct ether_header);
    650 
    651  reup:
    652 	mutex_enter(&sc->sc_mtx);
    653 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
    654 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
    655 	mutex_exit(&sc->sc_mtx);
    656 
    657 	busmem = sc->sc_busmem;
    658 
    659 	while (ifp->if_flags & IFF_RUNNING) {
    660 		struct shmif_pkthdr sp;
    661 
    662 		if (m == NULL) {
    663 			m = m_gethdr(M_WAIT, MT_DATA);
    664 			MCLGET(m, M_WAIT);
    665 			m->m_data += align;
    666 		}
    667 
    668 		DPRINTF(("waiting %d/%" PRIu64 "\n",
    669 		    sc->sc_nextpacket, sc->sc_devgen));
    670 		KASSERT(m->m_flags & M_EXT);
    671 
    672 		shmif_lockbus(busmem);
    673 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    674 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
    675 
    676 		/* need more data? */
    677 		if (sc->sc_devgen == busmem->shm_gen &&
    678 		    shmif_nextpktoff(busmem, busmem->shm_last)
    679 		     == sc->sc_nextpacket) {
    680 			shmif_unlockbus(busmem);
    681 			error = 0;
    682 			rumpcomp_shmif_watchwait(sc->sc_kq);
    683 			if (__predict_false(error))
    684 				printf("shmif_rcv: wait failed %d\n", error);
    685 			membar_consumer();
    686 			continue;
    687 		}
    688 
    689 		if (stillvalid_p(sc)) {
    690 			nextpkt = sc->sc_nextpacket;
    691 		} else {
    692 			KASSERT(busmem->shm_gen > 0);
    693 			nextpkt = busmem->shm_first;
    694 			if (busmem->shm_first > busmem->shm_last)
    695 				sc->sc_devgen = busmem->shm_gen - 1;
    696 			else
    697 				sc->sc_devgen = busmem->shm_gen;
    698 			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
    699 			    sc, nextpkt, sc->sc_devgen));
    700 		}
    701 
    702 		/*
    703 		 * If our read pointer is ahead the bus last write, our
    704 		 * generation must be one behind.
    705 		 */
    706 		KASSERT(!(nextpkt > busmem->shm_last
    707 		    && sc->sc_devgen == busmem->shm_gen));
    708 
    709 		wrap = false;
    710 		nextpkt = shmif_busread(busmem, &sp,
    711 		    nextpkt, sizeof(sp), &wrap);
    712 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
    713 		nextpkt = shmif_busread(busmem, mtod(m, void *),
    714 		    nextpkt, sp.sp_len, &wrap);
    715 
    716 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
    717 		    sp.sp_len, nextpkt));
    718 
    719 		sc->sc_nextpacket = nextpkt;
    720 		shmif_unlockbus(sc->sc_busmem);
    721 
    722 		if (wrap) {
    723 			sc->sc_devgen++;
    724 			DPRINTF(("dev %p generation now %" PRIu64 "\n",
    725 			    sc, sc->sc_devgen));
    726 		}
    727 
    728 		/*
    729 		 * Ignore packets too short to possibly be valid.
    730 		 * This is hit at least for the first frame on a new bus.
    731 		 */
    732 		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
    733 			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
    734 			    sp.sp_len));
    735 			continue;
    736 		}
    737 
    738 		m->m_len = m->m_pkthdr.len = sp.sp_len;
    739 		m->m_pkthdr.rcvif = ifp;
    740 
    741 		/*
    742 		 * Test if we want to pass the packet upwards
    743 		 */
    744 		eth = mtod(m, struct ether_header *);
    745 		if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
    746 		    ETHER_ADDR_LEN) == 0) {
    747 			passup = true;
    748 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
    749 			passup = true;
    750 		} else if (ifp->if_flags & IFF_PROMISC) {
    751 			m->m_flags |= M_PROMISC;
    752 			passup = true;
    753 		} else {
    754 			passup = false;
    755 		}
    756 
    757 		if (passup) {
    758 			KERNEL_LOCK(1, NULL);
    759 			bpf_mtap(ifp, m);
    760 			ifp->if_input(ifp, m);
    761 			KERNEL_UNLOCK_ONE(NULL);
    762 			m = NULL;
    763 		}
    764 		/* else: reuse mbuf for a future packet */
    765 	}
    766 	m_freem(m);
    767 	m = NULL;
    768 
    769 	if (!sc->sc_dying)
    770 		goto reup;
    771 
    772 	kthread_exit(0);
    773 }
    774