Home | History | Annotate | Line # | Download | only in libshmif
if_shmem.c revision 1.46
      1 /*	$NetBSD: if_shmem.c,v 1.46 2012/10/11 10:50:45 pooka 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 1.46 2012/10/11 10:50:45 pooka 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 
     56 static int shmif_clone(struct if_clone *, int);
     57 static int shmif_unclone(struct ifnet *);
     58 
     59 struct if_clone shmif_cloner =
     60     IF_CLONE_INITIALIZER("shmif", shmif_clone, shmif_unclone);
     61 
     62 /*
     63  * Do r/w prefault for backend pages when attaching the interface.
     64  * At least logically thinking improves performance (although no
     65  * mlocking is done, so they might go away).
     66  */
     67 #define PREFAULT_RW
     68 
     69 /*
     70  * A virtual ethernet interface which uses shared memory from a
     71  * memory mapped file as the bus.
     72  */
     73 
     74 static int	shmif_init(struct ifnet *);
     75 static int	shmif_ioctl(struct ifnet *, u_long, void *);
     76 static void	shmif_start(struct ifnet *);
     77 static void	shmif_stop(struct ifnet *, int);
     78 
     79 #include "shmifvar.h"
     80 
     81 struct shmif_sc {
     82 	struct ethercom sc_ec;
     83 	struct shmif_mem *sc_busmem;
     84 	int sc_memfd;
     85 	int sc_kq;
     86 	int sc_unit;
     87 
     88 	char *sc_backfile;
     89 	size_t sc_backfilelen;
     90 
     91 	uint64_t sc_devgen;
     92 	uint32_t sc_nextpacket;
     93 
     94 	kmutex_t sc_mtx;
     95 	kcondvar_t sc_cv;
     96 
     97 	struct lwp *sc_rcvl;
     98 	bool sc_dying;
     99 };
    100 
    101 static const uint32_t busversion = SHMIF_VERSION;
    102 
    103 static void shmif_rcv(void *);
    104 
    105 #define LOCK_UNLOCKED	0
    106 #define LOCK_LOCKED	1
    107 #define LOCK_COOLDOWN	1001
    108 
    109 vmem_t *shmif_units;
    110 
    111 /*
    112  * This locking needs work and will misbehave severely if:
    113  * 1) the backing memory has to be paged in
    114  * 2) some lockholder exits while holding the lock
    115  */
    116 static void
    117 shmif_lockbus(struct shmif_mem *busmem)
    118 {
    119 	int i = 0;
    120 
    121 	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
    122 	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
    123 		if (__predict_false(++i > LOCK_COOLDOWN)) {
    124 			uint64_t sec, nsec;
    125 			int error;
    126 
    127 			sec = 0;
    128 			nsec = 1000*1000; /* 1ms */
    129 			rumpuser_nanosleep(&sec, &nsec, &error);
    130 			i = 0;
    131 		}
    132 		continue;
    133 	}
    134 	membar_enter();
    135 }
    136 
    137 static void
    138 shmif_unlockbus(struct shmif_mem *busmem)
    139 {
    140 	unsigned int old;
    141 
    142 	membar_exit();
    143 	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
    144 	KASSERT(old == LOCK_LOCKED);
    145 }
    146 
    147 static int
    148 allocif(int unit, struct shmif_sc **scp)
    149 {
    150 	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
    151 	struct shmif_sc *sc;
    152 	struct ifnet *ifp;
    153 	uint32_t randnum;
    154 	int error;
    155 
    156 	randnum = cprng_fast32();
    157 	memcpy(&enaddr[2], &randnum, sizeof(randnum));
    158 
    159 	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
    160 	sc->sc_memfd = -1;
    161 	sc->sc_unit = unit;
    162 
    163 	ifp = &sc->sc_ec.ec_if;
    164 
    165 	sprintf(ifp->if_xname, "shmif%d", unit);
    166 	ifp->if_softc = sc;
    167 	ifp->if_flags = IFF_BROADCAST | IFF_MULTICAST;
    168 	ifp->if_init = shmif_init;
    169 	ifp->if_ioctl = shmif_ioctl;
    170 	ifp->if_start = shmif_start;
    171 	ifp->if_stop = shmif_stop;
    172 	ifp->if_mtu = ETHERMTU;
    173 	ifp->if_dlt = DLT_EN10MB;
    174 
    175 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
    176 	cv_init(&sc->sc_cv, "shmifcv");
    177 
    178 	if_attach(ifp);
    179 	ether_ifattach(ifp, enaddr);
    180 
    181 	aprint_verbose("shmif%d: Ethernet address %s\n",
    182 	    unit, ether_sprintf(enaddr));
    183 
    184 	if (scp)
    185 		*scp = sc;
    186 
    187 	error = 0;
    188 	if (rump_threads) {
    189 		error = kthread_create(PRI_NONE,
    190 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
    191 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
    192 	} else {
    193 		printf("WARNING: threads not enabled, shmif NOT working\n");
    194 	}
    195 
    196 	if (error) {
    197 		shmif_unclone(ifp);
    198 	}
    199 
    200 	return error;
    201 }
    202 
    203 static int
    204 initbackend(struct shmif_sc *sc, int memfd)
    205 {
    206 	volatile uint8_t v;
    207 	volatile uint8_t *p;
    208 	int error;
    209 
    210 	sc->sc_busmem = rumpuser_filemmap(memfd, 0, BUSMEM_SIZE,
    211 	    RUMPUSER_FILEMMAP_TRUNCATE | RUMPUSER_FILEMMAP_SHARED
    212 	    | RUMPUSER_FILEMMAP_READ | RUMPUSER_FILEMMAP_WRITE, &error);
    213 	if (error)
    214 		return error;
    215 
    216 	if (sc->sc_busmem->shm_magic
    217 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
    218 		printf("bus is not magical");
    219 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    220 		return ENOEXEC;
    221 	}
    222 
    223 	/*
    224 	 * Prefault in pages to minimize runtime penalty with buslock.
    225 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
    226 	 * rump kernel PAGE_SIZE > host page size.
    227 	 */
    228 	for (p = (uint8_t *)sc->sc_busmem;
    229 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    230 	    p += 512)
    231 		v = *p;
    232 
    233 	shmif_lockbus(sc->sc_busmem);
    234 	/* we're first?  initialize bus */
    235 	if (sc->sc_busmem->shm_magic == 0) {
    236 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
    237 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
    238 	}
    239 
    240 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    241 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    242 
    243 #ifdef PREFAULT_RW
    244 	for (p = (uint8_t *)sc->sc_busmem;
    245 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    246 	    p += PAGE_SIZE) {
    247 		v = *p;
    248 		*p = v;
    249 	}
    250 #endif
    251 	shmif_unlockbus(sc->sc_busmem);
    252 
    253 	sc->sc_kq = rumpuser_writewatchfile_setup(-1, memfd, 0, &error);
    254 	if (sc->sc_kq == -1) {
    255 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    256 		return error;
    257 	}
    258 
    259 	sc->sc_memfd = memfd;
    260 
    261 	return error;
    262 }
    263 
    264 static void
    265 finibackend(struct shmif_sc *sc)
    266 {
    267 
    268 	if (sc->sc_backfile == NULL)
    269 		return;
    270 
    271 	if (sc->sc_backfile) {
    272 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
    273 		sc->sc_backfile = NULL;
    274 		sc->sc_backfilelen = 0;
    275 	}
    276 
    277 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    278 	rumpuser_close(sc->sc_memfd, NULL);
    279 	rumpuser_close(sc->sc_kq, NULL);
    280 
    281 	sc->sc_memfd = -1;
    282 }
    283 
    284 int
    285 rump_shmif_create(const char *path, int *ifnum)
    286 {
    287 	struct shmif_sc *sc;
    288 	vmem_addr_t t;
    289 	int unit, error;
    290 	int memfd = -1; /* XXXgcc */
    291 
    292 	if (path) {
    293 		memfd = rumpuser_open(path,
    294 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &error);
    295 		if (memfd == -1)
    296 			return error;
    297 	}
    298 
    299 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
    300 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
    301 
    302 	if (error != 0) {
    303 		if (path)
    304 			rumpuser_close(memfd, NULL);
    305 		return error;
    306 	}
    307 
    308 	unit = t - 1;
    309 
    310 	if ((error = allocif(unit, &sc)) != 0) {
    311 		if (path)
    312 			rumpuser_close(memfd, NULL);
    313 		return error;
    314 	}
    315 
    316 	if (!path)
    317 		goto out;
    318 
    319 	error = initbackend(sc, memfd);
    320 	if (error) {
    321 		shmif_unclone(&sc->sc_ec.ec_if);
    322 		return error;
    323 	}
    324 
    325 	sc->sc_backfilelen = strlen(path)+1;
    326 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
    327 	strcpy(sc->sc_backfile, path);
    328 
    329  out:
    330 	if (ifnum)
    331 		*ifnum = unit;
    332 
    333 	return 0;
    334 }
    335 
    336 static int
    337 shmif_clone(struct if_clone *ifc, int unit)
    338 {
    339 	int rc;
    340 	vmem_addr_t unit2;
    341 
    342 	/*
    343 	 * Ok, we know the unit number, but we must still reserve it.
    344 	 * Otherwise the wildcard-side of things might get the same one.
    345 	 * This is slightly offset-happy due to vmem.  First, we offset
    346 	 * the range of unit numbers by +1 since vmem cannot deal with
    347 	 * ranges starting from 0.  Talk about uuuh.
    348 	 */
    349 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
    350 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
    351 	KASSERT(rc == 0 && unit2-1 == unit);
    352 
    353 	return allocif(unit, NULL);
    354 }
    355 
    356 static int
    357 shmif_unclone(struct ifnet *ifp)
    358 {
    359 	struct shmif_sc *sc = ifp->if_softc;
    360 
    361 	shmif_stop(ifp, 1);
    362 	if_down(ifp);
    363 	finibackend(sc);
    364 
    365 	mutex_enter(&sc->sc_mtx);
    366 	sc->sc_dying = true;
    367 	cv_broadcast(&sc->sc_cv);
    368 	mutex_exit(&sc->sc_mtx);
    369 
    370 	if (sc->sc_rcvl)
    371 		kthread_join(sc->sc_rcvl);
    372 	sc->sc_rcvl = NULL;
    373 
    374 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
    375 
    376 	ether_ifdetach(ifp);
    377 	if_detach(ifp);
    378 
    379 	cv_destroy(&sc->sc_cv);
    380 	mutex_destroy(&sc->sc_mtx);
    381 
    382 	kmem_free(sc, sizeof(*sc));
    383 
    384 	return 0;
    385 }
    386 
    387 static int
    388 shmif_init(struct ifnet *ifp)
    389 {
    390 	struct shmif_sc *sc = ifp->if_softc;
    391 	int error = 0;
    392 
    393 	if (sc->sc_memfd == -1)
    394 		return ENXIO;
    395 	KASSERT(sc->sc_busmem);
    396 
    397 	ifp->if_flags |= IFF_RUNNING;
    398 
    399 	mutex_enter(&sc->sc_mtx);
    400 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    401 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    402 
    403 	cv_broadcast(&sc->sc_cv);
    404 	mutex_exit(&sc->sc_mtx);
    405 
    406 	return error;
    407 }
    408 
    409 static int
    410 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
    411 {
    412 	struct shmif_sc *sc = ifp->if_softc;
    413 	struct ifdrv *ifd;
    414 	char *path;
    415 	int s, rv, memfd;
    416 
    417 	s = splnet();
    418 	switch (cmd) {
    419 	case SIOCGLINKSTR:
    420 		ifd = data;
    421 
    422 		if (sc->sc_backfilelen == 0) {
    423 			rv = ENOENT;
    424 			break;
    425 		}
    426 
    427 		ifd->ifd_len = sc->sc_backfilelen;
    428 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
    429 			rv = 0;
    430 			break;
    431 		}
    432 
    433 		if (ifd->ifd_cmd != 0) {
    434 			rv = EINVAL;
    435 			break;
    436 		}
    437 
    438 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
    439 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
    440 		break;
    441 	case SIOCSLINKSTR:
    442 		if (ifp->if_flags & IFF_UP) {
    443 			rv = EBUSY;
    444 			break;
    445 		}
    446 
    447 		ifd = data;
    448 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
    449 			finibackend(sc);
    450 			rv = 0;
    451 			break;
    452 		} else if (ifd->ifd_cmd != 0) {
    453 			rv = EINVAL;
    454 			break;
    455 		} else if (sc->sc_backfile) {
    456 			rv = EBUSY;
    457 			break;
    458 		}
    459 
    460 		if (ifd->ifd_len > MAXPATHLEN) {
    461 			rv = E2BIG;
    462 			break;
    463 		} else if (ifd->ifd_len < 1) {
    464 			rv = EINVAL;
    465 			break;
    466 		}
    467 
    468 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
    469 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
    470 		if (rv) {
    471 			kmem_free(path, ifd->ifd_len);
    472 			break;
    473 		}
    474 		memfd = rumpuser_open(path,
    475 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &rv);
    476 		if (memfd == -1) {
    477 			kmem_free(path, ifd->ifd_len);
    478 			break;
    479 		}
    480 		rv = initbackend(sc, memfd);
    481 		if (rv) {
    482 			kmem_free(path, ifd->ifd_len);
    483 			rumpuser_close(memfd, NULL);
    484 			break;
    485 		}
    486 		sc->sc_backfile = path;
    487 		sc->sc_backfilelen = ifd->ifd_len;
    488 
    489 		break;
    490 	default:
    491 		rv = ether_ioctl(ifp, cmd, data);
    492 		if (rv == ENETRESET)
    493 			rv = 0;
    494 		break;
    495 	}
    496 	splx(s);
    497 
    498 	return rv;
    499 }
    500 
    501 /* send everything in-context since it's just a matter of mem-to-mem copy */
    502 static void
    503 shmif_start(struct ifnet *ifp)
    504 {
    505 	struct shmif_sc *sc = ifp->if_softc;
    506 	struct shmif_mem *busmem = sc->sc_busmem;
    507 	struct mbuf *m, *m0;
    508 	uint32_t dataoff;
    509 	uint32_t pktsize, pktwrote;
    510 	bool wrote = false;
    511 	bool wrap;
    512 	int error;
    513 
    514 	ifp->if_flags |= IFF_OACTIVE;
    515 
    516 	for (;;) {
    517 		struct shmif_pkthdr sp;
    518 		struct timeval tv;
    519 
    520 		IF_DEQUEUE(&ifp->if_snd, m0);
    521 		if (m0 == NULL) {
    522 			break;
    523 		}
    524 
    525 		pktsize = 0;
    526 		for (m = m0; m != NULL; m = m->m_next) {
    527 			pktsize += m->m_len;
    528 		}
    529 		KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
    530 
    531 		getmicrouptime(&tv);
    532 		sp.sp_len = pktsize;
    533 		sp.sp_sec = tv.tv_sec;
    534 		sp.sp_usec = tv.tv_usec;
    535 
    536 		bpf_mtap(ifp, m0);
    537 
    538 		shmif_lockbus(busmem);
    539 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    540 		busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
    541 
    542 		wrap = false;
    543 		dataoff = shmif_buswrite(busmem,
    544 		    busmem->shm_last, &sp, sizeof(sp), &wrap);
    545 		pktwrote = 0;
    546 		for (m = m0; m != NULL; m = m->m_next) {
    547 			pktwrote += m->m_len;
    548 			dataoff = shmif_buswrite(busmem, dataoff,
    549 			    mtod(m, void *), m->m_len, &wrap);
    550 		}
    551 		KASSERT(pktwrote == pktsize);
    552 		if (wrap) {
    553 			busmem->shm_gen++;
    554 			DPRINTF(("bus generation now %d\n", busmem->shm_gen));
    555 		}
    556 		shmif_unlockbus(busmem);
    557 
    558 		m_freem(m0);
    559 		wrote = true;
    560 
    561 		DPRINTF(("shmif_start: send %d bytes at off %d\n",
    562 		    pktsize, busmem->shm_last));
    563 	}
    564 
    565 	ifp->if_flags &= ~IFF_OACTIVE;
    566 
    567 	/* wakeup? */
    568 	if (wrote)
    569 		rumpuser_pwrite(sc->sc_memfd,
    570 		    &busversion, sizeof(busversion), IFMEM_WAKEUP, &error);
    571 }
    572 
    573 static void
    574 shmif_stop(struct ifnet *ifp, int disable)
    575 {
    576 	struct shmif_sc *sc = ifp->if_softc;
    577 
    578 	ifp->if_flags &= ~IFF_RUNNING;
    579 	membar_producer();
    580 
    581 	/*
    582 	 * wakeup thread.  this will of course wake up all bus
    583 	 * listeners, but that's life.
    584 	 */
    585 	if (sc->sc_memfd != -1)
    586 		rumpuser_pwrite(sc->sc_memfd,
    587 		    &busversion, sizeof(busversion), IFMEM_WAKEUP, NULL);
    588 }
    589 
    590 
    591 /*
    592  * Check if we have been sleeping too long.  Basically,
    593  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
    594  * We use the fact that first is guaranteed to never overlap
    595  * with the last frame in the ring.
    596  */
    597 static __inline bool
    598 stillvalid_p(struct shmif_sc *sc)
    599 {
    600 	struct shmif_mem *busmem = sc->sc_busmem;
    601 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
    602 	uint32_t lastoff, devoff;
    603 
    604 	KASSERT(busmem->shm_first != busmem->shm_last);
    605 
    606 	/* normalize onto a 2x busmem chunk */
    607 	devoff = sc->sc_nextpacket;
    608 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
    609 
    610 	/* trivial case */
    611 	if (gendiff > 1)
    612 		return false;
    613 	KASSERT(gendiff <= 1);
    614 
    615 	/* Normalize onto 2x busmem chunk */
    616 	if (busmem->shm_first >= lastoff) {
    617 		lastoff += BUSMEM_DATASIZE;
    618 		if (gendiff == 0)
    619 			devoff += BUSMEM_DATASIZE;
    620 	} else {
    621 		if (gendiff)
    622 			return false;
    623 	}
    624 
    625 	return devoff >= busmem->shm_first && devoff <= lastoff;
    626 }
    627 
    628 static void
    629 shmif_rcv(void *arg)
    630 {
    631 	struct ifnet *ifp = arg;
    632 	struct shmif_sc *sc = ifp->if_softc;
    633 	struct shmif_mem *busmem;
    634 	struct mbuf *m = NULL;
    635 	struct ether_header *eth;
    636 	uint32_t nextpkt;
    637 	bool wrap, passup;
    638 	int error;
    639 
    640  reup:
    641 	mutex_enter(&sc->sc_mtx);
    642 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
    643 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
    644 	mutex_exit(&sc->sc_mtx);
    645 
    646 	busmem = sc->sc_busmem;
    647 
    648 	while (ifp->if_flags & IFF_RUNNING) {
    649 		struct shmif_pkthdr sp;
    650 
    651 		if (m == NULL) {
    652 			m = m_gethdr(M_WAIT, MT_DATA);
    653 			MCLGET(m, M_WAIT);
    654 		}
    655 
    656 		DPRINTF(("waiting %d/%d\n", sc->sc_nextpacket, sc->sc_devgen));
    657 		KASSERT(m->m_flags & M_EXT);
    658 
    659 		shmif_lockbus(busmem);
    660 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    661 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
    662 
    663 		/* need more data? */
    664 		if (sc->sc_devgen == busmem->shm_gen &&
    665 		    shmif_nextpktoff(busmem, busmem->shm_last)
    666 		     == sc->sc_nextpacket) {
    667 			shmif_unlockbus(busmem);
    668 			error = 0;
    669 			rumpuser_writewatchfile_wait(sc->sc_kq, NULL, &error);
    670 			if (__predict_false(error))
    671 				printf("shmif_rcv: wait failed %d\n", error);
    672 			membar_consumer();
    673 			continue;
    674 		}
    675 
    676 		if (stillvalid_p(sc)) {
    677 			nextpkt = sc->sc_nextpacket;
    678 		} else {
    679 			KASSERT(busmem->shm_gen > 0);
    680 			nextpkt = busmem->shm_first;
    681 			if (busmem->shm_first > busmem->shm_last)
    682 				sc->sc_devgen = busmem->shm_gen - 1;
    683 			else
    684 				sc->sc_devgen = busmem->shm_gen;
    685 			DPRINTF(("dev %p overrun, new data: %d/%d\n",
    686 			    sc, nextpkt, sc->sc_devgen));
    687 		}
    688 
    689 		/*
    690 		 * If our read pointer is ahead the bus last write, our
    691 		 * generation must be one behind.
    692 		 */
    693 		KASSERT(!(nextpkt > busmem->shm_last
    694 		    && sc->sc_devgen == busmem->shm_gen));
    695 
    696 		wrap = false;
    697 		nextpkt = shmif_busread(busmem, &sp,
    698 		    nextpkt, sizeof(sp), &wrap);
    699 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
    700 		nextpkt = shmif_busread(busmem, mtod(m, void *),
    701 		    nextpkt, sp.sp_len, &wrap);
    702 
    703 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
    704 		    sp.sp_len, nextpkt));
    705 
    706 		sc->sc_nextpacket = nextpkt;
    707 		shmif_unlockbus(sc->sc_busmem);
    708 
    709 		if (wrap) {
    710 			sc->sc_devgen++;
    711 			DPRINTF(("dev %p generation now %d\n",
    712 			    sc, sc->sc_devgen));
    713 		}
    714 
    715 		m->m_len = m->m_pkthdr.len = sp.sp_len;
    716 		m->m_pkthdr.rcvif = ifp;
    717 
    718 		/*
    719 		 * Test if we want to pass the packet upwards
    720 		 */
    721 		eth = mtod(m, struct ether_header *);
    722 		if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
    723 		    ETHER_ADDR_LEN) == 0) {
    724 			passup = true;
    725 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
    726 			passup = true;
    727 		} else if (ifp->if_flags & IFF_PROMISC) {
    728 			m->m_flags |= M_PROMISC;
    729 			passup = true;
    730 		} else {
    731 			passup = false;
    732 		}
    733 
    734 		if (passup) {
    735 			KERNEL_LOCK(1, NULL);
    736 			bpf_mtap(ifp, m);
    737 			ifp->if_input(ifp, m);
    738 			KERNEL_UNLOCK_ONE(NULL);
    739 			m = NULL;
    740 		}
    741 		/* else: reuse mbuf for a future packet */
    742 	}
    743 	m_freem(m);
    744 	m = NULL;
    745 
    746 	if (!sc->sc_dying)
    747 		goto reup;
    748 
    749 	kthread_exit(0);
    750 }
    751