Home | History | Annotate | Line # | Download | only in libshmif
if_shmem.c revision 1.55
      1 /*	$NetBSD: if_shmem.c,v 1.55 2013/06/14 05:56:29 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 1.55 2013/06/14 05:56:29 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 #include "rumpcomp_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 const uint32_t busversion = SHMIF_VERSION;
    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;
    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 
    173 	ifp = &sc->sc_ec.ec_if;
    174 
    175 	sprintf(ifp->if_xname, "shmif%d", unit);
    176 	ifp->if_softc = sc;
    177 	ifp->if_flags = IFF_BROADCAST | IFF_MULTICAST;
    178 	ifp->if_init = shmif_init;
    179 	ifp->if_ioctl = shmif_ioctl;
    180 	ifp->if_start = shmif_start;
    181 	ifp->if_stop = shmif_stop;
    182 	ifp->if_mtu = ETHERMTU;
    183 	ifp->if_dlt = DLT_EN10MB;
    184 
    185 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
    186 	cv_init(&sc->sc_cv, "shmifcv");
    187 
    188 	if_attach(ifp);
    189 	ether_ifattach(ifp, enaddr);
    190 
    191 	aprint_verbose("shmif%d: Ethernet address %s\n",
    192 	    unit, ether_sprintf(enaddr));
    193 
    194 	if (scp)
    195 		*scp = sc;
    196 
    197 	error = 0;
    198 	if (rump_threads) {
    199 		error = kthread_create(PRI_NONE,
    200 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
    201 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
    202 	} else {
    203 		printf("WARNING: threads not enabled, shmif NOT working\n");
    204 	}
    205 
    206 	if (error) {
    207 		shmif_unclone(ifp);
    208 	}
    209 
    210 	return error;
    211 }
    212 
    213 static int
    214 initbackend(struct shmif_sc *sc, int memfd)
    215 {
    216 	volatile uint8_t v;
    217 	volatile uint8_t *p;
    218 	void *mem;
    219 	int error;
    220 
    221 	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
    222 	if (error)
    223 		return error;
    224 	sc->sc_busmem = mem;
    225 
    226 	if (sc->sc_busmem->shm_magic
    227 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
    228 		printf("bus is not magical");
    229 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    230 		return ENOEXEC;
    231 	}
    232 
    233 	/*
    234 	 * Prefault in pages to minimize runtime penalty with buslock.
    235 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
    236 	 * rump kernel PAGE_SIZE > host page size.
    237 	 */
    238 	for (p = (uint8_t *)sc->sc_busmem;
    239 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    240 	    p += 512)
    241 		v = *p;
    242 
    243 	shmif_lockbus(sc->sc_busmem);
    244 	/* we're first?  initialize bus */
    245 	if (sc->sc_busmem->shm_magic == 0) {
    246 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
    247 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
    248 	}
    249 
    250 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    251 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    252 
    253 #ifdef PREFAULT_RW
    254 	for (p = (uint8_t *)sc->sc_busmem;
    255 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    256 	    p += PAGE_SIZE) {
    257 		v = *p;
    258 		*p = v;
    259 	}
    260 #endif
    261 	shmif_unlockbus(sc->sc_busmem);
    262 
    263 	sc->sc_kq = -1;
    264 	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
    265 	if (error) {
    266 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    267 		return error;
    268 	}
    269 
    270 	sc->sc_memfd = memfd;
    271 
    272 	return error;
    273 }
    274 
    275 static void
    276 finibackend(struct shmif_sc *sc)
    277 {
    278 
    279 	if (sc->sc_backfile == NULL)
    280 		return;
    281 
    282 	if (sc->sc_backfile) {
    283 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
    284 		sc->sc_backfile = NULL;
    285 		sc->sc_backfilelen = 0;
    286 	}
    287 
    288 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    289 	rumpuser_close(sc->sc_memfd);
    290 	rumpuser_close(sc->sc_kq);
    291 
    292 	sc->sc_memfd = -1;
    293 }
    294 
    295 int
    296 rump_shmif_create(const char *path, int *ifnum)
    297 {
    298 	struct shmif_sc *sc;
    299 	vmem_addr_t t;
    300 	int unit, error;
    301 	int memfd = -1; /* XXXgcc */
    302 
    303 	if (path) {
    304 		error = rumpuser_open(path,
    305 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    306 		if (error)
    307 			return error;
    308 	}
    309 
    310 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
    311 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
    312 
    313 	if (error != 0) {
    314 		if (path)
    315 			rumpuser_close(memfd);
    316 		return error;
    317 	}
    318 
    319 	unit = t - 1;
    320 
    321 	if ((error = allocif(unit, &sc)) != 0) {
    322 		if (path)
    323 			rumpuser_close(memfd);
    324 		return error;
    325 	}
    326 
    327 	if (!path)
    328 		goto out;
    329 
    330 	error = initbackend(sc, memfd);
    331 	if (error) {
    332 		shmif_unclone(&sc->sc_ec.ec_if);
    333 		return error;
    334 	}
    335 
    336 	sc->sc_backfilelen = strlen(path)+1;
    337 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
    338 	strcpy(sc->sc_backfile, path);
    339 
    340  out:
    341 	if (ifnum)
    342 		*ifnum = unit;
    343 
    344 	return 0;
    345 }
    346 
    347 static int
    348 shmif_clone(struct if_clone *ifc, int unit)
    349 {
    350 	int rc;
    351 	vmem_addr_t unit2;
    352 
    353 	/*
    354 	 * Ok, we know the unit number, but we must still reserve it.
    355 	 * Otherwise the wildcard-side of things might get the same one.
    356 	 * This is slightly offset-happy due to vmem.  First, we offset
    357 	 * the range of unit numbers by +1 since vmem cannot deal with
    358 	 * ranges starting from 0.  Talk about uuuh.
    359 	 */
    360 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
    361 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
    362 	KASSERT(rc == 0 && unit2-1 == unit);
    363 
    364 	return allocif(unit, NULL);
    365 }
    366 
    367 static int
    368 shmif_unclone(struct ifnet *ifp)
    369 {
    370 	struct shmif_sc *sc = ifp->if_softc;
    371 
    372 	shmif_stop(ifp, 1);
    373 	if_down(ifp);
    374 	finibackend(sc);
    375 
    376 	mutex_enter(&sc->sc_mtx);
    377 	sc->sc_dying = true;
    378 	cv_broadcast(&sc->sc_cv);
    379 	mutex_exit(&sc->sc_mtx);
    380 
    381 	if (sc->sc_rcvl)
    382 		kthread_join(sc->sc_rcvl);
    383 	sc->sc_rcvl = NULL;
    384 
    385 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
    386 
    387 	ether_ifdetach(ifp);
    388 	if_detach(ifp);
    389 
    390 	cv_destroy(&sc->sc_cv);
    391 	mutex_destroy(&sc->sc_mtx);
    392 
    393 	kmem_free(sc, sizeof(*sc));
    394 
    395 	return 0;
    396 }
    397 
    398 static int
    399 shmif_init(struct ifnet *ifp)
    400 {
    401 	struct shmif_sc *sc = ifp->if_softc;
    402 	int error = 0;
    403 
    404 	if (sc->sc_memfd == -1)
    405 		return ENXIO;
    406 	KASSERT(sc->sc_busmem);
    407 
    408 	ifp->if_flags |= IFF_RUNNING;
    409 
    410 	mutex_enter(&sc->sc_mtx);
    411 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    412 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    413 
    414 	cv_broadcast(&sc->sc_cv);
    415 	mutex_exit(&sc->sc_mtx);
    416 
    417 	return error;
    418 }
    419 
    420 static int
    421 shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
    422 {
    423 	struct shmif_sc *sc = ifp->if_softc;
    424 	struct ifdrv *ifd;
    425 	char *path;
    426 	int s, rv, memfd;
    427 
    428 	s = splnet();
    429 	switch (cmd) {
    430 	case SIOCGLINKSTR:
    431 		ifd = data;
    432 
    433 		if (sc->sc_backfilelen == 0) {
    434 			rv = ENOENT;
    435 			break;
    436 		}
    437 
    438 		ifd->ifd_len = sc->sc_backfilelen;
    439 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
    440 			rv = 0;
    441 			break;
    442 		}
    443 
    444 		if (ifd->ifd_cmd != 0) {
    445 			rv = EINVAL;
    446 			break;
    447 		}
    448 
    449 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
    450 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
    451 		break;
    452 	case SIOCSLINKSTR:
    453 		if (ifp->if_flags & IFF_UP) {
    454 			rv = EBUSY;
    455 			break;
    456 		}
    457 
    458 		ifd = data;
    459 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
    460 			finibackend(sc);
    461 			rv = 0;
    462 			break;
    463 		} else if (ifd->ifd_cmd != 0) {
    464 			rv = EINVAL;
    465 			break;
    466 		} else if (sc->sc_backfile) {
    467 			rv = EBUSY;
    468 			break;
    469 		}
    470 
    471 		if (ifd->ifd_len > MAXPATHLEN) {
    472 			rv = E2BIG;
    473 			break;
    474 		} else if (ifd->ifd_len < 1) {
    475 			rv = EINVAL;
    476 			break;
    477 		}
    478 
    479 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
    480 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
    481 		if (rv) {
    482 			kmem_free(path, ifd->ifd_len);
    483 			break;
    484 		}
    485 		rv = rumpuser_open(path,
    486 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    487 		if (rv) {
    488 			kmem_free(path, ifd->ifd_len);
    489 			break;
    490 		}
    491 		rv = initbackend(sc, memfd);
    492 		if (rv) {
    493 			kmem_free(path, ifd->ifd_len);
    494 			rumpuser_close(memfd);
    495 			break;
    496 		}
    497 		sc->sc_backfile = path;
    498 		sc->sc_backfilelen = ifd->ifd_len;
    499 
    500 		break;
    501 	default:
    502 		rv = ether_ioctl(ifp, cmd, data);
    503 		if (rv == ENETRESET)
    504 			rv = 0;
    505 		break;
    506 	}
    507 	splx(s);
    508 
    509 	return rv;
    510 }
    511 
    512 /* send everything in-context since it's just a matter of mem-to-mem copy */
    513 static void
    514 shmif_start(struct ifnet *ifp)
    515 {
    516 	struct shmif_sc *sc = ifp->if_softc;
    517 	struct shmif_mem *busmem = sc->sc_busmem;
    518 	struct mbuf *m, *m0;
    519 	uint32_t dataoff;
    520 	uint32_t pktsize, pktwrote;
    521 	bool wrote = false;
    522 	bool wrap;
    523 
    524 	ifp->if_flags |= IFF_OACTIVE;
    525 
    526 	for (;;) {
    527 		struct shmif_pkthdr sp;
    528 		struct timeval tv;
    529 
    530 		IF_DEQUEUE(&ifp->if_snd, m0);
    531 		if (m0 == NULL) {
    532 			break;
    533 		}
    534 
    535 		pktsize = 0;
    536 		for (m = m0; m != NULL; m = m->m_next) {
    537 			pktsize += m->m_len;
    538 		}
    539 		KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
    540 
    541 		getmicrouptime(&tv);
    542 		sp.sp_len = pktsize;
    543 		sp.sp_sec = tv.tv_sec;
    544 		sp.sp_usec = tv.tv_usec;
    545 
    546 		bpf_mtap(ifp, m0);
    547 
    548 		shmif_lockbus(busmem);
    549 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    550 		busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
    551 
    552 		wrap = false;
    553 		dataoff = shmif_buswrite(busmem,
    554 		    busmem->shm_last, &sp, sizeof(sp), &wrap);
    555 		pktwrote = 0;
    556 		for (m = m0; m != NULL; m = m->m_next) {
    557 			pktwrote += m->m_len;
    558 			dataoff = shmif_buswrite(busmem, dataoff,
    559 			    mtod(m, void *), m->m_len, &wrap);
    560 		}
    561 		KASSERT(pktwrote == pktsize);
    562 		if (wrap) {
    563 			busmem->shm_gen++;
    564 			DPRINTF(("bus generation now %" PRIu64 "\n",
    565 			    busmem->shm_gen));
    566 		}
    567 		shmif_unlockbus(busmem);
    568 
    569 		m_freem(m0);
    570 		wrote = true;
    571 
    572 		DPRINTF(("shmif_start: send %d bytes at off %d\n",
    573 		    pktsize, busmem->shm_last));
    574 	}
    575 
    576 	ifp->if_flags &= ~IFF_OACTIVE;
    577 
    578 	/* wakeup? */
    579 	if (wrote) {
    580 		dowakeup(sc);
    581 	}
    582 }
    583 
    584 static void
    585 shmif_stop(struct ifnet *ifp, int disable)
    586 {
    587 	struct shmif_sc *sc = ifp->if_softc;
    588 
    589 	ifp->if_flags &= ~IFF_RUNNING;
    590 	membar_producer();
    591 
    592 	/*
    593 	 * wakeup thread.  this will of course wake up all bus
    594 	 * listeners, but that's life.
    595 	 */
    596 	if (sc->sc_memfd != -1) {
    597 		dowakeup(sc);
    598 	}
    599 }
    600 
    601 
    602 /*
    603  * Check if we have been sleeping too long.  Basically,
    604  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
    605  * We use the fact that first is guaranteed to never overlap
    606  * with the last frame in the ring.
    607  */
    608 static __inline bool
    609 stillvalid_p(struct shmif_sc *sc)
    610 {
    611 	struct shmif_mem *busmem = sc->sc_busmem;
    612 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
    613 	uint32_t lastoff, devoff;
    614 
    615 	KASSERT(busmem->shm_first != busmem->shm_last);
    616 
    617 	/* normalize onto a 2x busmem chunk */
    618 	devoff = sc->sc_nextpacket;
    619 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
    620 
    621 	/* trivial case */
    622 	if (gendiff > 1)
    623 		return false;
    624 	KASSERT(gendiff <= 1);
    625 
    626 	/* Normalize onto 2x busmem chunk */
    627 	if (busmem->shm_first >= lastoff) {
    628 		lastoff += BUSMEM_DATASIZE;
    629 		if (gendiff == 0)
    630 			devoff += BUSMEM_DATASIZE;
    631 	} else {
    632 		if (gendiff)
    633 			return false;
    634 	}
    635 
    636 	return devoff >= busmem->shm_first && devoff <= lastoff;
    637 }
    638 
    639 static void
    640 shmif_rcv(void *arg)
    641 {
    642 	struct ifnet *ifp = arg;
    643 	struct shmif_sc *sc = ifp->if_softc;
    644 	struct shmif_mem *busmem;
    645 	struct mbuf *m = NULL;
    646 	struct ether_header *eth;
    647 	uint32_t nextpkt;
    648 	bool wrap, passup;
    649 	int error;
    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 		}
    666 
    667 		DPRINTF(("waiting %d/%" PRIu64 "\n",
    668 		    sc->sc_nextpacket, sc->sc_devgen));
    669 		KASSERT(m->m_flags & M_EXT);
    670 
    671 		shmif_lockbus(busmem);
    672 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    673 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
    674 
    675 		/* need more data? */
    676 		if (sc->sc_devgen == busmem->shm_gen &&
    677 		    shmif_nextpktoff(busmem, busmem->shm_last)
    678 		     == sc->sc_nextpacket) {
    679 			shmif_unlockbus(busmem);
    680 			error = 0;
    681 			rumpcomp_shmif_watchwait(sc->sc_kq);
    682 			if (__predict_false(error))
    683 				printf("shmif_rcv: wait failed %d\n", error);
    684 			membar_consumer();
    685 			continue;
    686 		}
    687 
    688 		if (stillvalid_p(sc)) {
    689 			nextpkt = sc->sc_nextpacket;
    690 		} else {
    691 			KASSERT(busmem->shm_gen > 0);
    692 			nextpkt = busmem->shm_first;
    693 			if (busmem->shm_first > busmem->shm_last)
    694 				sc->sc_devgen = busmem->shm_gen - 1;
    695 			else
    696 				sc->sc_devgen = busmem->shm_gen;
    697 			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
    698 			    sc, nextpkt, sc->sc_devgen));
    699 		}
    700 
    701 		/*
    702 		 * If our read pointer is ahead the bus last write, our
    703 		 * generation must be one behind.
    704 		 */
    705 		KASSERT(!(nextpkt > busmem->shm_last
    706 		    && sc->sc_devgen == busmem->shm_gen));
    707 
    708 		wrap = false;
    709 		nextpkt = shmif_busread(busmem, &sp,
    710 		    nextpkt, sizeof(sp), &wrap);
    711 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
    712 		nextpkt = shmif_busread(busmem, mtod(m, void *),
    713 		    nextpkt, sp.sp_len, &wrap);
    714 
    715 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
    716 		    sp.sp_len, nextpkt));
    717 
    718 		sc->sc_nextpacket = nextpkt;
    719 		shmif_unlockbus(sc->sc_busmem);
    720 
    721 		if (wrap) {
    722 			sc->sc_devgen++;
    723 			DPRINTF(("dev %p generation now %" PRIu64 "\n",
    724 			    sc, sc->sc_devgen));
    725 		}
    726 
    727 		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
    728 			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
    729 			    sp.sp_len));
    730 			continue;
    731 		}
    732 
    733 		m->m_len = m->m_pkthdr.len = sp.sp_len;
    734 		m->m_pkthdr.rcvif = ifp;
    735 
    736 		/*
    737 		 * Test if we want to pass the packet upwards
    738 		 */
    739 		eth = mtod(m, struct ether_header *);
    740 		if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
    741 		    ETHER_ADDR_LEN) == 0) {
    742 			passup = true;
    743 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
    744 			passup = true;
    745 		} else if (ifp->if_flags & IFF_PROMISC) {
    746 			m->m_flags |= M_PROMISC;
    747 			passup = true;
    748 		} else {
    749 			passup = false;
    750 		}
    751 
    752 		if (passup) {
    753 			KERNEL_LOCK(1, NULL);
    754 			bpf_mtap(ifp, m);
    755 			ifp->if_input(ifp, m);
    756 			KERNEL_UNLOCK_ONE(NULL);
    757 			m = NULL;
    758 		}
    759 		/* else: reuse mbuf for a future packet */
    760 	}
    761 	m_freem(m);
    762 	m = NULL;
    763 
    764 	if (!sc->sc_dying)
    765 		goto reup;
    766 
    767 	kthread_exit(0);
    768 }
    769