xref: /src/sys/rump/net/lib/libshmif/if_shmem.c

/*	$NetBSD: if_shmem.c,v 1.84.4.3 2024/10/03 16:19:59 martin Exp $	*/

/*
 * Copyright (c) 2009, 2010 Antti Kantee.  All Rights Reserved.
 *
 * Development of this software was supported by The Nokia Foundation.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
 * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 */

#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: if_shmem.c,v 1.84.4.3 2024/10/03 16:19:59 martin Exp $");

#include <sys/param.h>
#include <sys/atomic.h>
#include <sys/fcntl.h>
#include <sys/kmem.h>
#include <sys/kthread.h>
#include <sys/lock.h>
#include <sys/vmem.h>
#include <sys/cprng.h>

#include <net/bpf.h>
#include <net/if.h>
#include <net/if_dl.h>
#include <net/if_ether.h>
#include <net/if_media.h>
#include <net/ether_sw_offload.h>

#include <netinet/in.h>
#include <netinet/in_var.h>

#include <rump-sys/kern.h>
#include <rump-sys/net.h>

#include <rump/rump.h>
#include <rump/rumpuser.h>

#include "shmif_user.h"

static int shmif_clone(struct if_clone *, int);
static int shmif_unclone(struct ifnet *);

static int shmif_mediachange(struct ifnet *);
static void shmif_mediastatus(struct ifnet *, struct ifmediareq *);

struct if_clone shmif_cloner =
    IF_CLONE_INITIALIZER("shmif", shmif_clone, shmif_unclone);

/*
 * Do r/w prefault for backend pages when attaching the interface.
 * At least logically thinking improves performance (although no
 * mlocking is done, so they might go away).
 */
#define PREFAULT_RW

/*
 * A virtual ethernet interface which uses shared memory from a
 * memory mapped file as the bus.
 */

static int	shmif_init(struct ifnet *);
static int	shmif_ioctl(struct ifnet *, u_long, void *);
static void	shmif_start(struct ifnet *);
static void	shmif_snd(struct ifnet *, struct mbuf *);
static void	shmif_stop(struct ifnet *, int);

#include "shmifvar.h"

struct shmif_sc {
	struct ethercom sc_ec;
	struct ifmedia sc_im;
	struct shmif_mem *sc_busmem;
	int sc_memfd;
	int sc_kq;
	int sc_unit;

	char *sc_backfile;
	size_t sc_backfilelen;

	uint64_t sc_devgen;
	uint32_t sc_nextpacket;

	kmutex_t sc_mtx;
	kcondvar_t sc_cv;

	struct lwp *sc_rcvl;
	bool sc_dying;

	uint64_t sc_uid;
};

static void shmif_rcv(void *);

vmem_t *shmif_units;

static void
dowakeup(struct shmif_sc *sc)
{
	struct rumpuser_iovec iov;
	uint32_t ver = SHMIF_VERSION;
	size_t n;

	iov.iov_base = &ver;
	iov.iov_len = sizeof(ver);
	rumpuser_iovwrite(sc->sc_memfd, &iov, 1, IFMEM_WAKEUP, &n);
}

/*
 * This locking needs work and will misbehave severely if:
 * 1) the backing memory has to be paged in
 * 2) some lockholder exits while holding the lock
 */
static void
shmif_lockbus(struct shmif_mem *busmem)
{
	int i = 0;

	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
		if (__predict_false(++i > LOCK_COOLDOWN)) {
			/* wait 1ms */
			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL,
			    0, 1000*1000);
			i = 0;
		}
		continue;
	}
	membar_acquire();
}

static void
shmif_unlockbus(struct shmif_mem *busmem)
{
	unsigned int old __diagused;

	membar_release();
	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
	KASSERT(old == LOCK_LOCKED);
}

static int
allocif(int unit, struct shmif_sc **scp)
{
	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
	struct shmif_sc *sc;
	struct ifnet *ifp;
	uint64_t randnum;
	int error = 0;

	randnum = cprng_strong64();
	memcpy(&enaddr[2], &randnum, 4);

	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
	sc->sc_memfd = -1;
	sc->sc_unit = unit;
	sc->sc_uid = randnum;

	ifp = &sc->sc_ec.ec_if;

	ifmedia_init(&sc->sc_im, 0, shmif_mediachange, shmif_mediastatus);
	ifmedia_add(&sc->sc_im, IFM_ETHER|IFM_AUTO, 0, NULL);
	ifmedia_add(&sc->sc_im, IFM_ETHER|IFM_NONE, 0, NULL);
	ifmedia_set(&sc->sc_im, IFM_ETHER|IFM_AUTO);

	snprintf(ifp->if_xname, sizeof(ifp->if_xname), "shmif%d", unit);
	ifp->if_softc = sc;
	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
	ifp->if_init = shmif_init;
	ifp->if_ioctl = shmif_ioctl;
	ifp->if_start = shmif_start;
	ifp->if_stop = shmif_stop;
	ifp->if_mtu = ETHERMTU;
	ifp->if_dlt = DLT_EN10MB;
	ifp->if_capabilities = IFCAP_TSOv4 | IFCAP_TSOv6 |
	    IFCAP_CSUM_IPv4_Rx	| IFCAP_CSUM_IPv4_Tx |
	    IFCAP_CSUM_TCPv4_Rx	| IFCAP_CSUM_TCPv4_Tx |
	    IFCAP_CSUM_UDPv4_Rx	| IFCAP_CSUM_UDPv4_Tx |
	    IFCAP_CSUM_TCPv6_Rx	| IFCAP_CSUM_TCPv6_Tx |
	    IFCAP_CSUM_UDPv6_Rx	| IFCAP_CSUM_UDPv6_Tx;
	IFQ_SET_READY(&ifp->if_snd);

	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
	cv_init(&sc->sc_cv, "shmifcv");

	if_initialize(ifp);
#if 1
	char buf[256];

	if (rumpuser_getparam("RUMP_SHMIF_CAPENABLE", buf, sizeof(buf)) == 0) {
		uint64_t capen = strtoul(buf, NULL, 0);

		ifp->if_capenable = capen & ifp->if_capabilities;
	}
#endif

	if_deferred_start_init(ifp, NULL);
	ether_ifattach(ifp, enaddr);
	if_register(ifp);

	aprint_verbose("shmif%d: Ethernet address %s\n",
	    unit, ether_sprintf(enaddr));

	if (scp)
		*scp = sc;

	if (rump_threads) {
		error = kthread_create(PRI_NONE,
		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
	} else {
		printf("WARNING: threads not enabled, shmif NOT working\n");
	}

	if (error) {
		shmif_unclone(ifp);
	}

	return 0;
}

static int
initbackend(struct shmif_sc *sc, int memfd)
{
	volatile uint8_t v;
	volatile uint8_t *p;
	void *mem;
	int error;

	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
	if (error)
		return error;
	sc->sc_busmem = mem;

	if (sc->sc_busmem->shm_magic
	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
		printf("bus is not magical");
		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
		return ENOEXEC;
	}

	/*
	 * Prefault in pages to minimize runtime penalty with buslock.
	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
	 * rump kernel PAGE_SIZE > host page size.
	 */
	for (p = (uint8_t *)sc->sc_busmem;
	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
	    p += 512)
		v = *p;

	shmif_lockbus(sc->sc_busmem);
	/* we're first?  initialize bus */
	if (sc->sc_busmem->shm_magic == 0) {
		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
	}

	sc->sc_nextpacket = sc->sc_busmem->shm_last;
	sc->sc_devgen = sc->sc_busmem->shm_gen;

#ifdef PREFAULT_RW
	for (p = (uint8_t *)sc->sc_busmem;
	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
	    p += PAGE_SIZE) {
		v = *p;
		*p = v;
	}
#endif
	shmif_unlockbus(sc->sc_busmem);

	sc->sc_kq = -1;
	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
	if (error) {
		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
		return error;
	}

	sc->sc_memfd = memfd;

	return error;
}

static void
finibackend(struct shmif_sc *sc)
{

	if (sc->sc_backfile == NULL)
		return;

	if (sc->sc_backfile) {
		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
		sc->sc_backfile = NULL;
		sc->sc_backfilelen = 0;
	}

	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
	rumpuser_close(sc->sc_memfd);
	rumpuser_close(sc->sc_kq);

	sc->sc_memfd = -1;
}

int
rump_shmif_create(const char *path, int *ifnum)
{
	struct shmif_sc *sc;
	vmem_addr_t t;
	int unit, error;
	int memfd = -1; /* XXXgcc */

	if (path) {
		error = rumpuser_open(path,
		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
		if (error)
			return error;
	}

	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);

	if (error != 0) {
		if (path)
			rumpuser_close(memfd);
		return error;
	}

	unit = t - 1;

	if ((error = allocif(unit, &sc)) != 0) {
		if (path)
			rumpuser_close(memfd);
		return error;
	}

	if (!path)
		goto out;

	error = initbackend(sc, memfd);
	if (error) {
		shmif_unclone(&sc->sc_ec.ec_if);
		return error;
	}

	sc->sc_backfilelen = strlen(path)+1;
	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
	strcpy(sc->sc_backfile, path);

 out:
	if (ifnum)
		*ifnum = unit;

	return 0;
}

static int
shmif_clone(struct if_clone *ifc, int unit)
{
	int rc __diagused;
	vmem_addr_t unit2;

	/*
	 * Ok, we know the unit number, but we must still reserve it.
	 * Otherwise the wildcard-side of things might get the same one.
	 * This is slightly offset-happy due to vmem.  First, we offset
	 * the range of unit numbers by +1 since vmem cannot deal with
	 * ranges starting from 0.  Talk about uuuh.
	 */
	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
	    VM_SLEEP | VM_INSTANTFIT, &unit2);
	KASSERT(rc == 0 && unit2-1 == unit);

	return allocif(unit, NULL);
}

static int
shmif_unclone(struct ifnet *ifp)
{
	struct shmif_sc *sc = ifp->if_softc;

	shmif_stop(ifp, 1);
	if_down(ifp);

	mutex_enter(&sc->sc_mtx);
	sc->sc_dying = true;
	cv_broadcast(&sc->sc_cv);
	mutex_exit(&sc->sc_mtx);

	if (sc->sc_rcvl)
		kthread_join(sc->sc_rcvl);
	sc->sc_rcvl = NULL;

	/*
	 * Need to be called after the kthread left, otherwise closing kqueue
	 * (sc_kq) hangs sometimes perhaps because of a race condition between
	 * close and kevent in the kthread on the kqueue.
	 */
	finibackend(sc);

	vmem_xfree(shmif_units, sc->sc_unit+1, 1);

	ether_ifdetach(ifp);
	if_detach(ifp);

	cv_destroy(&sc->sc_cv);
	mutex_destroy(&sc->sc_mtx);

	kmem_free(sc, sizeof(*sc));

	return 0;
}

static int
shmif_init(struct ifnet *ifp)
{
	struct shmif_sc *sc = ifp->if_softc;
	int error = 0;

	if (sc->sc_memfd == -1)
		return ENXIO;
	KASSERT(sc->sc_busmem);

	ifp->if_flags |= IFF_RUNNING;

	mutex_enter(&sc->sc_mtx);
	sc->sc_nextpacket = sc->sc_busmem->shm_last;
	sc->sc_devgen = sc->sc_busmem->shm_gen;

	cv_broadcast(&sc->sc_cv);
	mutex_exit(&sc->sc_mtx);

	return error;
}

static int
shmif_mediachange(struct ifnet *ifp)
{
	struct shmif_sc *sc = ifp->if_softc;
	int link_state;

	if (IFM_SUBTYPE(sc->sc_im.ifm_cur->ifm_media) == IFM_NONE)
		link_state = LINK_STATE_DOWN;
	else
		link_state = LINK_STATE_UP;

	if_link_state_change(ifp, link_state);
	return 0;
}

static void
shmif_mediastatus(struct ifnet *ifp, struct ifmediareq *imr)
{
	struct shmif_sc *sc = ifp->if_softc;

	imr->ifm_active = sc->sc_im.ifm_cur->ifm_media;

	imr->ifm_status = IFM_AVALID;
	if (IFM_SUBTYPE(imr->ifm_active) != IFM_NONE)
		imr->ifm_status |= IFM_ACTIVE;
}

static int
shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
{
	struct shmif_sc *sc = ifp->if_softc;
	struct ifdrv *ifd;
	char *path;
	int s, rv, memfd;

	s = splnet();
	switch (cmd) {
	case SIOCGLINKSTR:
		ifd = data;

		if (sc->sc_backfilelen == 0) {
			rv = ENOENT;
			break;
		}

		ifd->ifd_len = sc->sc_backfilelen;
		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
			rv = 0;
			break;
		}

		if (ifd->ifd_cmd != 0) {
			rv = EINVAL;
			break;
		}

		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
		break;
	case SIOCSLINKSTR:
		if (ifp->if_flags & IFF_UP) {
			rv = EBUSY;
			break;
		}

		ifd = data;
		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
			finibackend(sc);
			/* Back to the default just in case */
			ifp->if_link_state = LINK_STATE_UNKNOWN;
			rv = 0;
			break;
		} else if (ifd->ifd_cmd != 0) {
			rv = EINVAL;
			break;
		} else if (sc->sc_backfile) {
			rv = EBUSY;
			break;
		}

		if (ifd->ifd_len > MAXPATHLEN) {
			rv = E2BIG;
			break;
		} else if (ifd->ifd_len < 1) {
			rv = EINVAL;
			break;
		}

		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
		if (rv) {
			kmem_free(path, ifd->ifd_len);
			break;
		}
		rv = rumpuser_open(path,
		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
		if (rv) {
			kmem_free(path, ifd->ifd_len);
			break;
		}
		rv = initbackend(sc, memfd);
		if (rv) {
			kmem_free(path, ifd->ifd_len);
			rumpuser_close(memfd);
			break;
		}
		sc->sc_backfile = path;
		sc->sc_backfilelen = ifd->ifd_len;

		if_link_state_change(ifp, LINK_STATE_UP);
		break;

#ifdef OSIOCSIFMEDIA
	case OSIOCSIFMEDIA:
#endif
	case SIOCSIFMEDIA:
	case SIOCGIFMEDIA:
		rv = ifmedia_ioctl(ifp, data, &sc->sc_im, cmd);
		break;

	default:
		rv = ether_ioctl(ifp, cmd, data);
		if (rv == ENETRESET)
			rv = 0;
		break;
	}
	splx(s);

	return rv;
}

static void
shmif_start(struct ifnet *ifp)
{
	struct shmif_sc *sc = ifp->if_softc;
	struct mbuf *m, *n;
	bool wrote = false;

	ifp->if_flags |= IFF_OACTIVE;

	for (;;) {
		IFQ_DEQUEUE(&ifp->if_snd, m);
		if (m == NULL)
			break;

		m = ether_sw_offload_tx(ifp, m);
		if (m == NULL) {
			if_statinc(ifp, if_oerrors);
			break;
		}

		do {
			n = m->m_nextpkt;
			shmif_snd(ifp, m);
			m = n;
		} while (m != NULL);

		wrote = true;
	}

	ifp->if_flags &= ~IFF_OACTIVE;

	/* wakeup? */
	if (wrote) {
		dowakeup(sc);
	}
}

/* send everything in-context since it's just a matter of mem-to-mem copy */
static void
shmif_snd(struct ifnet *ifp, struct mbuf *m0)
{
	struct shmif_sc *sc = ifp->if_softc;
	struct shmif_mem *busmem = sc->sc_busmem;
	struct shmif_pkthdr sp;
	struct timeval tv;
	struct mbuf *m;
	uint32_t dataoff;
	uint32_t pktsize, pktwrote;
	bool wrap;

	pktsize = 0;
	for (m = m0; m != NULL; m = m->m_next) {
		pktsize += m->m_len;
	}
	KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);

	getmicrouptime(&tv);
	sp.sp_len = pktsize;
	sp.sp_sec = tv.tv_sec;
	sp.sp_usec = tv.tv_usec;
	sp.sp_sender = sc->sc_uid;

	bpf_mtap(ifp, m0, BPF_D_OUT);

	/*
	 * Compare with DOWN to allow UNKNOWN (the default value),
	 * which is required by some ATF tests using rump servers
	 * written in C.
	 */
	if (ifp->if_link_state == LINK_STATE_DOWN)
		goto dontsend;

	shmif_lockbus(busmem);
	KASSERT(busmem->shm_magic == SHMIF_MAGIC);
	busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);

	wrap = false;
	dataoff =
	    shmif_buswrite(busmem, busmem->shm_last, &sp, sizeof(sp), &wrap);
	pktwrote = 0;
	for (m = m0; m != NULL; m = m->m_next) {
		pktwrote += m->m_len;
		dataoff = shmif_buswrite(busmem, dataoff, mtod(m, void *),
		    m->m_len, &wrap);
	}
	KASSERT(pktwrote == pktsize);
	if (wrap) {
		busmem->shm_gen++;
		DPRINTF(("bus generation now %" PRIu64 "\n", busmem->shm_gen));
	}
	shmif_unlockbus(busmem);

dontsend:
	m_freem(m0);
	if_statinc(ifp, if_opackets);

	DPRINTF(("shmif_start: send %d bytes at off %d\n", pktsize,
	    busmem->shm_last));
}

static void
shmif_stop(struct ifnet *ifp, int disable)
{
	struct shmif_sc *sc = ifp->if_softc;

	ifp->if_flags &= ~IFF_RUNNING;
	membar_producer();

	/*
	 * wakeup thread.  this will of course wake up all bus
	 * listeners, but that's life.
	 */
	if (sc->sc_memfd != -1) {
		dowakeup(sc);
	}
}


/*
 * Check if we have been sleeping too long.  Basically,
 * our in-sc nextpkt must by first <= nextpkt <= last"+1".
 * We use the fact that first is guaranteed to never overlap
 * with the last frame in the ring.
 */
static __inline bool
stillvalid_p(struct shmif_sc *sc)
{
	struct shmif_mem *busmem = sc->sc_busmem;
	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
	uint32_t lastoff, devoff;

	KASSERT(busmem->shm_first != busmem->shm_last);

	/* normalize onto a 2x busmem chunk */
	devoff = sc->sc_nextpacket;
	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);

	/* trivial case */
	if (gendiff > 1)
		return false;
	KASSERT(gendiff <= 1);

	/* Normalize onto 2x busmem chunk */
	if (busmem->shm_first >= lastoff) {
		lastoff += BUSMEM_DATASIZE;
		if (gendiff == 0)
			devoff += BUSMEM_DATASIZE;
	} else {
		if (gendiff)
			return false;
	}

	return devoff >= busmem->shm_first && devoff <= lastoff;
}

static void
shmif_rcv(void *arg)
{
	struct ifnet *ifp = arg;
	struct shmif_sc *sc = ifp->if_softc;
	struct shmif_mem *busmem;
	struct mbuf *m = NULL;
	struct ether_header *eth;
	uint32_t nextpkt;
	bool wrap, passup;
	int error;
	const int align
	    = ALIGN(sizeof(struct ether_header)) - sizeof(struct ether_header);

 reup:
	mutex_enter(&sc->sc_mtx);
	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
		cv_wait(&sc->sc_cv, &sc->sc_mtx);
	mutex_exit(&sc->sc_mtx);

	busmem = sc->sc_busmem;

	while (ifp->if_flags & IFF_RUNNING) {
		struct shmif_pkthdr sp;

		if (m == NULL) {
			m = m_gethdr(M_WAIT, MT_DATA);
			MCLGET(m, M_WAIT);
			m->m_data += align;
		}

		DPRINTF(("waiting %d/%" PRIu64 "\n",
		    sc->sc_nextpacket, sc->sc_devgen));
		KASSERT(m->m_flags & M_EXT);

		shmif_lockbus(busmem);
		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
		KASSERT(busmem->shm_gen >= sc->sc_devgen);

		/* need more data? */
		if (sc->sc_devgen == busmem->shm_gen &&
		    shmif_nextpktoff(busmem, busmem->shm_last)
		     == sc->sc_nextpacket) {
			shmif_unlockbus(busmem);
			error = rumpcomp_shmif_watchwait(sc->sc_kq);
			if (__predict_false(error))
				printf("shmif_rcv: wait failed %d\n", error);
			membar_consumer();
			continue;
		}

		if (stillvalid_p(sc)) {
			nextpkt = sc->sc_nextpacket;
		} else {
			KASSERT(busmem->shm_gen > 0);
			nextpkt = busmem->shm_first;
			if (busmem->shm_first > busmem->shm_last)
				sc->sc_devgen = busmem->shm_gen - 1;
			else
				sc->sc_devgen = busmem->shm_gen;
			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
			    sc, nextpkt, sc->sc_devgen));
		}

		/*
		 * If our read pointer is ahead the bus last write, our
		 * generation must be one behind.
		 */
		KASSERT(!(nextpkt > busmem->shm_last
		    && sc->sc_devgen == busmem->shm_gen));

		wrap = false;
		nextpkt = shmif_busread(busmem, &sp,
		    nextpkt, sizeof(sp), &wrap);
		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
		nextpkt = shmif_busread(busmem, mtod(m, void *),
		    nextpkt, sp.sp_len, &wrap);

		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
		    sp.sp_len, nextpkt));

		sc->sc_nextpacket = nextpkt;
		shmif_unlockbus(sc->sc_busmem);

		if (wrap) {
			sc->sc_devgen++;
			DPRINTF(("dev %p generation now %" PRIu64 "\n",
			    sc, sc->sc_devgen));
		}

		/*
		 * Ignore packets too short to possibly be valid.
		 * This is hit at least for the first frame on a new bus.
		 */
		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
			    sp.sp_len));
			continue;
		}

		m->m_len = m->m_pkthdr.len = sp.sp_len;
		m_set_rcvif(m, ifp);

		/*
		 * Test if we want to pass the packet upwards
		 */
		eth = mtod(m, struct ether_header *);
		/*
		 * Compare with DOWN to allow UNKNOWN (the default value),
		 * which is required by some ATF tests using rump servers
		 * written in C.
		 */
		if (ifp->if_link_state == LINK_STATE_DOWN) {
			passup = false;
		} else if (sp.sp_sender == sc->sc_uid) {
			passup = false;
		} else if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
		    ETHER_ADDR_LEN) == 0) {
			passup = true;
		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
			passup = true;
		} else if (ifp->if_flags & IFF_PROMISC) {
			m->m_flags |= M_PROMISC;
			passup = true;
		} else {
			passup = false;
		}

		if (passup) {
			int bound;

			m = ether_sw_offload_rx(ifp, m);

			KERNEL_LOCK(1, NULL);
			/* Prevent LWP migrations between CPUs for psref(9) */
			bound = curlwp_bind();
			if_input(ifp, m);
			curlwp_bindx(bound);
			KERNEL_UNLOCK_ONE(NULL);

			m = NULL;
		}
		/* else: reuse mbuf for a future packet */
	}
	m_freem(m);
	m = NULL;

	if (!sc->sc_dying)
		goto reup;

	kthread_exit(0);
}