Home | History | Annotate | Line # | Download | only in libshmif
if_shmem.c revision 1.81.10.1
      1  1.81.10.1   thorpej /*	$NetBSD: if_shmem.c,v 1.81.10.1 2021/06/17 04:46:36 thorpej Exp $	*/
      2        1.1     pooka 
      3        1.1     pooka /*
      4       1.39     pooka  * Copyright (c) 2009, 2010 Antti Kantee.  All Rights Reserved.
      5        1.1     pooka  *
      6        1.1     pooka  * Development of this software was supported by The Nokia Foundation.
      7        1.1     pooka  *
      8        1.1     pooka  * Redistribution and use in source and binary forms, with or without
      9        1.1     pooka  * modification, are permitted provided that the following conditions
     10        1.1     pooka  * are met:
     11        1.1     pooka  * 1. Redistributions of source code must retain the above copyright
     12        1.1     pooka  *    notice, this list of conditions and the following disclaimer.
     13        1.1     pooka  * 2. Redistributions in binary form must reproduce the above copyright
     14        1.1     pooka  *    notice, this list of conditions and the following disclaimer in the
     15        1.1     pooka  *    documentation and/or other materials provided with the distribution.
     16        1.1     pooka  *
     17        1.1     pooka  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
     18        1.1     pooka  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     19        1.1     pooka  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
     20        1.1     pooka  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     21        1.1     pooka  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     22        1.1     pooka  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     23        1.1     pooka  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     24        1.1     pooka  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     25        1.1     pooka  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     26        1.1     pooka  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     27        1.1     pooka  * SUCH DAMAGE.
     28        1.1     pooka  */
     29        1.1     pooka 
     30        1.1     pooka #include <sys/cdefs.h>
     31  1.81.10.1   thorpej __KERNEL_RCSID(0, "$NetBSD: if_shmem.c,v 1.81.10.1 2021/06/17 04:46:36 thorpej Exp $");
     32        1.1     pooka 
     33        1.1     pooka #include <sys/param.h>
     34       1.13     pooka #include <sys/atomic.h>
     35        1.1     pooka #include <sys/fcntl.h>
     36        1.1     pooka #include <sys/kmem.h>
     37        1.1     pooka #include <sys/kthread.h>
     38        1.1     pooka #include <sys/lock.h>
     39       1.31     pooka #include <sys/vmem.h>
     40       1.44       tls #include <sys/cprng.h>
     41        1.1     pooka 
     42       1.35     pooka #include <net/bpf.h>
     43        1.1     pooka #include <net/if.h>
     44       1.37     pooka #include <net/if_dl.h>
     45        1.1     pooka #include <net/if_ether.h>
     46       1.76       rin #include <net/ether_sw_offload.h>
     47        1.1     pooka 
     48        1.1     pooka #include <netinet/in.h>
     49        1.1     pooka #include <netinet/in_var.h>
     50        1.1     pooka 
     51       1.64     pooka #include <rump-sys/kern.h>
     52       1.64     pooka #include <rump-sys/net.h>
     53       1.64     pooka 
     54        1.1     pooka #include <rump/rump.h>
     55        1.1     pooka #include <rump/rumpuser.h>
     56        1.1     pooka 
     57       1.59     pooka #include "shmif_user.h"
     58        1.1     pooka 
     59       1.29     pooka static int shmif_clone(struct if_clone *, int);
     60       1.29     pooka static int shmif_unclone(struct ifnet *);
     61       1.29     pooka 
     62       1.29     pooka struct if_clone shmif_cloner =
     63       1.29     pooka     IF_CLONE_INITIALIZER("shmif", shmif_clone, shmif_unclone);
     64       1.29     pooka 
     65        1.1     pooka /*
     66       1.28     pooka  * Do r/w prefault for backend pages when attaching the interface.
     67       1.29     pooka  * At least logically thinking improves performance (although no
     68       1.29     pooka  * mlocking is done, so they might go away).
     69       1.28     pooka  */
     70       1.28     pooka #define PREFAULT_RW
     71       1.28     pooka 
     72       1.28     pooka /*
     73        1.1     pooka  * A virtual ethernet interface which uses shared memory from a
     74        1.1     pooka  * memory mapped file as the bus.
     75        1.1     pooka  */
     76        1.1     pooka 
     77        1.1     pooka static int	shmif_init(struct ifnet *);
     78        1.1     pooka static int	shmif_ioctl(struct ifnet *, u_long, void *);
     79        1.1     pooka static void	shmif_start(struct ifnet *);
     80       1.76       rin static void	shmif_snd(struct ifnet *, struct mbuf *);
     81        1.1     pooka static void	shmif_stop(struct ifnet *, int);
     82        1.1     pooka 
     83       1.16     pooka #include "shmifvar.h"
     84       1.16     pooka 
     85        1.1     pooka struct shmif_sc {
     86        1.1     pooka 	struct ethercom sc_ec;
     87       1.16     pooka 	struct shmif_mem *sc_busmem;
     88        1.1     pooka 	int sc_memfd;
     89        1.1     pooka 	int sc_kq;
     90       1.32     pooka 	int sc_unit;
     91        1.1     pooka 
     92       1.29     pooka 	char *sc_backfile;
     93       1.29     pooka 	size_t sc_backfilelen;
     94       1.29     pooka 
     95       1.26     pooka 	uint64_t sc_devgen;
     96        1.1     pooka 	uint32_t sc_nextpacket;
     97       1.32     pooka 
     98       1.32     pooka 	kmutex_t sc_mtx;
     99       1.32     pooka 	kcondvar_t sc_cv;
    100       1.32     pooka 
    101       1.32     pooka 	struct lwp *sc_rcvl;
    102       1.32     pooka 	bool sc_dying;
    103       1.63     ozaki 
    104       1.80     ozaki 	uint64_t sc_uid;
    105        1.1     pooka };
    106        1.1     pooka 
    107        1.1     pooka static void shmif_rcv(void *);
    108        1.1     pooka 
    109       1.23     pooka #define LOCK_UNLOCKED	0
    110       1.23     pooka #define LOCK_LOCKED	1
    111       1.23     pooka #define LOCK_COOLDOWN	1001
    112       1.23     pooka 
    113       1.31     pooka vmem_t *shmif_units;
    114       1.31     pooka 
    115       1.52     pooka static void
    116       1.52     pooka dowakeup(struct shmif_sc *sc)
    117       1.52     pooka {
    118       1.52     pooka 	struct rumpuser_iovec iov;
    119       1.52     pooka 	uint32_t ver = SHMIF_VERSION;
    120       1.53     pooka 	size_t n;
    121       1.52     pooka 
    122       1.52     pooka 	iov.iov_base = &ver;
    123       1.52     pooka 	iov.iov_len = sizeof(ver);
    124       1.53     pooka 	rumpuser_iovwrite(sc->sc_memfd, &iov, 1, IFMEM_WAKEUP, &n);
    125       1.52     pooka }
    126       1.52     pooka 
    127       1.23     pooka /*
    128       1.23     pooka  * This locking needs work and will misbehave severely if:
    129       1.23     pooka  * 1) the backing memory has to be paged in
    130       1.23     pooka  * 2) some lockholder exits while holding the lock
    131       1.23     pooka  */
    132       1.23     pooka static void
    133       1.23     pooka shmif_lockbus(struct shmif_mem *busmem)
    134       1.23     pooka {
    135       1.23     pooka 	int i = 0;
    136       1.23     pooka 
    137       1.23     pooka 	while (__predict_false(atomic_cas_32(&busmem->shm_lock,
    138       1.23     pooka 	    LOCK_UNLOCKED, LOCK_LOCKED) == LOCK_LOCKED)) {
    139       1.23     pooka 		if (__predict_false(++i > LOCK_COOLDOWN)) {
    140       1.50     pooka 			/* wait 1ms */
    141       1.54     pooka 			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL,
    142       1.54     pooka 			    0, 1000*1000);
    143       1.23     pooka 			i = 0;
    144       1.23     pooka 		}
    145       1.23     pooka 		continue;
    146       1.23     pooka 	}
    147       1.23     pooka 	membar_enter();
    148       1.23     pooka }
    149       1.23     pooka 
    150       1.23     pooka static void
    151       1.23     pooka shmif_unlockbus(struct shmif_mem *busmem)
    152       1.23     pooka {
    153       1.61    justin 	unsigned int old __diagused;
    154       1.23     pooka 
    155       1.23     pooka 	membar_exit();
    156       1.23     pooka 	old = atomic_swap_32(&busmem->shm_lock, LOCK_UNLOCKED);
    157       1.23     pooka 	KASSERT(old == LOCK_LOCKED);
    158       1.23     pooka }
    159       1.23     pooka 
    160       1.29     pooka static int
    161       1.29     pooka allocif(int unit, struct shmif_sc **scp)
    162        1.1     pooka {
    163       1.29     pooka 	uint8_t enaddr[ETHER_ADDR_LEN] = { 0xb2, 0xa0, 0x00, 0x00, 0x00, 0x00 };
    164        1.1     pooka 	struct shmif_sc *sc;
    165        1.1     pooka 	struct ifnet *ifp;
    166       1.81     ozaki 	uint64_t randnum;
    167  1.81.10.1   thorpej 	int error = 0;
    168        1.1     pooka 
    169       1.81     ozaki 	randnum = cprng_strong64();
    170       1.81     ozaki 	memcpy(&enaddr[2], &randnum, 4);
    171        1.1     pooka 
    172        1.1     pooka 	sc = kmem_zalloc(sizeof(*sc), KM_SLEEP);
    173       1.29     pooka 	sc->sc_memfd = -1;
    174       1.32     pooka 	sc->sc_unit = unit;
    175       1.81     ozaki 	sc->sc_uid = randnum;
    176       1.29     pooka 
    177        1.1     pooka 	ifp = &sc->sc_ec.ec_if;
    178        1.1     pooka 
    179       1.60  christos 	snprintf(ifp->if_xname, sizeof(ifp->if_xname), "shmif%d", unit);
    180       1.29     pooka 	ifp->if_softc = sc;
    181       1.63     ozaki 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
    182       1.29     pooka 	ifp->if_init = shmif_init;
    183       1.29     pooka 	ifp->if_ioctl = shmif_ioctl;
    184       1.29     pooka 	ifp->if_start = shmif_start;
    185       1.29     pooka 	ifp->if_stop = shmif_stop;
    186       1.29     pooka 	ifp->if_mtu = ETHERMTU;
    187       1.37     pooka 	ifp->if_dlt = DLT_EN10MB;
    188       1.76       rin 	ifp->if_capabilities = IFCAP_TSOv4 | IFCAP_TSOv6 |
    189       1.76       rin 	    IFCAP_CSUM_IPv4_Rx	| IFCAP_CSUM_IPv4_Tx |
    190       1.76       rin 	    IFCAP_CSUM_TCPv4_Rx	| IFCAP_CSUM_TCPv4_Tx |
    191       1.76       rin 	    IFCAP_CSUM_UDPv4_Rx	| IFCAP_CSUM_UDPv4_Tx |
    192       1.76       rin 	    IFCAP_CSUM_TCPv6_Rx	| IFCAP_CSUM_TCPv6_Tx |
    193       1.76       rin 	    IFCAP_CSUM_UDPv6_Rx	| IFCAP_CSUM_UDPv6_Tx;
    194       1.29     pooka 
    195       1.32     pooka 	mutex_init(&sc->sc_mtx, MUTEX_DEFAULT, IPL_NONE);
    196       1.32     pooka 	cv_init(&sc->sc_cv, "shmifcv");
    197       1.32     pooka 
    198  1.81.10.1   thorpej 	if_initialize(ifp);
    199       1.76       rin #if 1
    200       1.76       rin 	char buf[256];
    201       1.76       rin 
    202       1.76       rin 	if (rumpuser_getparam("RUMP_SHMIF_CAPENABLE", buf, sizeof(buf)) == 0) {
    203       1.76       rin 		uint64_t capen = strtoul(buf, NULL, 0);
    204       1.76       rin 
    205       1.76       rin 		ifp->if_capenable = capen & ifp->if_capabilities;
    206       1.76       rin 	}
    207       1.76       rin #endif
    208       1.76       rin 
    209       1.29     pooka 	ether_ifattach(ifp, enaddr);
    210       1.65     ozaki 	if_register(ifp);
    211       1.29     pooka 
    212       1.29     pooka 	aprint_verbose("shmif%d: Ethernet address %s\n",
    213       1.32     pooka 	    unit, ether_sprintf(enaddr));
    214       1.29     pooka 
    215       1.29     pooka 	if (scp)
    216       1.29     pooka 		*scp = sc;
    217       1.29     pooka 
    218       1.32     pooka 	if (rump_threads) {
    219       1.32     pooka 		error = kthread_create(PRI_NONE,
    220       1.40     rmind 		    KTHREAD_MPSAFE | KTHREAD_MUSTJOIN, NULL,
    221       1.32     pooka 		    shmif_rcv, ifp, &sc->sc_rcvl, "shmif");
    222       1.32     pooka 	} else {
    223       1.32     pooka 		printf("WARNING: threads not enabled, shmif NOT working\n");
    224       1.32     pooka 	}
    225       1.32     pooka 
    226       1.32     pooka 	if (error) {
    227       1.32     pooka 		shmif_unclone(ifp);
    228       1.32     pooka 	}
    229       1.32     pooka 
    230  1.81.10.1   thorpej 	return 0;
    231       1.29     pooka }
    232       1.29     pooka 
    233       1.29     pooka static int
    234       1.29     pooka initbackend(struct shmif_sc *sc, int memfd)
    235       1.29     pooka {
    236       1.29     pooka 	volatile uint8_t v;
    237       1.29     pooka 	volatile uint8_t *p;
    238       1.53     pooka 	void *mem;
    239       1.29     pooka 	int error;
    240       1.29     pooka 
    241       1.53     pooka 	error = rumpcomp_shmif_mmap(memfd, BUSMEM_SIZE, &mem);
    242        1.1     pooka 	if (error)
    243       1.29     pooka 		return error;
    244       1.53     pooka 	sc->sc_busmem = mem;
    245       1.17     pooka 
    246       1.29     pooka 	if (sc->sc_busmem->shm_magic
    247       1.29     pooka 	    && sc->sc_busmem->shm_magic != SHMIF_MAGIC) {
    248       1.29     pooka 		printf("bus is not magical");
    249       1.29     pooka 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    250       1.69   msaitoh 		return ENOEXEC;
    251       1.29     pooka 	}
    252       1.28     pooka 
    253       1.36     pooka 	/*
    254       1.36     pooka 	 * Prefault in pages to minimize runtime penalty with buslock.
    255       1.36     pooka 	 * Use 512 instead of PAGE_SIZE to make sure we catch cases where
    256       1.36     pooka 	 * rump kernel PAGE_SIZE > host page size.
    257       1.36     pooka 	 */
    258       1.28     pooka 	for (p = (uint8_t *)sc->sc_busmem;
    259       1.28     pooka 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    260       1.36     pooka 	    p += 512)
    261       1.28     pooka 		v = *p;
    262       1.28     pooka 
    263       1.19     pooka 	shmif_lockbus(sc->sc_busmem);
    264       1.19     pooka 	/* we're first?  initialize bus */
    265       1.19     pooka 	if (sc->sc_busmem->shm_magic == 0) {
    266       1.19     pooka 		sc->sc_busmem->shm_magic = SHMIF_MAGIC;
    267       1.19     pooka 		sc->sc_busmem->shm_first = BUSMEM_DATASIZE;
    268       1.19     pooka 	}
    269       1.19     pooka 
    270       1.16     pooka 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    271       1.26     pooka 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    272       1.28     pooka 
    273       1.28     pooka #ifdef PREFAULT_RW
    274       1.28     pooka 	for (p = (uint8_t *)sc->sc_busmem;
    275       1.28     pooka 	    p < (uint8_t *)sc->sc_busmem + BUSMEM_SIZE;
    276       1.28     pooka 	    p += PAGE_SIZE) {
    277       1.28     pooka 		v = *p;
    278       1.28     pooka 		*p = v;
    279       1.28     pooka 	}
    280       1.28     pooka #endif
    281       1.19     pooka 	shmif_unlockbus(sc->sc_busmem);
    282        1.1     pooka 
    283       1.53     pooka 	sc->sc_kq = -1;
    284       1.53     pooka 	error = rumpcomp_shmif_watchsetup(&sc->sc_kq, memfd);
    285       1.53     pooka 	if (error) {
    286       1.32     pooka 		rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    287       1.29     pooka 		return error;
    288       1.32     pooka 	}
    289        1.1     pooka 
    290       1.29     pooka 	sc->sc_memfd = memfd;
    291       1.32     pooka 
    292       1.32     pooka 	return error;
    293       1.29     pooka }
    294       1.29     pooka 
    295       1.29     pooka static void
    296       1.29     pooka finibackend(struct shmif_sc *sc)
    297       1.29     pooka {
    298       1.29     pooka 
    299       1.32     pooka 	if (sc->sc_backfile == NULL)
    300       1.32     pooka 		return;
    301       1.32     pooka 
    302       1.32     pooka 	if (sc->sc_backfile) {
    303       1.32     pooka 		kmem_free(sc->sc_backfile, sc->sc_backfilelen);
    304       1.32     pooka 		sc->sc_backfile = NULL;
    305       1.32     pooka 		sc->sc_backfilelen = 0;
    306       1.32     pooka 	}
    307       1.29     pooka 
    308       1.29     pooka 	rumpuser_unmap(sc->sc_busmem, BUSMEM_SIZE);
    309       1.53     pooka 	rumpuser_close(sc->sc_memfd);
    310       1.53     pooka 	rumpuser_close(sc->sc_kq);
    311       1.32     pooka 
    312       1.32     pooka 	sc->sc_memfd = -1;
    313       1.29     pooka }
    314       1.29     pooka 
    315       1.29     pooka int
    316       1.29     pooka rump_shmif_create(const char *path, int *ifnum)
    317       1.29     pooka {
    318       1.29     pooka 	struct shmif_sc *sc;
    319       1.43    dyoung 	vmem_addr_t t;
    320       1.33     pooka 	int unit, error;
    321       1.33     pooka 	int memfd = -1; /* XXXgcc */
    322       1.29     pooka 
    323       1.33     pooka 	if (path) {
    324       1.53     pooka 		error = rumpuser_open(path,
    325       1.53     pooka 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    326       1.53     pooka 		if (error)
    327       1.33     pooka 			return error;
    328       1.33     pooka 	}
    329        1.1     pooka 
    330       1.43    dyoung 	error = vmem_xalloc(shmif_units, 1, 0, 0, 0,
    331       1.43    dyoung 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX, VM_INSTANTFIT | VM_SLEEP, &t);
    332       1.43    dyoung 
    333       1.43    dyoung 	if (error != 0) {
    334       1.43    dyoung 		if (path)
    335       1.53     pooka 			rumpuser_close(memfd);
    336       1.43    dyoung 		return error;
    337       1.43    dyoung 	}
    338       1.43    dyoung 
    339       1.43    dyoung 	unit = t - 1;
    340       1.31     pooka 
    341       1.32     pooka 	if ((error = allocif(unit, &sc)) != 0) {
    342       1.33     pooka 		if (path)
    343       1.53     pooka 			rumpuser_close(memfd);
    344       1.29     pooka 		return error;
    345       1.29     pooka 	}
    346       1.33     pooka 
    347       1.33     pooka 	if (!path)
    348       1.33     pooka 		goto out;
    349       1.33     pooka 
    350       1.29     pooka 	error = initbackend(sc, memfd);
    351       1.29     pooka 	if (error) {
    352       1.32     pooka 		shmif_unclone(&sc->sc_ec.ec_if);
    353       1.29     pooka 		return error;
    354       1.29     pooka 	}
    355        1.1     pooka 
    356       1.29     pooka 	sc->sc_backfilelen = strlen(path)+1;
    357       1.29     pooka 	sc->sc_backfile = kmem_alloc(sc->sc_backfilelen, KM_SLEEP);
    358       1.29     pooka 	strcpy(sc->sc_backfile, path);
    359       1.12     pooka 
    360       1.33     pooka  out:
    361        1.2     pooka 	if (ifnum)
    362       1.32     pooka 		*ifnum = unit;
    363       1.29     pooka 
    364        1.1     pooka 	return 0;
    365       1.29     pooka }
    366        1.1     pooka 
    367       1.29     pooka static int
    368       1.29     pooka shmif_clone(struct if_clone *ifc, int unit)
    369       1.29     pooka {
    370       1.61    justin 	int rc __diagused;
    371       1.43    dyoung 	vmem_addr_t unit2;
    372       1.29     pooka 
    373       1.31     pooka 	/*
    374       1.31     pooka 	 * Ok, we know the unit number, but we must still reserve it.
    375       1.31     pooka 	 * Otherwise the wildcard-side of things might get the same one.
    376       1.31     pooka 	 * This is slightly offset-happy due to vmem.  First, we offset
    377       1.31     pooka 	 * the range of unit numbers by +1 since vmem cannot deal with
    378       1.41    dyoung 	 * ranges starting from 0.  Talk about uuuh.
    379       1.31     pooka 	 */
    380       1.43    dyoung 	rc = vmem_xalloc(shmif_units, 1, 0, 0, 0, unit+1, unit+1,
    381       1.43    dyoung 	    VM_SLEEP | VM_INSTANTFIT, &unit2);
    382       1.43    dyoung 	KASSERT(rc == 0 && unit2-1 == unit);
    383       1.29     pooka 
    384       1.29     pooka 	return allocif(unit, NULL);
    385       1.29     pooka }
    386       1.29     pooka 
    387       1.29     pooka static int
    388       1.29     pooka shmif_unclone(struct ifnet *ifp)
    389       1.29     pooka {
    390       1.32     pooka 	struct shmif_sc *sc = ifp->if_softc;
    391       1.32     pooka 
    392       1.32     pooka 	shmif_stop(ifp, 1);
    393       1.32     pooka 	if_down(ifp);
    394       1.32     pooka 
    395       1.32     pooka 	mutex_enter(&sc->sc_mtx);
    396       1.32     pooka 	sc->sc_dying = true;
    397       1.32     pooka 	cv_broadcast(&sc->sc_cv);
    398       1.32     pooka 	mutex_exit(&sc->sc_mtx);
    399       1.29     pooka 
    400       1.32     pooka 	if (sc->sc_rcvl)
    401       1.32     pooka 		kthread_join(sc->sc_rcvl);
    402       1.32     pooka 	sc->sc_rcvl = NULL;
    403       1.32     pooka 
    404       1.72     ozaki 	/*
    405       1.72     ozaki 	 * Need to be called after the kthread left, otherwise closing kqueue
    406       1.72     ozaki 	 * (sc_kq) hangs sometimes perhaps because of a race condition between
    407       1.72     ozaki 	 * close and kevent in the kthread on the kqueue.
    408       1.72     ozaki 	 */
    409       1.72     ozaki 	finibackend(sc);
    410       1.72     ozaki 
    411       1.32     pooka 	vmem_xfree(shmif_units, sc->sc_unit+1, 1);
    412       1.32     pooka 
    413       1.32     pooka 	ether_ifdetach(ifp);
    414       1.32     pooka 	if_detach(ifp);
    415       1.32     pooka 
    416       1.32     pooka 	cv_destroy(&sc->sc_cv);
    417       1.32     pooka 	mutex_destroy(&sc->sc_mtx);
    418       1.32     pooka 
    419       1.32     pooka 	kmem_free(sc, sizeof(*sc));
    420       1.32     pooka 
    421       1.32     pooka 	return 0;
    422        1.1     pooka }
    423        1.1     pooka 
    424        1.1     pooka static int
    425        1.1     pooka shmif_init(struct ifnet *ifp)
    426        1.1     pooka {
    427       1.29     pooka 	struct shmif_sc *sc = ifp->if_softc;
    428        1.4     pooka 	int error = 0;
    429        1.4     pooka 
    430       1.29     pooka 	if (sc->sc_memfd == -1)
    431       1.29     pooka 		return ENXIO;
    432       1.32     pooka 	KASSERT(sc->sc_busmem);
    433       1.29     pooka 
    434       1.32     pooka 	ifp->if_flags |= IFF_RUNNING;
    435       1.32     pooka 
    436       1.32     pooka 	mutex_enter(&sc->sc_mtx);
    437       1.32     pooka 	sc->sc_nextpacket = sc->sc_busmem->shm_last;
    438       1.32     pooka 	sc->sc_devgen = sc->sc_busmem->shm_gen;
    439       1.32     pooka 
    440       1.32     pooka 	cv_broadcast(&sc->sc_cv);
    441       1.32     pooka 	mutex_exit(&sc->sc_mtx);
    442        1.1     pooka 
    443        1.4     pooka 	return error;
    444        1.1     pooka }
    445        1.1     pooka 
    446        1.1     pooka static int
    447        1.1     pooka shmif_ioctl(struct ifnet *ifp, u_long cmd, void *data)
    448        1.1     pooka {
    449       1.29     pooka 	struct shmif_sc *sc = ifp->if_softc;
    450       1.29     pooka 	struct ifdrv *ifd;
    451       1.29     pooka 	char *path;
    452       1.32     pooka 	int s, rv, memfd;
    453        1.1     pooka 
    454        1.1     pooka 	s = splnet();
    455       1.29     pooka 	switch (cmd) {
    456       1.29     pooka 	case SIOCGLINKSTR:
    457       1.29     pooka 		ifd = data;
    458       1.29     pooka 
    459       1.29     pooka 		if (sc->sc_backfilelen == 0) {
    460       1.29     pooka 			rv = ENOENT;
    461       1.29     pooka 			break;
    462       1.29     pooka 		}
    463       1.29     pooka 
    464       1.29     pooka 		ifd->ifd_len = sc->sc_backfilelen;
    465       1.29     pooka 		if (ifd->ifd_cmd == IFLINKSTR_QUERYLEN) {
    466       1.29     pooka 			rv = 0;
    467       1.29     pooka 			break;
    468       1.29     pooka 		}
    469       1.29     pooka 
    470       1.29     pooka 		if (ifd->ifd_cmd != 0) {
    471       1.29     pooka 			rv = EINVAL;
    472       1.29     pooka 			break;
    473       1.29     pooka 		}
    474       1.29     pooka 
    475       1.29     pooka 		rv = copyoutstr(sc->sc_backfile, ifd->ifd_data,
    476       1.29     pooka 		    MIN(sc->sc_backfilelen, ifd->ifd_len), NULL);
    477       1.29     pooka 		break;
    478       1.29     pooka 	case SIOCSLINKSTR:
    479       1.29     pooka 		if (ifp->if_flags & IFF_UP) {
    480       1.29     pooka 			rv = EBUSY;
    481       1.29     pooka 			break;
    482       1.29     pooka 		}
    483       1.29     pooka 
    484       1.29     pooka 		ifd = data;
    485       1.29     pooka 		if (ifd->ifd_cmd == IFLINKSTR_UNSET) {
    486       1.29     pooka 			finibackend(sc);
    487       1.29     pooka 			rv = 0;
    488       1.29     pooka 			break;
    489       1.29     pooka 		} else if (ifd->ifd_cmd != 0) {
    490       1.29     pooka 			rv = EINVAL;
    491       1.29     pooka 			break;
    492       1.29     pooka 		} else if (sc->sc_backfile) {
    493       1.29     pooka 			rv = EBUSY;
    494       1.29     pooka 			break;
    495       1.29     pooka 		}
    496       1.29     pooka 
    497       1.29     pooka 		if (ifd->ifd_len > MAXPATHLEN) {
    498       1.29     pooka 			rv = E2BIG;
    499       1.29     pooka 			break;
    500       1.29     pooka 		} else if (ifd->ifd_len < 1) {
    501       1.29     pooka 			rv = EINVAL;
    502       1.29     pooka 			break;
    503       1.29     pooka 		}
    504       1.29     pooka 
    505       1.29     pooka 		path = kmem_alloc(ifd->ifd_len, KM_SLEEP);
    506       1.29     pooka 		rv = copyinstr(ifd->ifd_data, path, ifd->ifd_len, NULL);
    507       1.29     pooka 		if (rv) {
    508       1.29     pooka 			kmem_free(path, ifd->ifd_len);
    509       1.29     pooka 			break;
    510       1.29     pooka 		}
    511       1.53     pooka 		rv = rumpuser_open(path,
    512       1.53     pooka 		    RUMPUSER_OPEN_RDWR | RUMPUSER_OPEN_CREATE, &memfd);
    513       1.53     pooka 		if (rv) {
    514       1.29     pooka 			kmem_free(path, ifd->ifd_len);
    515       1.29     pooka 			break;
    516       1.29     pooka 		}
    517       1.29     pooka 		rv = initbackend(sc, memfd);
    518       1.29     pooka 		if (rv) {
    519       1.29     pooka 			kmem_free(path, ifd->ifd_len);
    520       1.53     pooka 			rumpuser_close(memfd);
    521       1.29     pooka 			break;
    522       1.29     pooka 		}
    523       1.29     pooka 		sc->sc_backfile = path;
    524       1.29     pooka 		sc->sc_backfilelen = ifd->ifd_len;
    525       1.29     pooka 
    526       1.29     pooka 		break;
    527       1.29     pooka 	default:
    528       1.29     pooka 		rv = ether_ioctl(ifp, cmd, data);
    529       1.29     pooka 		if (rv == ENETRESET)
    530       1.29     pooka 			rv = 0;
    531       1.29     pooka 		break;
    532       1.29     pooka 	}
    533        1.1     pooka 	splx(s);
    534        1.1     pooka 
    535        1.1     pooka 	return rv;
    536        1.1     pooka }
    537        1.1     pooka 
    538        1.1     pooka static void
    539        1.1     pooka shmif_start(struct ifnet *ifp)
    540        1.1     pooka {
    541        1.1     pooka 	struct shmif_sc *sc = ifp->if_softc;
    542       1.76       rin 	struct mbuf *m, *n;
    543        1.1     pooka 	bool wrote = false;
    544        1.1     pooka 
    545       1.26     pooka 	ifp->if_flags |= IFF_OACTIVE;
    546       1.26     pooka 
    547        1.1     pooka 	for (;;) {
    548       1.76       rin 		IF_DEQUEUE(&ifp->if_snd, m);
    549       1.76       rin 		if (m == NULL)
    550       1.76       rin 			break;
    551       1.20     pooka 
    552       1.76       rin 		m = ether_sw_offload_tx(ifp, m);
    553       1.76       rin 		if (m == NULL) {
    554       1.77   thorpej 			if_statinc(ifp, if_oerrors);
    555        1.1     pooka 			break;
    556        1.1     pooka 		}
    557        1.1     pooka 
    558       1.76       rin 		do {
    559       1.76       rin 			n = m->m_nextpkt;
    560       1.76       rin 			shmif_snd(ifp, m);
    561       1.76       rin 			m = n;
    562       1.76       rin 		} while (m != NULL);
    563        1.1     pooka 
    564        1.1     pooka 		wrote = true;
    565        1.1     pooka 	}
    566       1.26     pooka 
    567       1.26     pooka 	ifp->if_flags &= ~IFF_OACTIVE;
    568       1.26     pooka 
    569       1.32     pooka 	/* wakeup? */
    570       1.52     pooka 	if (wrote) {
    571       1.52     pooka 		dowakeup(sc);
    572       1.52     pooka 	}
    573        1.1     pooka }
    574        1.1     pooka 
    575       1.76       rin /* send everything in-context since it's just a matter of mem-to-mem copy */
    576       1.76       rin static void
    577       1.76       rin shmif_snd(struct ifnet *ifp, struct mbuf *m0)
    578       1.76       rin {
    579       1.76       rin 	struct shmif_sc *sc = ifp->if_softc;
    580       1.76       rin 	struct shmif_mem *busmem = sc->sc_busmem;
    581       1.76       rin 	struct shmif_pkthdr sp;
    582       1.76       rin 	struct timeval tv;
    583       1.76       rin 	struct mbuf *m;
    584       1.76       rin 	uint32_t dataoff;
    585       1.76       rin 	uint32_t pktsize, pktwrote;
    586       1.76       rin 	bool wrap;
    587       1.76       rin 
    588       1.76       rin 	pktsize = 0;
    589       1.76       rin 	for (m = m0; m != NULL; m = m->m_next) {
    590       1.76       rin 		pktsize += m->m_len;
    591       1.76       rin 	}
    592       1.76       rin 	KASSERT(pktsize <= ETHERMTU + ETHER_HDR_LEN);
    593       1.76       rin 
    594       1.76       rin 	getmicrouptime(&tv);
    595       1.76       rin 	sp.sp_len = pktsize;
    596       1.76       rin 	sp.sp_sec = tv.tv_sec;
    597       1.76       rin 	sp.sp_usec = tv.tv_usec;
    598       1.80     ozaki 	sp.sp_sender = sc->sc_uid;
    599       1.76       rin 
    600       1.76       rin 	bpf_mtap(ifp, m0, BPF_D_OUT);
    601       1.76       rin 
    602       1.76       rin 	shmif_lockbus(busmem);
    603       1.76       rin 	KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    604       1.76       rin 	busmem->shm_last = shmif_nextpktoff(busmem, busmem->shm_last);
    605       1.76       rin 
    606       1.76       rin 	wrap = false;
    607       1.76       rin 	dataoff =
    608       1.76       rin 	    shmif_buswrite(busmem, busmem->shm_last, &sp, sizeof(sp), &wrap);
    609       1.76       rin 	pktwrote = 0;
    610       1.76       rin 	for (m = m0; m != NULL; m = m->m_next) {
    611       1.76       rin 		pktwrote += m->m_len;
    612       1.76       rin 		dataoff = shmif_buswrite(busmem, dataoff, mtod(m, void *),
    613       1.76       rin 		    m->m_len, &wrap);
    614       1.76       rin 	}
    615       1.76       rin 	KASSERT(pktwrote == pktsize);
    616       1.76       rin 	if (wrap) {
    617       1.76       rin 		busmem->shm_gen++;
    618       1.76       rin 		DPRINTF(("bus generation now %" PRIu64 "\n", busmem->shm_gen));
    619       1.76       rin 	}
    620       1.76       rin 	shmif_unlockbus(busmem);
    621       1.76       rin 
    622       1.76       rin 	m_freem(m0);
    623       1.77   thorpej 	if_statinc(ifp, if_opackets);
    624       1.76       rin 
    625       1.76       rin 	DPRINTF(("shmif_start: send %d bytes at off %d\n", pktsize,
    626       1.76       rin 	    busmem->shm_last));
    627       1.76       rin }
    628       1.76       rin 
    629        1.1     pooka static void
    630        1.1     pooka shmif_stop(struct ifnet *ifp, int disable)
    631        1.1     pooka {
    632       1.32     pooka 	struct shmif_sc *sc = ifp->if_softc;
    633        1.1     pooka 
    634       1.32     pooka 	ifp->if_flags &= ~IFF_RUNNING;
    635       1.32     pooka 	membar_producer();
    636       1.32     pooka 
    637       1.32     pooka 	/*
    638       1.32     pooka 	 * wakeup thread.  this will of course wake up all bus
    639       1.32     pooka 	 * listeners, but that's life.
    640       1.32     pooka 	 */
    641       1.52     pooka 	if (sc->sc_memfd != -1) {
    642       1.52     pooka 		dowakeup(sc);
    643       1.52     pooka 	}
    644        1.1     pooka }
    645        1.1     pooka 
    646       1.27     pooka 
    647       1.27     pooka /*
    648       1.27     pooka  * Check if we have been sleeping too long.  Basically,
    649       1.27     pooka  * our in-sc nextpkt must by first <= nextpkt <= last"+1".
    650       1.27     pooka  * We use the fact that first is guaranteed to never overlap
    651       1.27     pooka  * with the last frame in the ring.
    652       1.27     pooka  */
    653       1.27     pooka static __inline bool
    654       1.27     pooka stillvalid_p(struct shmif_sc *sc)
    655       1.27     pooka {
    656       1.27     pooka 	struct shmif_mem *busmem = sc->sc_busmem;
    657       1.27     pooka 	unsigned gendiff = busmem->shm_gen - sc->sc_devgen;
    658       1.27     pooka 	uint32_t lastoff, devoff;
    659       1.27     pooka 
    660       1.27     pooka 	KASSERT(busmem->shm_first != busmem->shm_last);
    661       1.27     pooka 
    662       1.27     pooka 	/* normalize onto a 2x busmem chunk */
    663       1.27     pooka 	devoff = sc->sc_nextpacket;
    664       1.27     pooka 	lastoff = shmif_nextpktoff(busmem, busmem->shm_last);
    665       1.27     pooka 
    666       1.27     pooka 	/* trivial case */
    667       1.27     pooka 	if (gendiff > 1)
    668       1.27     pooka 		return false;
    669       1.27     pooka 	KASSERT(gendiff <= 1);
    670       1.27     pooka 
    671       1.27     pooka 	/* Normalize onto 2x busmem chunk */
    672       1.27     pooka 	if (busmem->shm_first >= lastoff) {
    673       1.27     pooka 		lastoff += BUSMEM_DATASIZE;
    674       1.27     pooka 		if (gendiff == 0)
    675       1.27     pooka 			devoff += BUSMEM_DATASIZE;
    676       1.27     pooka 	} else {
    677       1.27     pooka 		if (gendiff)
    678       1.27     pooka 			return false;
    679       1.27     pooka 	}
    680       1.27     pooka 
    681       1.27     pooka 	return devoff >= busmem->shm_first && devoff <= lastoff;
    682       1.27     pooka }
    683       1.27     pooka 
    684        1.1     pooka static void
    685        1.1     pooka shmif_rcv(void *arg)
    686        1.1     pooka {
    687        1.1     pooka 	struct ifnet *ifp = arg;
    688        1.1     pooka 	struct shmif_sc *sc = ifp->if_softc;
    689       1.32     pooka 	struct shmif_mem *busmem;
    690        1.1     pooka 	struct mbuf *m = NULL;
    691        1.1     pooka 	struct ether_header *eth;
    692       1.27     pooka 	uint32_t nextpkt;
    693       1.37     pooka 	bool wrap, passup;
    694        1.1     pooka 	int error;
    695       1.57     pooka 	const int align
    696       1.57     pooka 	    = ALIGN(sizeof(struct ether_header)) - sizeof(struct ether_header);
    697        1.1     pooka 
    698       1.32     pooka  reup:
    699       1.32     pooka 	mutex_enter(&sc->sc_mtx);
    700       1.32     pooka 	while ((ifp->if_flags & IFF_RUNNING) == 0 && !sc->sc_dying)
    701       1.32     pooka 		cv_wait(&sc->sc_cv, &sc->sc_mtx);
    702       1.32     pooka 	mutex_exit(&sc->sc_mtx);
    703       1.32     pooka 
    704       1.32     pooka 	busmem = sc->sc_busmem;
    705       1.32     pooka 
    706       1.32     pooka 	while (ifp->if_flags & IFF_RUNNING) {
    707       1.20     pooka 		struct shmif_pkthdr sp;
    708       1.20     pooka 
    709        1.1     pooka 		if (m == NULL) {
    710        1.1     pooka 			m = m_gethdr(M_WAIT, MT_DATA);
    711        1.1     pooka 			MCLGET(m, M_WAIT);
    712       1.57     pooka 			m->m_data += align;
    713        1.1     pooka 		}
    714        1.1     pooka 
    715       1.47     pooka 		DPRINTF(("waiting %d/%" PRIu64 "\n",
    716       1.47     pooka 		    sc->sc_nextpacket, sc->sc_devgen));
    717       1.26     pooka 		KASSERT(m->m_flags & M_EXT);
    718        1.1     pooka 
    719       1.26     pooka 		shmif_lockbus(busmem);
    720       1.26     pooka 		KASSERT(busmem->shm_magic == SHMIF_MAGIC);
    721       1.27     pooka 		KASSERT(busmem->shm_gen >= sc->sc_devgen);
    722        1.1     pooka 
    723        1.1     pooka 		/* need more data? */
    724       1.27     pooka 		if (sc->sc_devgen == busmem->shm_gen &&
    725       1.26     pooka 		    shmif_nextpktoff(busmem, busmem->shm_last)
    726       1.26     pooka 		     == sc->sc_nextpacket) {
    727       1.26     pooka 			shmif_unlockbus(busmem);
    728       1.71     ozaki 			error = rumpcomp_shmif_watchwait(sc->sc_kq);
    729        1.1     pooka 			if (__predict_false(error))
    730        1.1     pooka 				printf("shmif_rcv: wait failed %d\n", error);
    731       1.32     pooka 			membar_consumer();
    732        1.1     pooka 			continue;
    733        1.1     pooka 		}
    734        1.1     pooka 
    735       1.27     pooka 		if (stillvalid_p(sc)) {
    736       1.27     pooka 			nextpkt = sc->sc_nextpacket;
    737       1.27     pooka 		} else {
    738       1.27     pooka 			KASSERT(busmem->shm_gen > 0);
    739       1.26     pooka 			nextpkt = busmem->shm_first;
    740       1.26     pooka 			if (busmem->shm_first > busmem->shm_last)
    741       1.27     pooka 				sc->sc_devgen = busmem->shm_gen - 1;
    742       1.26     pooka 			else
    743       1.27     pooka 				sc->sc_devgen = busmem->shm_gen;
    744       1.47     pooka 			DPRINTF(("dev %p overrun, new data: %d/%" PRIu64 "\n",
    745       1.27     pooka 			    sc, nextpkt, sc->sc_devgen));
    746       1.26     pooka 		}
    747       1.26     pooka 
    748       1.26     pooka 		/*
    749       1.26     pooka 		 * If our read pointer is ahead the bus last write, our
    750       1.26     pooka 		 * generation must be one behind.
    751       1.26     pooka 		 */
    752       1.26     pooka 		KASSERT(!(nextpkt > busmem->shm_last
    753       1.27     pooka 		    && sc->sc_devgen == busmem->shm_gen));
    754       1.26     pooka 
    755       1.24     pooka 		wrap = false;
    756       1.26     pooka 		nextpkt = shmif_busread(busmem, &sp,
    757       1.26     pooka 		    nextpkt, sizeof(sp), &wrap);
    758       1.26     pooka 		KASSERT(sp.sp_len <= ETHERMTU + ETHER_HDR_LEN);
    759       1.26     pooka 		nextpkt = shmif_busread(busmem, mtod(m, void *),
    760       1.26     pooka 		    nextpkt, sp.sp_len, &wrap);
    761        1.1     pooka 
    762        1.1     pooka 		DPRINTF(("shmif_rcv: read packet of length %d at %d\n",
    763       1.20     pooka 		    sp.sp_len, nextpkt));
    764        1.1     pooka 
    765       1.26     pooka 		sc->sc_nextpacket = nextpkt;
    766       1.19     pooka 		shmif_unlockbus(sc->sc_busmem);
    767        1.1     pooka 
    768       1.27     pooka 		if (wrap) {
    769       1.26     pooka 			sc->sc_devgen++;
    770       1.47     pooka 			DPRINTF(("dev %p generation now %" PRIu64 "\n",
    771       1.27     pooka 			    sc, sc->sc_devgen));
    772       1.27     pooka 		}
    773       1.26     pooka 
    774       1.56     pooka 		/*
    775       1.56     pooka 		 * Ignore packets too short to possibly be valid.
    776       1.56     pooka 		 * This is hit at least for the first frame on a new bus.
    777       1.56     pooka 		 */
    778       1.55     pooka 		if (__predict_false(sp.sp_len < ETHER_HDR_LEN)) {
    779       1.55     pooka 			DPRINTF(("shmif read packet len %d < ETHER_HDR_LEN\n",
    780       1.55     pooka 			    sp.sp_len));
    781       1.55     pooka 			continue;
    782       1.55     pooka 		}
    783       1.55     pooka 
    784       1.20     pooka 		m->m_len = m->m_pkthdr.len = sp.sp_len;
    785       1.67     ozaki 		m_set_rcvif(m, ifp);
    786        1.1     pooka 
    787       1.37     pooka 		/*
    788       1.37     pooka 		 * Test if we want to pass the packet upwards
    789       1.37     pooka 		 */
    790        1.1     pooka 		eth = mtod(m, struct ether_header *);
    791       1.80     ozaki 		if (sp.sp_sender == sc->sc_uid) {
    792       1.63     ozaki 			passup = false;
    793       1.63     ozaki 		} else if (memcmp(eth->ether_dhost, CLLADDR(ifp->if_sadl),
    794       1.37     pooka 		    ETHER_ADDR_LEN) == 0) {
    795       1.37     pooka 			passup = true;
    796       1.46     pooka 		} else if (ETHER_IS_MULTICAST(eth->ether_dhost)) {
    797       1.37     pooka 			passup = true;
    798       1.37     pooka 		} else if (ifp->if_flags & IFF_PROMISC) {
    799       1.37     pooka 			m->m_flags |= M_PROMISC;
    800       1.37     pooka 			passup = true;
    801       1.38     pooka 		} else {
    802       1.38     pooka 			passup = false;
    803       1.37     pooka 		}
    804       1.37     pooka 
    805       1.37     pooka 		if (passup) {
    806       1.68     ozaki 			int bound;
    807       1.76       rin 
    808       1.76       rin 			m = ether_sw_offload_rx(ifp, m);
    809       1.76       rin 
    810       1.22     pooka 			KERNEL_LOCK(1, NULL);
    811       1.66     ozaki 			/* Prevent LWP migrations between CPUs for psref(9) */
    812       1.68     ozaki 			bound = curlwp_bind();
    813       1.65     ozaki 			if_input(ifp, m);
    814       1.68     ozaki 			curlwp_bindx(bound);
    815       1.22     pooka 			KERNEL_UNLOCK_ONE(NULL);
    816       1.76       rin 
    817        1.1     pooka 			m = NULL;
    818        1.1     pooka 		}
    819       1.37     pooka 		/* else: reuse mbuf for a future packet */
    820        1.1     pooka 	}
    821       1.32     pooka 	m_freem(m);
    822       1.32     pooka 	m = NULL;
    823       1.32     pooka 
    824       1.32     pooka 	if (!sc->sc_dying)
    825       1.32     pooka 		goto reup;
    826        1.1     pooka 
    827       1.32     pooka 	kthread_exit(0);
    828        1.1     pooka }
    829