Home | History | Annotate | Line # | Download | only in librumphijack
hijack.c revision 1.18
      1 /*      $NetBSD: hijack.c,v 1.18 2011/01/25 12:21:36 pooka Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2011 Antti Kantee.  All Rights Reserved.
      5  *
      6  * Redistribution and use in source and binary forms, with or without
      7  * modification, are permitted provided that the following conditions
      8  * are met:
      9  * 1. Redistributions of source code must retain the above copyright
     10  *    notice, this list of conditions and the following disclaimer.
     11  * 2. Redistributions in binary form must reproduce the above copyright
     12  *    notice, this list of conditions and the following disclaimer in the
     13  *    documentation and/or other materials provided with the distribution.
     14  *
     15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
     16  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     17  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
     18  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     21  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     25  * SUCH DAMAGE.
     26  */
     27 
     28 #include <sys/cdefs.h>
     29 __RCSID("$NetBSD: hijack.c,v 1.18 2011/01/25 12:21:36 pooka Exp $");
     30 
     31 #include <sys/param.h>
     32 #include <sys/types.h>
     33 #include <sys/event.h>
     34 #include <sys/ioctl.h>
     35 #include <sys/socket.h>
     36 #include <sys/poll.h>
     37 
     38 #include <rump/rumpclient.h>
     39 #include <rump/rump_syscalls.h>
     40 
     41 #include <assert.h>
     42 #include <dlfcn.h>
     43 #include <err.h>
     44 #include <errno.h>
     45 #include <fcntl.h>
     46 #include <poll.h>
     47 #include <pthread.h>
     48 #include <signal.h>
     49 #include <stdarg.h>
     50 #include <stdbool.h>
     51 #include <stdio.h>
     52 #include <stdlib.h>
     53 #include <time.h>
     54 #include <unistd.h>
     55 
     56 enum dualcall {
     57 	DUALCALL_WRITE, DUALCALL_WRITEV,
     58 	DUALCALL_IOCTL, DUALCALL_FCNTL,
     59 	DUALCALL_SOCKET, DUALCALL_ACCEPT, DUALCALL_BIND, DUALCALL_CONNECT,
     60 	DUALCALL_GETPEERNAME, DUALCALL_GETSOCKNAME, DUALCALL_LISTEN,
     61 	DUALCALL_RECVFROM, DUALCALL_RECVMSG,
     62 	DUALCALL_SENDTO, DUALCALL_SENDMSG,
     63 	DUALCALL_GETSOCKOPT, DUALCALL_SETSOCKOPT,
     64 	DUALCALL_SHUTDOWN,
     65 	DUALCALL_READ, DUALCALL_READV,
     66 	DUALCALL_DUP2, DUALCALL_CLOSE,
     67 	DUALCALL_POLLTS,
     68 	DUALCALL__NUM
     69 };
     70 
     71 #define RSYS_STRING(a) __STRING(a)
     72 #define RSYS_NAME(a) RSYS_STRING(__CONCAT(RUMP_SYS_RENAME_,a))
     73 
     74 /*
     75  * Would be nice to get this automatically in sync with libc.
     76  * Also, this does not work for compat-using binaries!
     77  */
     78 #if !__NetBSD_Prereq__(5,99,7)
     79 #define LIBCSELECT select
     80 #define LIBCPOLLTS pollts
     81 #define LIBCPOLL poll
     82 #else
     83 #define LIBCSELECT __select50
     84 #define LIBCPOLLTS __pollts50
     85 #define LIBCPOLL __poll50
     86 #endif
     87 
     88 int SELECT(int, fd_set *, fd_set *, fd_set *, struct timeval *);
     89 int POLLTS(struct pollfd *, nfds_t, const struct timespec *, const sigset_t *);
     90 int POLL(struct pollfd *, nfds_t, int);
     91 
     92 #define S(a) __STRING(a)
     93 struct sysnames {
     94 	enum dualcall scm_callnum;
     95 	const char *scm_hostname;
     96 	const char *scm_rumpname;
     97 } syscnames[] = {
     98 	{ DUALCALL_SOCKET,	"__socket30",	RSYS_NAME(SOCKET)	},
     99 	{ DUALCALL_ACCEPT,	"accept",	RSYS_NAME(ACCEPT)	},
    100 	{ DUALCALL_BIND,	"bind",		RSYS_NAME(BIND)		},
    101 	{ DUALCALL_CONNECT,	"connect",	RSYS_NAME(CONNECT)	},
    102 	{ DUALCALL_GETPEERNAME,	"getpeername",	RSYS_NAME(GETPEERNAME)	},
    103 	{ DUALCALL_GETSOCKNAME,	"getsockname",	RSYS_NAME(GETSOCKNAME)	},
    104 	{ DUALCALL_LISTEN,	"listen",	RSYS_NAME(LISTEN)	},
    105 	{ DUALCALL_RECVFROM,	"recvfrom",	RSYS_NAME(RECVFROM)	},
    106 	{ DUALCALL_RECVMSG,	"recvmsg",	RSYS_NAME(RECVMSG)	},
    107 	{ DUALCALL_SENDTO,	"sendto",	RSYS_NAME(SENDTO)	},
    108 	{ DUALCALL_SENDMSG,	"sendmsg",	RSYS_NAME(SENDMSG)	},
    109 	{ DUALCALL_GETSOCKOPT,	"getsockopt",	RSYS_NAME(GETSOCKOPT)	},
    110 	{ DUALCALL_SETSOCKOPT,	"setsockopt",	RSYS_NAME(SETSOCKOPT)	},
    111 	{ DUALCALL_SHUTDOWN,	"shutdown",	RSYS_NAME(SHUTDOWN)	},
    112 	{ DUALCALL_READ,	"read",		RSYS_NAME(READ)		},
    113 	{ DUALCALL_READV,	"readv",	RSYS_NAME(READV)	},
    114 	{ DUALCALL_WRITE,	"write",	RSYS_NAME(WRITE)	},
    115 	{ DUALCALL_WRITEV,	"writev",	RSYS_NAME(WRITEV)	},
    116 	{ DUALCALL_IOCTL,	"ioctl",	RSYS_NAME(IOCTL)	},
    117 	{ DUALCALL_FCNTL,	"fcntl",	RSYS_NAME(FCNTL)	},
    118 	{ DUALCALL_DUP2,	"dup2",		RSYS_NAME(DUP2)		},
    119 	{ DUALCALL_CLOSE,	"close",	RSYS_NAME(CLOSE)	},
    120 	{ DUALCALL_POLLTS,	S(LIBCPOLLTS),	RSYS_NAME(POLLTS)	},
    121 };
    122 #undef S
    123 
    124 struct bothsys {
    125 	void *bs_host;
    126 	void *bs_rump;
    127 } syscalls[DUALCALL__NUM];
    128 #define GETSYSCALL(which, name) syscalls[DUALCALL_##name].bs_##which
    129 
    130 pid_t (*host_fork)(void);
    131 
    132 static unsigned dup2mask;
    133 #define ISDUP2D(fd) (1<<(fd) & dup2mask)
    134 
    135 //#define DEBUGJACK
    136 #ifdef DEBUGJACK
    137 #define DPRINTF(x) mydprintf x
    138 static void
    139 mydprintf(const char *fmt, ...)
    140 {
    141 	va_list ap;
    142 
    143 	if (ISDUP2D(STDERR_FILENO))
    144 		return;
    145 
    146 	va_start(ap, fmt);
    147 	vfprintf(stderr, fmt, ap);
    148 	va_end(ap);
    149 }
    150 
    151 #else
    152 #define DPRINTF(x)
    153 #endif
    154 
    155 #define FDCALL(type, name, rcname, args, proto, vars)			\
    156 type name args								\
    157 {									\
    158 	type (*fun) proto;						\
    159 									\
    160 	if (fd_isrump(fd)) {						\
    161 		fun = syscalls[rcname].bs_rump;				\
    162 		fd = fd_host2rump(fd);					\
    163 	} else {							\
    164 		fun = syscalls[rcname].bs_host;				\
    165 	}								\
    166 									\
    167 	return fun vars;						\
    168 }
    169 
    170 /*
    171  * This is called from librumpclient in case of LD_PRELOAD.
    172  * It ensures correct RTLD_NEXT.
    173  */
    174 static void *
    175 hijackdlsym(void *handle, const char *symbol)
    176 {
    177 
    178 	return dlsym(handle, symbol);
    179 }
    180 
    181 /* low calorie sockets? */
    182 static bool hostlocalsockets = true;
    183 
    184 static void __attribute__((constructor))
    185 rcinit(void)
    186 {
    187 	int (*rumpcinit)(void);
    188 	void **rumpcdlsym;
    189 	void *hand;
    190 	int i, j;
    191 
    192 	hand = dlopen("librumpclient.so", RTLD_LAZY|RTLD_GLOBAL);
    193 	if (!hand)
    194 		err(1, "cannot open librumpclient.so");
    195 	rumpcinit = dlsym(hand, "rumpclient_init");
    196 	_DIAGASSERT(rumpcinit);
    197 
    198 	rumpcdlsym = dlsym(hand, "rumpclient_dlsym");
    199 	*rumpcdlsym = hijackdlsym;
    200 	host_fork = dlsym(RTLD_NEXT, "fork");
    201 
    202 	/*
    203 	 * In theory cannot print anything during lookups because
    204 	 * we might not have the call vector set up.  so, the errx()
    205 	 * is a bit of a strech, but it might work.
    206 	 */
    207 
    208 	for (i = 0; i < DUALCALL__NUM; i++) {
    209 		/* build runtime O(1) access */
    210 		for (j = 0; j < __arraycount(syscnames); j++) {
    211 			if (syscnames[j].scm_callnum == i)
    212 				break;
    213 		}
    214 
    215 		if (j == __arraycount(syscnames))
    216 			errx(1, "rumphijack error: syscall pos %d missing", i);
    217 
    218 		syscalls[i].bs_host = dlsym(hand,syscnames[j].scm_hostname);
    219 		if (syscalls[i].bs_host == NULL)
    220 			errx(1, "hostcall %s not found missing",
    221 			    syscnames[j].scm_hostname);
    222 
    223 		syscalls[i].bs_rump = dlsym(hand,syscnames[j].scm_rumpname);
    224 		if (syscalls[i].bs_rump == NULL)
    225 			errx(1, "rumpcall %s not found missing",
    226 			    syscnames[j].scm_rumpname);
    227 	}
    228 
    229 	if (rumpcinit() == -1)
    230 		err(1, "rumpclient init");
    231 }
    232 
    233 /* XXX: need runtime selection.  low for now due to FD_SETSIZE */
    234 #define HIJACK_FDOFF 128
    235 #define HIJACK_SELECT 128 /* XXX */
    236 #define HIJACK_ASSERT 128 /* XXX */
    237 static int
    238 fd_rump2host(int fd)
    239 {
    240 
    241 	if (fd == -1)
    242 		return fd;
    243 
    244 	if (!ISDUP2D(fd))
    245 		fd += HIJACK_FDOFF;
    246 
    247 	return fd;
    248 }
    249 
    250 static int
    251 fd_host2rump(int fd)
    252 {
    253 
    254 	if (!ISDUP2D(fd))
    255 		fd -= HIJACK_FDOFF;
    256 	return fd;
    257 }
    258 
    259 static bool
    260 fd_isrump(int fd)
    261 {
    262 
    263 	return ISDUP2D(fd) || fd >= HIJACK_FDOFF;
    264 }
    265 
    266 #define assertfd(_fd_) assert(ISDUP2D(_fd_) || (_fd_) >= HIJACK_ASSERT)
    267 #undef HIJACK_FDOFF
    268 
    269 int __socket30(int, int, int);
    270 int
    271 __socket30(int domain, int type, int protocol)
    272 {
    273 	int (*op_socket)(int, int, int);
    274 	int fd;
    275 	bool dohost;
    276 
    277 	dohost = hostlocalsockets && (domain == AF_LOCAL);
    278 
    279 	if (dohost)
    280 		op_socket = GETSYSCALL(host, SOCKET);
    281 	else
    282 		op_socket = GETSYSCALL(rump, SOCKET);
    283 	fd = op_socket(domain, type, protocol);
    284 
    285 	if (!dohost)
    286 		fd = fd_rump2host(fd);
    287 	DPRINTF(("socket <- %d\n", fd));
    288 
    289 	return fd;
    290 }
    291 
    292 int
    293 accept(int s, struct sockaddr *addr, socklen_t *addrlen)
    294 {
    295 	int (*op_accept)(int, struct sockaddr *, socklen_t *);
    296 	int fd;
    297 	bool isrump;
    298 
    299 	isrump = fd_isrump(s);
    300 
    301 	DPRINTF(("accept -> %d", s));
    302 	if (isrump) {
    303 		op_accept = GETSYSCALL(rump, ACCEPT);
    304 		s = fd_host2rump(s);
    305 	} else {
    306 		op_accept = GETSYSCALL(host, ACCEPT);
    307 	}
    308 	fd = op_accept(s, addr, addrlen);
    309 	if (fd != -1 && isrump)
    310 		fd = fd_rump2host(fd);
    311 
    312 	DPRINTF((" <- %d\n", fd));
    313 
    314 	return fd;
    315 }
    316 
    317 /*
    318  * ioctl and fcntl are varargs calls and need special treatment
    319  */
    320 int
    321 ioctl(int fd, unsigned long cmd, ...)
    322 {
    323 	int (*op_ioctl)(int, unsigned long cmd, ...);
    324 	va_list ap;
    325 	int rv;
    326 
    327 	DPRINTF(("ioctl -> %d\n", fd));
    328 	if (fd_isrump(fd)) {
    329 		fd = fd_host2rump(fd);
    330 		op_ioctl = GETSYSCALL(rump, IOCTL);
    331 	} else {
    332 		op_ioctl = GETSYSCALL(host, IOCTL);
    333 	}
    334 
    335 	va_start(ap, cmd);
    336 	rv = op_ioctl(fd, cmd, va_arg(ap, void *));
    337 	va_end(ap);
    338 	return rv;
    339 }
    340 
    341 int
    342 fcntl(int fd, int cmd, ...)
    343 {
    344 	int (*op_fcntl)(int, int, ...);
    345 	va_list ap;
    346 	int rv;
    347 
    348 	DPRINTF(("fcntl -> %d\n", fd));
    349 	if (fd_isrump(fd)) {
    350 		fd = fd_host2rump(fd);
    351 		op_fcntl = GETSYSCALL(rump, FCNTL);
    352 	} else {
    353 		op_fcntl = GETSYSCALL(host, FCNTL);
    354 	}
    355 
    356 	va_start(ap, cmd);
    357 	rv = op_fcntl(fd, cmd, va_arg(ap, void *));
    358 	va_end(ap);
    359 	return rv;
    360 }
    361 
    362 /*
    363  * write cannot issue a standard debug printf due to recursion
    364  */
    365 ssize_t
    366 write(int fd, const void *buf, size_t blen)
    367 {
    368 	ssize_t (*op_write)(int, const void *, size_t);
    369 
    370 	if (fd_isrump(fd)) {
    371 		fd = fd_host2rump(fd);
    372 		op_write = GETSYSCALL(rump, WRITE);
    373 	} else {
    374 		op_write = GETSYSCALL(host, WRITE);
    375 	}
    376 
    377 	return op_write(fd, buf, blen);
    378 }
    379 
    380 /*
    381  * dup2 is special.  we allow dup2 of a rump kernel fd to 0-2 since
    382  * many programs do that.  dup2 of a rump kernel fd to another value
    383  * not >= fdoff is an error.
    384  *
    385  * Note: cannot rump2host newd, because it is often hardcoded.
    386  */
    387 int
    388 dup2(int oldd, int newd)
    389 {
    390 	int (*host_dup2)(int, int);
    391 	int rv;
    392 
    393 	DPRINTF(("dup2 -> %d (o) -> %d (n)\n", oldd, newd));
    394 
    395 	if (fd_isrump(oldd)) {
    396 		if (!(newd >= 0 && newd <= 2))
    397 			return EBADF;
    398 		oldd = fd_host2rump(oldd);
    399 		rv = rump_sys_dup2(oldd, newd);
    400 		if (rv != -1)
    401 			dup2mask |= 1<<newd;
    402 	} else {
    403 		host_dup2 = syscalls[DUALCALL_DUP2].bs_host;
    404 		rv = host_dup2(oldd, newd);
    405 	}
    406 
    407 	return rv;
    408 }
    409 
    410 /*
    411  * We just wrap fork the appropriate rump client calls to preserve
    412  * the file descriptors of the forked parent in the child, but
    413  * prevent double use of connection fd.
    414  */
    415 pid_t
    416 fork()
    417 {
    418 	struct rumpclient_fork *rf;
    419 	pid_t rv;
    420 
    421 	DPRINTF(("fork\n"));
    422 
    423 	if ((rf = rumpclient_prefork()) == NULL)
    424 		return -1;
    425 
    426 	switch ((rv = host_fork())) {
    427 	case -1:
    428 		/* XXX: cancel rf */
    429 		break;
    430 	case 0:
    431 		if (rumpclient_fork_init(rf) == -1)
    432 			rv = -1;
    433 		break;
    434 	default:
    435 		break;
    436 	}
    437 
    438 	DPRINTF(("fork returns %d\n", rv));
    439 	return rv;
    440 }
    441 
    442 /*
    443  * select is done by calling poll.
    444  */
    445 int
    446 LIBCSELECT(int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
    447 	struct timeval *timeout)
    448 {
    449 	struct pollfd *pfds;
    450 	struct timespec ts, *tsp = NULL;
    451 	nfds_t i, j, realnfds;
    452 	int rv, incr;
    453 
    454 	DPRINTF(("select\n"));
    455 
    456 	/*
    457 	 * Well, first we must scan the fds to figure out how many
    458 	 * fds there really are.  This is because up to and including
    459 	 * nb5 poll() silently refuses nfds > process_maxopen_fds.
    460 	 * Seems to be fixed in current, thank the maker.
    461 	 * god damn cluster...bomb.
    462 	 */
    463 
    464 	for (i = 0, realnfds = 0; i < nfds; i++) {
    465 		if (readfds && FD_ISSET(i, readfds)) {
    466 			realnfds++;
    467 			continue;
    468 		}
    469 		if (writefds && FD_ISSET(i, writefds)) {
    470 			realnfds++;
    471 			continue;
    472 		}
    473 		if (exceptfds && FD_ISSET(i, exceptfds)) {
    474 			realnfds++;
    475 			continue;
    476 		}
    477 	}
    478 
    479 	if (realnfds) {
    480 		pfds = malloc(sizeof(*pfds) * realnfds);
    481 		if (!pfds)
    482 			return -1;
    483 	} else {
    484 		pfds = NULL;
    485 	}
    486 
    487 	for (i = 0, j = 0; i < nfds; i++) {
    488 		incr = 0;
    489 		pfds[j].events = pfds[j].revents = 0;
    490 		if (readfds && FD_ISSET(i, readfds)) {
    491 			pfds[j].fd = i;
    492 			pfds[j].events |= POLLIN;
    493 			incr=1;
    494 		}
    495 		if (writefds && FD_ISSET(i, writefds)) {
    496 			pfds[j].fd = i;
    497 			pfds[j].events |= POLLOUT;
    498 			incr=1;
    499 		}
    500 		if (exceptfds && FD_ISSET(i, exceptfds)) {
    501 			pfds[j].fd = i;
    502 			pfds[j].events |= POLLHUP|POLLERR;
    503 			incr=1;
    504 		}
    505 		if (incr)
    506 			j++;
    507 	}
    508 
    509 	if (timeout) {
    510 		TIMEVAL_TO_TIMESPEC(timeout, &ts);
    511 		tsp = &ts;
    512 	}
    513 	rv = pollts(pfds, realnfds, tsp, NULL);
    514 	if (rv <= 0)
    515 		goto out;
    516 
    517 	/*
    518 	 * ok, harvest results.  first zero out entries (can't use
    519 	 * FD_ZERO for the obvious select-me-not reason).  whee.
    520 	 */
    521 	for (i = 0; i < nfds; i++) {
    522 		if (readfds)
    523 			FD_CLR(i, readfds);
    524 		if (writefds)
    525 			FD_CLR(i, writefds);
    526 		if (exceptfds)
    527 			FD_CLR(i, exceptfds);
    528 	}
    529 
    530 	/* and then plug in the results */
    531 	for (i = 0; i < realnfds; i++) {
    532 		if (readfds) {
    533 			if (pfds[i].revents & POLLIN) {
    534 				FD_SET(pfds[i].fd, readfds);
    535 			}
    536 		}
    537 		if (writefds) {
    538 			if (pfds[i].revents & POLLOUT) {
    539 				FD_SET(pfds[i].fd, writefds);
    540 			}
    541 		}
    542 		if (exceptfds) {
    543 			if (pfds[i].revents & (POLLHUP|POLLERR)) {
    544 				FD_SET(pfds[i].fd, exceptfds);
    545 			}
    546 		}
    547 	}
    548 
    549  out:
    550 	free(pfds);
    551 	return rv;
    552 }
    553 
    554 static void
    555 checkpoll(struct pollfd *fds, nfds_t nfds, int *hostcall, int *rumpcall)
    556 {
    557 	nfds_t i;
    558 
    559 	for (i = 0; i < nfds; i++) {
    560 		if (fds[i].fd == -1)
    561 			continue;
    562 
    563 		if (fd_isrump(fds[i].fd))
    564 			(*rumpcall)++;
    565 		else
    566 			(*hostcall)++;
    567 	}
    568 }
    569 
    570 static void
    571 adjustpoll(struct pollfd *fds, nfds_t nfds, int (*fdadj)(int))
    572 {
    573 	nfds_t i;
    574 
    575 	for (i = 0; i < nfds; i++) {
    576 		fds[i].fd = fdadj(fds[i].fd);
    577 	}
    578 }
    579 
    580 /*
    581  * poll is easy as long as the call comes in the fds only in one
    582  * kernel.  otherwise its quite tricky...
    583  */
    584 struct pollarg {
    585 	struct pollfd *pfds;
    586 	nfds_t nfds;
    587 	const struct timespec *ts;
    588 	const sigset_t *sigmask;
    589 	int pipefd;
    590 	int errnum;
    591 };
    592 
    593 static void *
    594 hostpoll(void *arg)
    595 {
    596 	int (*op_pollts)(struct pollfd *, nfds_t, const struct timespec *,
    597 			 const sigset_t *);
    598 	struct pollarg *parg = arg;
    599 	intptr_t rv;
    600 
    601 	op_pollts = syscalls[DUALCALL_POLLTS].bs_host;
    602 	rv = op_pollts(parg->pfds, parg->nfds, parg->ts, parg->sigmask);
    603 	if (rv == -1)
    604 		parg->errnum = errno;
    605 	rump_sys_write(parg->pipefd, &rv, sizeof(rv));
    606 
    607 	return (void *)(intptr_t)rv;
    608 }
    609 
    610 int
    611 LIBCPOLLTS(struct pollfd *fds, nfds_t nfds, const struct timespec *ts,
    612 	const sigset_t *sigmask)
    613 {
    614 	int (*op_pollts)(struct pollfd *, nfds_t, const struct timespec *,
    615 			 const sigset_t *);
    616 	int (*host_close)(int);
    617 	int hostcall = 0, rumpcall = 0;
    618 	pthread_t pt;
    619 	nfds_t i;
    620 	int rv;
    621 
    622 	DPRINTF(("poll\n"));
    623 	checkpoll(fds, nfds, &hostcall, &rumpcall);
    624 
    625 	if (hostcall && rumpcall) {
    626 		struct pollfd *pfd_host = NULL, *pfd_rump = NULL;
    627 		int rpipe[2] = {-1,-1}, hpipe[2] = {-1,-1};
    628 		struct pollarg parg;
    629 		uintptr_t lrv;
    630 		int sverrno = 0, trv;
    631 
    632 		/*
    633 		 * ok, this is where it gets tricky.  We must support
    634 		 * this since it's a very common operation in certain
    635 		 * types of software (telnet, netcat, etc).  We allocate
    636 		 * two vectors and run two poll commands in separate
    637 		 * threads.  Whichever returns first "wins" and the
    638 		 * other kernel's fds won't show activity.
    639 		 */
    640 		rv = -1;
    641 
    642 		/* allocate full vector for O(n) joining after call */
    643 		pfd_host = malloc(sizeof(*pfd_host)*(nfds+1));
    644 		if (!pfd_host)
    645 			goto out;
    646 		pfd_rump = malloc(sizeof(*pfd_rump)*(nfds+1));
    647 		if (!pfd_rump) {
    648 			goto out;
    649 		}
    650 
    651 		/* split vectors */
    652 		for (i = 0; i < nfds; i++) {
    653 			if (fds[i].fd == -1) {
    654 				pfd_host[i].fd = -1;
    655 				pfd_rump[i].fd = -1;
    656 			} else if (fd_isrump(fds[i].fd)) {
    657 				pfd_host[i].fd = -1;
    658 				pfd_rump[i].fd = fd_host2rump(fds[i].fd);
    659 				pfd_rump[i].events = fds[i].events;
    660 			} else {
    661 				pfd_rump[i].fd = -1;
    662 				pfd_host[i].fd = fds[i].fd;
    663 				pfd_host[i].events = fds[i].events;
    664 			}
    665 			fds[i].revents = 0;
    666 		}
    667 
    668 		/*
    669 		 * then, open two pipes, one for notifications
    670 		 * to each kernel.
    671 		 */
    672 		if (rump_sys_pipe(rpipe) == -1)
    673 			goto out;
    674 		if (pipe(hpipe) == -1)
    675 			goto out;
    676 
    677 		pfd_host[nfds].fd = hpipe[0];
    678 		pfd_host[nfds].events = POLLIN;
    679 		pfd_rump[nfds].fd = rpipe[0];
    680 		pfd_rump[nfds].events = POLLIN;
    681 
    682 		/*
    683 		 * then, create a thread to do host part and meanwhile
    684 		 * do rump kernel part right here
    685 		 */
    686 
    687 		parg.pfds = pfd_host;
    688 		parg.nfds = nfds+1;
    689 		parg.ts = ts;
    690 		parg.sigmask = sigmask;
    691 		parg.pipefd = rpipe[1];
    692 		pthread_create(&pt, NULL, hostpoll, &parg);
    693 
    694 		op_pollts = syscalls[DUALCALL_POLLTS].bs_rump;
    695 		lrv = op_pollts(pfd_rump, nfds+1, ts, NULL);
    696 		sverrno = errno;
    697 		write(hpipe[1], &rv, sizeof(rv));
    698 		pthread_join(pt, (void *)&trv);
    699 
    700 		/* check who "won" and merge results */
    701 		if (lrv != 0 && pfd_host[nfds].revents & POLLIN) {
    702 			rv = trv;
    703 
    704 			for (i = 0; i < nfds; i++) {
    705 				if (pfd_rump[i].fd != -1)
    706 					fds[i].revents = pfd_rump[i].revents;
    707 			}
    708 			sverrno = parg.errnum;
    709 		} else if (trv != 0 && pfd_rump[nfds].revents & POLLIN) {
    710 			rv = trv;
    711 
    712 			for (i = 0; i < nfds; i++) {
    713 				if (pfd_host[i].fd != -1)
    714 					fds[i].revents = pfd_host[i].revents;
    715 			}
    716 		} else {
    717 			rv = 0;
    718 		}
    719 
    720  out:
    721 		host_close = syscalls[DUALCALL_CLOSE].bs_host;
    722 		if (rpipe[0] != -1)
    723 			rump_sys_close(rpipe[0]);
    724 		if (rpipe[1] != -1)
    725 			rump_sys_close(rpipe[1]);
    726 		if (hpipe[0] != -1)
    727 			host_close(hpipe[0]);
    728 		if (hpipe[1] != -1)
    729 			host_close(hpipe[1]);
    730 		free(pfd_host);
    731 		free(pfd_rump);
    732 		errno = sverrno;
    733 	} else {
    734 		if (hostcall) {
    735 			op_pollts = syscalls[DUALCALL_POLLTS].bs_host;
    736 		} else {
    737 			op_pollts = syscalls[DUALCALL_POLLTS].bs_rump;
    738 			adjustpoll(fds, nfds, fd_host2rump);
    739 		}
    740 
    741 		rv = op_pollts(fds, nfds, ts, sigmask);
    742 		if (rumpcall)
    743 			adjustpoll(fds, nfds, fd_rump2host);
    744 	}
    745 
    746 	return rv;
    747 }
    748 
    749 int
    750 LIBCPOLL(struct pollfd *fds, nfds_t nfds, int timeout)
    751 {
    752 	struct timespec ts;
    753 	struct timespec *tsp = NULL;
    754 
    755 	if (timeout != INFTIM) {
    756 		ts.tv_sec = timeout / 1000;
    757 		ts.tv_nsec = (timeout % 1000) * 1000*1000;
    758 
    759 		tsp = &ts;
    760 	}
    761 
    762 	return pollts(fds, nfds, tsp, NULL);
    763 }
    764 
    765 int
    766 kqueue(void)
    767 {
    768 
    769 	fprintf(stderr, "kqueue unsupported");
    770 	abort();
    771 	/*NOTREACHED*/
    772 }
    773 
    774 /*ARGSUSED*/
    775 int
    776 kevent(int kq, const struct kevent *changelist, size_t nchanges,
    777 	struct kevent *eventlist, size_t nevents,
    778 	const struct timespec *timeout)
    779 {
    780 
    781 	fprintf(stderr, "kqueue unsupported");
    782 	abort();
    783 	/*NOTREACHED*/
    784 }
    785 
    786 /*
    787  * Rest are std type calls.
    788  */
    789 
    790 FDCALL(int, bind, DUALCALL_BIND,					\
    791 	(int fd, const struct sockaddr *name, socklen_t namelen),	\
    792 	(int, const struct sockaddr *, socklen_t),			\
    793 	(fd, name, namelen))
    794 
    795 FDCALL(int, connect, DUALCALL_CONNECT,					\
    796 	(int fd, const struct sockaddr *name, socklen_t namelen),	\
    797 	(int, const struct sockaddr *, socklen_t),			\
    798 	(fd, name, namelen))
    799 
    800 FDCALL(int, getpeername, DUALCALL_GETPEERNAME,				\
    801 	(int fd, struct sockaddr *name, socklen_t *namelen),		\
    802 	(int, struct sockaddr *, socklen_t *),				\
    803 	(fd, name, namelen))
    804 
    805 FDCALL(int, getsockname, DUALCALL_GETSOCKNAME, 				\
    806 	(int fd, struct sockaddr *name, socklen_t *namelen),		\
    807 	(int, struct sockaddr *, socklen_t *),				\
    808 	(fd, name, namelen))
    809 
    810 FDCALL(int, listen, DUALCALL_LISTEN,	 				\
    811 	(int fd, int backlog),						\
    812 	(int, int),							\
    813 	(fd, backlog))
    814 
    815 FDCALL(ssize_t, recvfrom, DUALCALL_RECVFROM, 				\
    816 	(int fd, void *buf, size_t len, int flags,			\
    817 	    struct sockaddr *from, socklen_t *fromlen),			\
    818 	(int, void *, size_t, int, struct sockaddr *, socklen_t *),	\
    819 	(fd, buf, len, flags, from, fromlen))
    820 
    821 FDCALL(ssize_t, sendto, DUALCALL_SENDTO, 				\
    822 	(int fd, const void *buf, size_t len, int flags,		\
    823 	    const struct sockaddr *to, socklen_t tolen),		\
    824 	(int, const void *, size_t, int,				\
    825 	    const struct sockaddr *, socklen_t),			\
    826 	(fd, buf, len, flags, to, tolen))
    827 
    828 FDCALL(ssize_t, recvmsg, DUALCALL_RECVMSG, 				\
    829 	(int fd, struct msghdr *msg, int flags),			\
    830 	(int, struct msghdr *, int),					\
    831 	(fd, msg, flags))
    832 
    833 FDCALL(ssize_t, sendmsg, DUALCALL_SENDMSG, 				\
    834 	(int fd, const struct msghdr *msg, int flags),			\
    835 	(int, const struct msghdr *, int),				\
    836 	(fd, msg, flags))
    837 
    838 FDCALL(int, getsockopt, DUALCALL_GETSOCKOPT, 				\
    839 	(int fd, int level, int optn, void *optval, socklen_t *optlen),	\
    840 	(int, int, int, void *, socklen_t *),				\
    841 	(fd, level, optn, optval, optlen))
    842 
    843 FDCALL(int, setsockopt, DUALCALL_SETSOCKOPT, 				\
    844 	(int fd, int level, int optn,					\
    845 	    const void *optval, socklen_t optlen),			\
    846 	(int, int, int, const void *, socklen_t),			\
    847 	(fd, level, optn, optval, optlen))
    848 
    849 FDCALL(int, shutdown, DUALCALL_SHUTDOWN, 				\
    850 	(int fd, int how),						\
    851 	(int, int),							\
    852 	(fd, how))
    853 
    854 FDCALL(ssize_t, read, DUALCALL_READ,	 				\
    855 	(int fd, void *buf, size_t buflen),				\
    856 	(int, void *, size_t),						\
    857 	(fd, buf, buflen))
    858 
    859 FDCALL(ssize_t, readv, DUALCALL_READV, 					\
    860 	(int fd, const struct iovec *iov, int iovcnt),			\
    861 	(int, const struct iovec *, int),				\
    862 	(fd, iov, iovcnt))
    863 
    864 FDCALL(ssize_t, writev, DUALCALL_WRITEV, 				\
    865 	(int fd, const struct iovec *iov, int iovcnt),			\
    866 	(int, const struct iovec *, int),				\
    867 	(fd, iov, iovcnt))
    868 
    869 FDCALL(int, close, DUALCALL_CLOSE,	 				\
    870 	(int fd),							\
    871 	(int),								\
    872 	(fd))
    873