Home | History | Annotate | Line # | Download | only in librumphijack
hijack.c revision 1.22
      1 /*      $NetBSD: hijack.c,v 1.22 2011/01/27 18:05:16 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 1.22 2011/01/27 18:05:16 pooka Exp $");
     30 
     31 #define __ssp_weak_name(fun) _hijack_ ## fun
     32 
     33 #include <sys/param.h>
     34 #include <sys/types.h>
     35 #include <sys/event.h>
     36 #include <sys/ioctl.h>
     37 #include <sys/socket.h>
     38 #include <sys/poll.h>
     39 
     40 #include <rump/rumpclient.h>
     41 #include <rump/rump_syscalls.h>
     42 
     43 #include <assert.h>
     44 #include <dlfcn.h>
     45 #include <err.h>
     46 #include <errno.h>
     47 #include <fcntl.h>
     48 #include <poll.h>
     49 #include <pthread.h>
     50 #include <signal.h>
     51 #include <stdarg.h>
     52 #include <stdbool.h>
     53 #include <stdio.h>
     54 #include <stdlib.h>
     55 #include <time.h>
     56 #include <unistd.h>
     57 
     58 enum dualcall {
     59 	DUALCALL_WRITE, DUALCALL_WRITEV,
     60 	DUALCALL_IOCTL, DUALCALL_FCNTL,
     61 	DUALCALL_SOCKET, DUALCALL_ACCEPT, DUALCALL_BIND, DUALCALL_CONNECT,
     62 	DUALCALL_GETPEERNAME, DUALCALL_GETSOCKNAME, DUALCALL_LISTEN,
     63 	DUALCALL_RECVFROM, DUALCALL_RECVMSG,
     64 	DUALCALL_SENDTO, DUALCALL_SENDMSG,
     65 	DUALCALL_GETSOCKOPT, DUALCALL_SETSOCKOPT,
     66 	DUALCALL_SHUTDOWN,
     67 	DUALCALL_READ, DUALCALL_READV,
     68 	DUALCALL_DUP2, DUALCALL_CLOSE,
     69 	DUALCALL_POLLTS,
     70 	DUALCALL__NUM
     71 };
     72 
     73 #define RSYS_STRING(a) __STRING(a)
     74 #define RSYS_NAME(a) RSYS_STRING(__CONCAT(RUMP_SYS_RENAME_,a))
     75 
     76 /*
     77  * Would be nice to get this automatically in sync with libc.
     78  * Also, this does not work for compat-using binaries!
     79  */
     80 #if !__NetBSD_Prereq__(5,99,7)
     81 #define LIBCSELECT select
     82 #define LIBCPOLLTS pollts
     83 #define LIBCPOLL poll
     84 #else
     85 #define LIBCSELECT __select50
     86 #define LIBCPOLLTS __pollts50
     87 #define LIBCPOLL __poll50
     88 #endif
     89 
     90 int LIBCSELECT(int, fd_set *, fd_set *, fd_set *, struct timeval *);
     91 int LIBCPOLLTS(struct pollfd *, nfds_t,
     92 	       const struct timespec *, const sigset_t *);
     93 int LIBCPOLL(struct pollfd *, nfds_t, int);
     94 
     95 #define S(a) __STRING(a)
     96 struct sysnames {
     97 	enum dualcall scm_callnum;
     98 	const char *scm_hostname;
     99 	const char *scm_rumpname;
    100 } syscnames[] = {
    101 	{ DUALCALL_SOCKET,	"__socket30",	RSYS_NAME(SOCKET)	},
    102 	{ DUALCALL_ACCEPT,	"accept",	RSYS_NAME(ACCEPT)	},
    103 	{ DUALCALL_BIND,	"bind",		RSYS_NAME(BIND)		},
    104 	{ DUALCALL_CONNECT,	"connect",	RSYS_NAME(CONNECT)	},
    105 	{ DUALCALL_GETPEERNAME,	"getpeername",	RSYS_NAME(GETPEERNAME)	},
    106 	{ DUALCALL_GETSOCKNAME,	"getsockname",	RSYS_NAME(GETSOCKNAME)	},
    107 	{ DUALCALL_LISTEN,	"listen",	RSYS_NAME(LISTEN)	},
    108 	{ DUALCALL_RECVFROM,	"recvfrom",	RSYS_NAME(RECVFROM)	},
    109 	{ DUALCALL_RECVMSG,	"recvmsg",	RSYS_NAME(RECVMSG)	},
    110 	{ DUALCALL_SENDTO,	"sendto",	RSYS_NAME(SENDTO)	},
    111 	{ DUALCALL_SENDMSG,	"sendmsg",	RSYS_NAME(SENDMSG)	},
    112 	{ DUALCALL_GETSOCKOPT,	"getsockopt",	RSYS_NAME(GETSOCKOPT)	},
    113 	{ DUALCALL_SETSOCKOPT,	"setsockopt",	RSYS_NAME(SETSOCKOPT)	},
    114 	{ DUALCALL_SHUTDOWN,	"shutdown",	RSYS_NAME(SHUTDOWN)	},
    115 	{ DUALCALL_READ,	"read",		RSYS_NAME(READ)		},
    116 	{ DUALCALL_READV,	"readv",	RSYS_NAME(READV)	},
    117 	{ DUALCALL_WRITE,	"write",	RSYS_NAME(WRITE)	},
    118 	{ DUALCALL_WRITEV,	"writev",	RSYS_NAME(WRITEV)	},
    119 	{ DUALCALL_IOCTL,	"ioctl",	RSYS_NAME(IOCTL)	},
    120 	{ DUALCALL_FCNTL,	"fcntl",	RSYS_NAME(FCNTL)	},
    121 	{ DUALCALL_DUP2,	"dup2",		RSYS_NAME(DUP2)		},
    122 	{ DUALCALL_CLOSE,	"close",	RSYS_NAME(CLOSE)	},
    123 	{ DUALCALL_POLLTS,	S(LIBCPOLLTS),	RSYS_NAME(POLLTS)	},
    124 };
    125 #undef S
    126 
    127 struct bothsys {
    128 	void *bs_host;
    129 	void *bs_rump;
    130 } syscalls[DUALCALL__NUM];
    131 #define GETSYSCALL(which, name) syscalls[DUALCALL_##name].bs_##which
    132 
    133 pid_t (*host_fork)(void);
    134 
    135 static unsigned dup2mask;
    136 #define ISDUP2D(fd) (1<<(fd) & dup2mask)
    137 
    138 //#define DEBUGJACK
    139 #ifdef DEBUGJACK
    140 #define DPRINTF(x) mydprintf x
    141 static void
    142 mydprintf(const char *fmt, ...)
    143 {
    144 	va_list ap;
    145 
    146 	if (ISDUP2D(STDERR_FILENO))
    147 		return;
    148 
    149 	va_start(ap, fmt);
    150 	vfprintf(stderr, fmt, ap);
    151 	va_end(ap);
    152 }
    153 
    154 #else
    155 #define DPRINTF(x)
    156 #endif
    157 
    158 #define FDCALL(type, name, rcname, args, proto, vars)			\
    159 type name args								\
    160 {									\
    161 	type (*fun) proto;						\
    162 									\
    163 	if (fd_isrump(fd)) {						\
    164 		fun = syscalls[rcname].bs_rump;				\
    165 		fd = fd_host2rump(fd);					\
    166 	} else {							\
    167 		fun = syscalls[rcname].bs_host;				\
    168 	}								\
    169 									\
    170 	return fun vars;						\
    171 }
    172 
    173 /*
    174  * This is called from librumpclient in case of LD_PRELOAD.
    175  * It ensures correct RTLD_NEXT.
    176  */
    177 static void *
    178 hijackdlsym(void *handle, const char *symbol)
    179 {
    180 
    181 	return dlsym(handle, symbol);
    182 }
    183 
    184 /* low calorie sockets? */
    185 static bool hostlocalsockets = true;
    186 
    187 static void __attribute__((constructor))
    188 rcinit(void)
    189 {
    190 	void **rumpcdlsym;
    191 	void *hand;
    192 	unsigned i, j;
    193 
    194 	hand = dlopen("librumpclient.so", RTLD_LAZY|RTLD_GLOBAL);
    195 	if (!hand)
    196 		err(1, "cannot open librumpclient.so");
    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 (rumpclient_init() == -1)
    230 		err(1, "rumpclient init");
    231 	rumpclient_setconnretry(RUMPCLIENT_RETRYCONN_INFTIME);
    232 }
    233 
    234 /* XXX: need runtime selection.  low for now due to FD_SETSIZE */
    235 #define HIJACK_FDOFF 128
    236 #define HIJACK_SELECT 128 /* XXX */
    237 #define HIJACK_ASSERT 128 /* XXX */
    238 static int
    239 fd_rump2host(int fd)
    240 {
    241 
    242 	if (fd == -1)
    243 		return fd;
    244 
    245 	if (!ISDUP2D(fd))
    246 		fd += HIJACK_FDOFF;
    247 
    248 	return fd;
    249 }
    250 
    251 static int
    252 fd_host2rump(int fd)
    253 {
    254 
    255 	if (!ISDUP2D(fd))
    256 		fd -= HIJACK_FDOFF;
    257 	return fd;
    258 }
    259 
    260 static bool
    261 fd_isrump(int fd)
    262 {
    263 
    264 	return ISDUP2D(fd) || fd >= HIJACK_FDOFF;
    265 }
    266 
    267 #define assertfd(_fd_) assert(ISDUP2D(_fd_) || (_fd_) >= HIJACK_ASSERT)
    268 #undef HIJACK_FDOFF
    269 
    270 int __socket30(int, int, int);
    271 int
    272 __socket30(int domain, int type, int protocol)
    273 {
    274 	int (*op_socket)(int, int, int);
    275 	int fd;
    276 	bool dohost;
    277 
    278 	dohost = hostlocalsockets && (domain == AF_LOCAL);
    279 
    280 	if (dohost)
    281 		op_socket = GETSYSCALL(host, SOCKET);
    282 	else
    283 		op_socket = GETSYSCALL(rump, SOCKET);
    284 	fd = op_socket(domain, type, protocol);
    285 
    286 	if (!dohost)
    287 		fd = fd_rump2host(fd);
    288 	DPRINTF(("socket <- %d\n", fd));
    289 
    290 	return fd;
    291 }
    292 
    293 int
    294 accept(int s, struct sockaddr *addr, socklen_t *addrlen)
    295 {
    296 	int (*op_accept)(int, struct sockaddr *, socklen_t *);
    297 	int fd;
    298 	bool isrump;
    299 
    300 	isrump = fd_isrump(s);
    301 
    302 	DPRINTF(("accept -> %d", s));
    303 	if (isrump) {
    304 		op_accept = GETSYSCALL(rump, ACCEPT);
    305 		s = fd_host2rump(s);
    306 	} else {
    307 		op_accept = GETSYSCALL(host, ACCEPT);
    308 	}
    309 	fd = op_accept(s, addr, addrlen);
    310 	if (fd != -1 && isrump)
    311 		fd = fd_rump2host(fd);
    312 
    313 	DPRINTF((" <- %d\n", fd));
    314 
    315 	return fd;
    316 }
    317 
    318 /*
    319  * ioctl and fcntl are varargs calls and need special treatment
    320  */
    321 int
    322 ioctl(int fd, unsigned long cmd, ...)
    323 {
    324 	int (*op_ioctl)(int, unsigned long cmd, ...);
    325 	va_list ap;
    326 	int rv;
    327 
    328 	DPRINTF(("ioctl -> %d\n", fd));
    329 	if (fd_isrump(fd)) {
    330 		fd = fd_host2rump(fd);
    331 		op_ioctl = GETSYSCALL(rump, IOCTL);
    332 	} else {
    333 		op_ioctl = GETSYSCALL(host, IOCTL);
    334 	}
    335 
    336 	va_start(ap, cmd);
    337 	rv = op_ioctl(fd, cmd, va_arg(ap, void *));
    338 	va_end(ap);
    339 	return rv;
    340 }
    341 
    342 int
    343 fcntl(int fd, int cmd, ...)
    344 {
    345 	int (*op_fcntl)(int, int, ...);
    346 	va_list ap;
    347 	int rv;
    348 
    349 	DPRINTF(("fcntl -> %d\n", fd));
    350 	if (fd_isrump(fd)) {
    351 		fd = fd_host2rump(fd);
    352 		op_fcntl = GETSYSCALL(rump, FCNTL);
    353 	} else {
    354 		op_fcntl = GETSYSCALL(host, FCNTL);
    355 	}
    356 
    357 	va_start(ap, cmd);
    358 	rv = op_fcntl(fd, cmd, va_arg(ap, void *));
    359 	va_end(ap);
    360 	return rv;
    361 }
    362 
    363 /*
    364  * write cannot issue a standard debug printf due to recursion
    365  */
    366 ssize_t
    367 write(int fd, const void *buf, size_t blen)
    368 {
    369 	ssize_t (*op_write)(int, const void *, size_t);
    370 
    371 	if (fd_isrump(fd)) {
    372 		fd = fd_host2rump(fd);
    373 		op_write = GETSYSCALL(rump, WRITE);
    374 	} else {
    375 		op_write = GETSYSCALL(host, WRITE);
    376 	}
    377 
    378 	return op_write(fd, buf, blen);
    379 }
    380 
    381 /*
    382  * dup2 is special.  we allow dup2 of a rump kernel fd to 0-2 since
    383  * many programs do that.  dup2 of a rump kernel fd to another value
    384  * not >= fdoff is an error.
    385  *
    386  * Note: cannot rump2host newd, because it is often hardcoded.
    387  */
    388 int
    389 dup2(int oldd, int newd)
    390 {
    391 	int (*host_dup2)(int, int);
    392 	int rv;
    393 
    394 	DPRINTF(("dup2 -> %d (o) -> %d (n)\n", oldd, newd));
    395 
    396 	if (fd_isrump(oldd)) {
    397 		if (!(newd >= 0 && newd <= 2))
    398 			return EBADF;
    399 		oldd = fd_host2rump(oldd);
    400 		rv = rump_sys_dup2(oldd, newd);
    401 		if (rv != -1)
    402 			dup2mask |= 1<<newd;
    403 	} else {
    404 		host_dup2 = syscalls[DUALCALL_DUP2].bs_host;
    405 		rv = host_dup2(oldd, newd);
    406 	}
    407 
    408 	return rv;
    409 }
    410 
    411 /*
    412  * We just wrap fork the appropriate rump client calls to preserve
    413  * the file descriptors of the forked parent in the child, but
    414  * prevent double use of connection fd.
    415  */
    416 pid_t
    417 fork()
    418 {
    419 	struct rumpclient_fork *rf;
    420 	pid_t rv;
    421 
    422 	DPRINTF(("fork\n"));
    423 
    424 	if ((rf = rumpclient_prefork()) == NULL)
    425 		return -1;
    426 
    427 	switch ((rv = host_fork())) {
    428 	case -1:
    429 		/* XXX: cancel rf */
    430 		break;
    431 	case 0:
    432 		if (rumpclient_fork_init(rf) == -1)
    433 			rv = -1;
    434 		break;
    435 	default:
    436 		break;
    437 	}
    438 
    439 	DPRINTF(("fork returns %d\n", rv));
    440 	return rv;
    441 }
    442 
    443 /*
    444  * select is done by calling poll.
    445  */
    446 int
    447 LIBCSELECT(int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
    448 	struct timeval *timeout)
    449 {
    450 	struct pollfd *pfds;
    451 	struct timespec ts, *tsp = NULL;
    452 	nfds_t realnfds;
    453 	int i, j;
    454 	int rv, incr;
    455 
    456 	DPRINTF(("select\n"));
    457 
    458 	/*
    459 	 * Well, first we must scan the fds to figure out how many
    460 	 * fds there really are.  This is because up to and including
    461 	 * nb5 poll() silently refuses nfds > process_maxopen_fds.
    462 	 * Seems to be fixed in current, thank the maker.
    463 	 * god damn cluster...bomb.
    464 	 */
    465 
    466 	for (i = 0, realnfds = 0; i < nfds; i++) {
    467 		if (readfds && FD_ISSET(i, readfds)) {
    468 			realnfds++;
    469 			continue;
    470 		}
    471 		if (writefds && FD_ISSET(i, writefds)) {
    472 			realnfds++;
    473 			continue;
    474 		}
    475 		if (exceptfds && FD_ISSET(i, exceptfds)) {
    476 			realnfds++;
    477 			continue;
    478 		}
    479 	}
    480 
    481 	if (realnfds) {
    482 		pfds = malloc(sizeof(*pfds) * realnfds);
    483 		if (!pfds)
    484 			return -1;
    485 	} else {
    486 		pfds = NULL;
    487 	}
    488 
    489 	for (i = 0, j = 0; i < nfds; i++) {
    490 		incr = 0;
    491 		pfds[j].events = pfds[j].revents = 0;
    492 		if (readfds && FD_ISSET(i, readfds)) {
    493 			pfds[j].fd = i;
    494 			pfds[j].events |= POLLIN;
    495 			incr=1;
    496 		}
    497 		if (writefds && FD_ISSET(i, writefds)) {
    498 			pfds[j].fd = i;
    499 			pfds[j].events |= POLLOUT;
    500 			incr=1;
    501 		}
    502 		if (exceptfds && FD_ISSET(i, exceptfds)) {
    503 			pfds[j].fd = i;
    504 			pfds[j].events |= POLLHUP|POLLERR;
    505 			incr=1;
    506 		}
    507 		if (incr)
    508 			j++;
    509 	}
    510 
    511 	if (timeout) {
    512 		TIMEVAL_TO_TIMESPEC(timeout, &ts);
    513 		tsp = &ts;
    514 	}
    515 	rv = pollts(pfds, realnfds, tsp, NULL);
    516 	if (rv <= 0)
    517 		goto out;
    518 
    519 	/*
    520 	 * ok, harvest results.  first zero out entries (can't use
    521 	 * FD_ZERO for the obvious select-me-not reason).  whee.
    522 	 */
    523 	for (i = 0; i < nfds; i++) {
    524 		if (readfds)
    525 			FD_CLR(i, readfds);
    526 		if (writefds)
    527 			FD_CLR(i, writefds);
    528 		if (exceptfds)
    529 			FD_CLR(i, exceptfds);
    530 	}
    531 
    532 	/* and then plug in the results */
    533 	for (i = 0; i < (int)realnfds; i++) {
    534 		if (readfds) {
    535 			if (pfds[i].revents & POLLIN) {
    536 				FD_SET(pfds[i].fd, readfds);
    537 			}
    538 		}
    539 		if (writefds) {
    540 			if (pfds[i].revents & POLLOUT) {
    541 				FD_SET(pfds[i].fd, writefds);
    542 			}
    543 		}
    544 		if (exceptfds) {
    545 			if (pfds[i].revents & (POLLHUP|POLLERR)) {
    546 				FD_SET(pfds[i].fd, exceptfds);
    547 			}
    548 		}
    549 	}
    550 
    551  out:
    552 	free(pfds);
    553 	return rv;
    554 }
    555 
    556 static void
    557 checkpoll(struct pollfd *fds, nfds_t nfds, int *hostcall, int *rumpcall)
    558 {
    559 	nfds_t i;
    560 
    561 	for (i = 0; i < nfds; i++) {
    562 		if (fds[i].fd == -1)
    563 			continue;
    564 
    565 		if (fd_isrump(fds[i].fd))
    566 			(*rumpcall)++;
    567 		else
    568 			(*hostcall)++;
    569 	}
    570 }
    571 
    572 static void
    573 adjustpoll(struct pollfd *fds, nfds_t nfds, int (*fdadj)(int))
    574 {
    575 	nfds_t i;
    576 
    577 	for (i = 0; i < nfds; i++) {
    578 		fds[i].fd = fdadj(fds[i].fd);
    579 	}
    580 }
    581 
    582 /*
    583  * poll is easy as long as the call comes in the fds only in one
    584  * kernel.  otherwise its quite tricky...
    585  */
    586 struct pollarg {
    587 	struct pollfd *pfds;
    588 	nfds_t nfds;
    589 	const struct timespec *ts;
    590 	const sigset_t *sigmask;
    591 	int pipefd;
    592 	int errnum;
    593 };
    594 
    595 static void *
    596 hostpoll(void *arg)
    597 {
    598 	int (*op_pollts)(struct pollfd *, nfds_t, const struct timespec *,
    599 			 const sigset_t *);
    600 	struct pollarg *parg = arg;
    601 	intptr_t rv;
    602 
    603 	op_pollts = syscalls[DUALCALL_POLLTS].bs_host;
    604 	rv = op_pollts(parg->pfds, parg->nfds, parg->ts, parg->sigmask);
    605 	if (rv == -1)
    606 		parg->errnum = errno;
    607 	rump_sys_write(parg->pipefd, &rv, sizeof(rv));
    608 
    609 	return (void *)(intptr_t)rv;
    610 }
    611 
    612 int
    613 LIBCPOLLTS(struct pollfd *fds, nfds_t nfds, const struct timespec *ts,
    614 	const sigset_t *sigmask)
    615 {
    616 	int (*op_pollts)(struct pollfd *, nfds_t, const struct timespec *,
    617 			 const sigset_t *);
    618 	int (*host_close)(int);
    619 	int hostcall = 0, rumpcall = 0;
    620 	pthread_t pt;
    621 	nfds_t i;
    622 	int rv;
    623 
    624 	DPRINTF(("poll\n"));
    625 	checkpoll(fds, nfds, &hostcall, &rumpcall);
    626 
    627 	if (hostcall && rumpcall) {
    628 		struct pollfd *pfd_host = NULL, *pfd_rump = NULL;
    629 		int rpipe[2] = {-1,-1}, hpipe[2] = {-1,-1};
    630 		struct pollarg parg;
    631 		uintptr_t lrv;
    632 		int sverrno = 0, trv;
    633 
    634 		/*
    635 		 * ok, this is where it gets tricky.  We must support
    636 		 * this since it's a very common operation in certain
    637 		 * types of software (telnet, netcat, etc).  We allocate
    638 		 * two vectors and run two poll commands in separate
    639 		 * threads.  Whichever returns first "wins" and the
    640 		 * other kernel's fds won't show activity.
    641 		 */
    642 		rv = -1;
    643 
    644 		/* allocate full vector for O(n) joining after call */
    645 		pfd_host = malloc(sizeof(*pfd_host)*(nfds+1));
    646 		if (!pfd_host)
    647 			goto out;
    648 		pfd_rump = malloc(sizeof(*pfd_rump)*(nfds+1));
    649 		if (!pfd_rump) {
    650 			goto out;
    651 		}
    652 
    653 		/* split vectors */
    654 		for (i = 0; i < nfds; i++) {
    655 			if (fds[i].fd == -1) {
    656 				pfd_host[i].fd = -1;
    657 				pfd_rump[i].fd = -1;
    658 			} else if (fd_isrump(fds[i].fd)) {
    659 				pfd_host[i].fd = -1;
    660 				pfd_rump[i].fd = fd_host2rump(fds[i].fd);
    661 				pfd_rump[i].events = fds[i].events;
    662 			} else {
    663 				pfd_rump[i].fd = -1;
    664 				pfd_host[i].fd = fds[i].fd;
    665 				pfd_host[i].events = fds[i].events;
    666 			}
    667 			fds[i].revents = 0;
    668 		}
    669 
    670 		/*
    671 		 * then, open two pipes, one for notifications
    672 		 * to each kernel.
    673 		 */
    674 		if (rump_sys_pipe(rpipe) == -1)
    675 			goto out;
    676 		if (pipe(hpipe) == -1)
    677 			goto out;
    678 
    679 		pfd_host[nfds].fd = hpipe[0];
    680 		pfd_host[nfds].events = POLLIN;
    681 		pfd_rump[nfds].fd = rpipe[0];
    682 		pfd_rump[nfds].events = POLLIN;
    683 
    684 		/*
    685 		 * then, create a thread to do host part and meanwhile
    686 		 * do rump kernel part right here
    687 		 */
    688 
    689 		parg.pfds = pfd_host;
    690 		parg.nfds = nfds+1;
    691 		parg.ts = ts;
    692 		parg.sigmask = sigmask;
    693 		parg.pipefd = rpipe[1];
    694 		pthread_create(&pt, NULL, hostpoll, &parg);
    695 
    696 		op_pollts = syscalls[DUALCALL_POLLTS].bs_rump;
    697 		lrv = op_pollts(pfd_rump, nfds+1, ts, NULL);
    698 		sverrno = errno;
    699 		write(hpipe[1], &rv, sizeof(rv));
    700 		pthread_join(pt, (void *)&trv);
    701 
    702 		/* check who "won" and merge results */
    703 		if (lrv != 0 && pfd_host[nfds].revents & POLLIN) {
    704 			rv = trv;
    705 
    706 			for (i = 0; i < nfds; i++) {
    707 				if (pfd_rump[i].fd != -1)
    708 					fds[i].revents = pfd_rump[i].revents;
    709 			}
    710 			sverrno = parg.errnum;
    711 		} else if (trv != 0 && pfd_rump[nfds].revents & POLLIN) {
    712 			rv = trv;
    713 
    714 			for (i = 0; i < nfds; i++) {
    715 				if (pfd_host[i].fd != -1)
    716 					fds[i].revents = pfd_host[i].revents;
    717 			}
    718 		} else {
    719 			rv = 0;
    720 		}
    721 
    722  out:
    723 		host_close = syscalls[DUALCALL_CLOSE].bs_host;
    724 		if (rpipe[0] != -1)
    725 			rump_sys_close(rpipe[0]);
    726 		if (rpipe[1] != -1)
    727 			rump_sys_close(rpipe[1]);
    728 		if (hpipe[0] != -1)
    729 			host_close(hpipe[0]);
    730 		if (hpipe[1] != -1)
    731 			host_close(hpipe[1]);
    732 		free(pfd_host);
    733 		free(pfd_rump);
    734 		errno = sverrno;
    735 	} else {
    736 		if (hostcall) {
    737 			op_pollts = syscalls[DUALCALL_POLLTS].bs_host;
    738 		} else {
    739 			op_pollts = syscalls[DUALCALL_POLLTS].bs_rump;
    740 			adjustpoll(fds, nfds, fd_host2rump);
    741 		}
    742 
    743 		rv = op_pollts(fds, nfds, ts, sigmask);
    744 		if (rumpcall)
    745 			adjustpoll(fds, nfds, fd_rump2host);
    746 	}
    747 
    748 	return rv;
    749 }
    750 
    751 int
    752 LIBCPOLL(struct pollfd *fds, nfds_t nfds, int timeout)
    753 {
    754 	struct timespec ts;
    755 	struct timespec *tsp = NULL;
    756 
    757 	if (timeout != INFTIM) {
    758 		ts.tv_sec = timeout / 1000;
    759 		ts.tv_nsec = (timeout % 1000) * 1000*1000;
    760 
    761 		tsp = &ts;
    762 	}
    763 
    764 	return pollts(fds, nfds, tsp, NULL);
    765 }
    766 
    767 int
    768 kqueue(void)
    769 {
    770 
    771 	fprintf(stderr, "kqueue unsupported");
    772 	abort();
    773 	/*NOTREACHED*/
    774 }
    775 
    776 /*ARGSUSED*/
    777 int
    778 kevent(int kq, const struct kevent *changelist, size_t nchanges,
    779 	struct kevent *eventlist, size_t nevents,
    780 	const struct timespec *timeout)
    781 {
    782 
    783 	fprintf(stderr, "kqueue unsupported");
    784 	abort();
    785 	/*NOTREACHED*/
    786 }
    787 
    788 /*
    789  * Rest are std type calls.
    790  */
    791 
    792 FDCALL(int, bind, DUALCALL_BIND,					\
    793 	(int fd, const struct sockaddr *name, socklen_t namelen),	\
    794 	(int, const struct sockaddr *, socklen_t),			\
    795 	(fd, name, namelen))
    796 
    797 FDCALL(int, connect, DUALCALL_CONNECT,					\
    798 	(int fd, const struct sockaddr *name, socklen_t namelen),	\
    799 	(int, const struct sockaddr *, socklen_t),			\
    800 	(fd, name, namelen))
    801 
    802 FDCALL(int, getpeername, DUALCALL_GETPEERNAME,				\
    803 	(int fd, struct sockaddr *name, socklen_t *namelen),		\
    804 	(int, struct sockaddr *, socklen_t *),				\
    805 	(fd, name, namelen))
    806 
    807 FDCALL(int, getsockname, DUALCALL_GETSOCKNAME, 				\
    808 	(int fd, struct sockaddr *name, socklen_t *namelen),		\
    809 	(int, struct sockaddr *, socklen_t *),				\
    810 	(fd, name, namelen))
    811 
    812 FDCALL(int, listen, DUALCALL_LISTEN,	 				\
    813 	(int fd, int backlog),						\
    814 	(int, int),							\
    815 	(fd, backlog))
    816 
    817 FDCALL(ssize_t, recvfrom, DUALCALL_RECVFROM, 				\
    818 	(int fd, void *buf, size_t len, int flags,			\
    819 	    struct sockaddr *from, socklen_t *fromlen),			\
    820 	(int, void *, size_t, int, struct sockaddr *, socklen_t *),	\
    821 	(fd, buf, len, flags, from, fromlen))
    822 
    823 FDCALL(ssize_t, sendto, DUALCALL_SENDTO, 				\
    824 	(int fd, const void *buf, size_t len, int flags,		\
    825 	    const struct sockaddr *to, socklen_t tolen),		\
    826 	(int, const void *, size_t, int,				\
    827 	    const struct sockaddr *, socklen_t),			\
    828 	(fd, buf, len, flags, to, tolen))
    829 
    830 FDCALL(ssize_t, recvmsg, DUALCALL_RECVMSG, 				\
    831 	(int fd, struct msghdr *msg, int flags),			\
    832 	(int, struct msghdr *, int),					\
    833 	(fd, msg, flags))
    834 
    835 FDCALL(ssize_t, sendmsg, DUALCALL_SENDMSG, 				\
    836 	(int fd, const struct msghdr *msg, int flags),			\
    837 	(int, const struct msghdr *, int),				\
    838 	(fd, msg, flags))
    839 
    840 FDCALL(int, getsockopt, DUALCALL_GETSOCKOPT, 				\
    841 	(int fd, int level, int optn, void *optval, socklen_t *optlen),	\
    842 	(int, int, int, void *, socklen_t *),				\
    843 	(fd, level, optn, optval, optlen))
    844 
    845 FDCALL(int, setsockopt, DUALCALL_SETSOCKOPT, 				\
    846 	(int fd, int level, int optn,					\
    847 	    const void *optval, socklen_t optlen),			\
    848 	(int, int, int, const void *, socklen_t),			\
    849 	(fd, level, optn, optval, optlen))
    850 
    851 FDCALL(int, shutdown, DUALCALL_SHUTDOWN, 				\
    852 	(int fd, int how),						\
    853 	(int, int),							\
    854 	(fd, how))
    855 
    856 #if _FORTIFY_SOURCE > 0
    857 #define STUB(fun) __ssp_weak_name(fun)
    858 ssize_t _sys_readlink(const char * __restrict, char * __restrict, size_t);
    859 ssize_t
    860 STUB(readlink)(const char * __restrict path, char * __restrict buf,
    861     size_t bufsiz)
    862 {
    863 	return _sys_readlink(path, buf, bufsiz);
    864 }
    865 
    866 char *_sys_getcwd(char *, size_t);
    867 char *
    868 STUB(getcwd)(char *buf, size_t size)
    869 {
    870 	return _sys_getcwd(buf, size);
    871 }
    872 #else
    873 #define STUB(fun) fun
    874 #endif
    875 
    876 FDCALL(ssize_t, STUB(read), DUALCALL_READ,				\
    877 	(int fd, void *buf, size_t buflen),				\
    878 	(int, void *, size_t),						\
    879 	(fd, buf, buflen))
    880 
    881 FDCALL(ssize_t, readv, DUALCALL_READV, 					\
    882 	(int fd, const struct iovec *iov, int iovcnt),			\
    883 	(int, const struct iovec *, int),				\
    884 	(fd, iov, iovcnt))
    885 
    886 FDCALL(ssize_t, writev, DUALCALL_WRITEV, 				\
    887 	(int fd, const struct iovec *iov, int iovcnt),			\
    888 	(int, const struct iovec *, int),				\
    889 	(fd, iov, iovcnt))
    890 
    891 FDCALL(int, close, DUALCALL_CLOSE,	 				\
    892 	(int fd),							\
    893 	(int),								\
    894 	(fd))
    895