Home | History | Annotate | Line # | Download | only in librumphijack
hijack.c revision 1.21
      1 /*      $NetBSD: hijack.c,v 1.21 2011/01/26 18:48:32 christos 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 1.21 2011/01/26 18:48:32 christos 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 	int (*rumpcinit)(void);
    191 	void **rumpcdlsym;
    192 	void *hand;
    193 	unsigned i, j;
    194 
    195 	hand = dlopen("librumpclient.so", RTLD_LAZY|RTLD_GLOBAL);
    196 	if (!hand)
    197 		err(1, "cannot open librumpclient.so");
    198 	rumpcinit = dlsym(hand, "rumpclient_init");
    199 	_DIAGASSERT(rumpcinit);
    200 
    201 	rumpcdlsym = dlsym(hand, "rumpclient_dlsym");
    202 	*rumpcdlsym = hijackdlsym;
    203 	host_fork = dlsym(RTLD_NEXT, "fork");
    204 
    205 	/*
    206 	 * In theory cannot print anything during lookups because
    207 	 * we might not have the call vector set up.  so, the errx()
    208 	 * is a bit of a strech, but it might work.
    209 	 */
    210 
    211 	for (i = 0; i < DUALCALL__NUM; i++) {
    212 		/* build runtime O(1) access */
    213 		for (j = 0; j < __arraycount(syscnames); j++) {
    214 			if (syscnames[j].scm_callnum == i)
    215 				break;
    216 		}
    217 
    218 		if (j == __arraycount(syscnames))
    219 			errx(1, "rumphijack error: syscall pos %d missing", i);
    220 
    221 		syscalls[i].bs_host = dlsym(hand,syscnames[j].scm_hostname);
    222 		if (syscalls[i].bs_host == NULL)
    223 			errx(1, "hostcall %s not found missing",
    224 			    syscnames[j].scm_hostname);
    225 
    226 		syscalls[i].bs_rump = dlsym(hand,syscnames[j].scm_rumpname);
    227 		if (syscalls[i].bs_rump == NULL)
    228 			errx(1, "rumpcall %s not found missing",
    229 			    syscnames[j].scm_rumpname);
    230 	}
    231 
    232 	if (rumpcinit() == -1)
    233 		err(1, "rumpclient init");
    234 }
    235 
    236 /* XXX: need runtime selection.  low for now due to FD_SETSIZE */
    237 #define HIJACK_FDOFF 128
    238 #define HIJACK_SELECT 128 /* XXX */
    239 #define HIJACK_ASSERT 128 /* XXX */
    240 static int
    241 fd_rump2host(int fd)
    242 {
    243 
    244 	if (fd == -1)
    245 		return fd;
    246 
    247 	if (!ISDUP2D(fd))
    248 		fd += HIJACK_FDOFF;
    249 
    250 	return fd;
    251 }
    252 
    253 static int
    254 fd_host2rump(int fd)
    255 {
    256 
    257 	if (!ISDUP2D(fd))
    258 		fd -= HIJACK_FDOFF;
    259 	return fd;
    260 }
    261 
    262 static bool
    263 fd_isrump(int fd)
    264 {
    265 
    266 	return ISDUP2D(fd) || fd >= HIJACK_FDOFF;
    267 }
    268 
    269 #define assertfd(_fd_) assert(ISDUP2D(_fd_) || (_fd_) >= HIJACK_ASSERT)
    270 #undef HIJACK_FDOFF
    271 
    272 int __socket30(int, int, int);
    273 int
    274 __socket30(int domain, int type, int protocol)
    275 {
    276 	int (*op_socket)(int, int, int);
    277 	int fd;
    278 	bool dohost;
    279 
    280 	dohost = hostlocalsockets && (domain == AF_LOCAL);
    281 
    282 	if (dohost)
    283 		op_socket = GETSYSCALL(host, SOCKET);
    284 	else
    285 		op_socket = GETSYSCALL(rump, SOCKET);
    286 	fd = op_socket(domain, type, protocol);
    287 
    288 	if (!dohost)
    289 		fd = fd_rump2host(fd);
    290 	DPRINTF(("socket <- %d\n", fd));
    291 
    292 	return fd;
    293 }
    294 
    295 int
    296 accept(int s, struct sockaddr *addr, socklen_t *addrlen)
    297 {
    298 	int (*op_accept)(int, struct sockaddr *, socklen_t *);
    299 	int fd;
    300 	bool isrump;
    301 
    302 	isrump = fd_isrump(s);
    303 
    304 	DPRINTF(("accept -> %d", s));
    305 	if (isrump) {
    306 		op_accept = GETSYSCALL(rump, ACCEPT);
    307 		s = fd_host2rump(s);
    308 	} else {
    309 		op_accept = GETSYSCALL(host, ACCEPT);
    310 	}
    311 	fd = op_accept(s, addr, addrlen);
    312 	if (fd != -1 && isrump)
    313 		fd = fd_rump2host(fd);
    314 
    315 	DPRINTF((" <- %d\n", fd));
    316 
    317 	return fd;
    318 }
    319 
    320 /*
    321  * ioctl and fcntl are varargs calls and need special treatment
    322  */
    323 int
    324 ioctl(int fd, unsigned long cmd, ...)
    325 {
    326 	int (*op_ioctl)(int, unsigned long cmd, ...);
    327 	va_list ap;
    328 	int rv;
    329 
    330 	DPRINTF(("ioctl -> %d\n", fd));
    331 	if (fd_isrump(fd)) {
    332 		fd = fd_host2rump(fd);
    333 		op_ioctl = GETSYSCALL(rump, IOCTL);
    334 	} else {
    335 		op_ioctl = GETSYSCALL(host, IOCTL);
    336 	}
    337 
    338 	va_start(ap, cmd);
    339 	rv = op_ioctl(fd, cmd, va_arg(ap, void *));
    340 	va_end(ap);
    341 	return rv;
    342 }
    343 
    344 int
    345 fcntl(int fd, int cmd, ...)
    346 {
    347 	int (*op_fcntl)(int, int, ...);
    348 	va_list ap;
    349 	int rv;
    350 
    351 	DPRINTF(("fcntl -> %d\n", fd));
    352 	if (fd_isrump(fd)) {
    353 		fd = fd_host2rump(fd);
    354 		op_fcntl = GETSYSCALL(rump, FCNTL);
    355 	} else {
    356 		op_fcntl = GETSYSCALL(host, FCNTL);
    357 	}
    358 
    359 	va_start(ap, cmd);
    360 	rv = op_fcntl(fd, cmd, va_arg(ap, void *));
    361 	va_end(ap);
    362 	return rv;
    363 }
    364 
    365 /*
    366  * write cannot issue a standard debug printf due to recursion
    367  */
    368 ssize_t
    369 write(int fd, const void *buf, size_t blen)
    370 {
    371 	ssize_t (*op_write)(int, const void *, size_t);
    372 
    373 	if (fd_isrump(fd)) {
    374 		fd = fd_host2rump(fd);
    375 		op_write = GETSYSCALL(rump, WRITE);
    376 	} else {
    377 		op_write = GETSYSCALL(host, WRITE);
    378 	}
    379 
    380 	return op_write(fd, buf, blen);
    381 }
    382 
    383 /*
    384  * dup2 is special.  we allow dup2 of a rump kernel fd to 0-2 since
    385  * many programs do that.  dup2 of a rump kernel fd to another value
    386  * not >= fdoff is an error.
    387  *
    388  * Note: cannot rump2host newd, because it is often hardcoded.
    389  */
    390 int
    391 dup2(int oldd, int newd)
    392 {
    393 	int (*host_dup2)(int, int);
    394 	int rv;
    395 
    396 	DPRINTF(("dup2 -> %d (o) -> %d (n)\n", oldd, newd));
    397 
    398 	if (fd_isrump(oldd)) {
    399 		if (!(newd >= 0 && newd <= 2))
    400 			return EBADF;
    401 		oldd = fd_host2rump(oldd);
    402 		rv = rump_sys_dup2(oldd, newd);
    403 		if (rv != -1)
    404 			dup2mask |= 1<<newd;
    405 	} else {
    406 		host_dup2 = syscalls[DUALCALL_DUP2].bs_host;
    407 		rv = host_dup2(oldd, newd);
    408 	}
    409 
    410 	return rv;
    411 }
    412 
    413 /*
    414  * We just wrap fork the appropriate rump client calls to preserve
    415  * the file descriptors of the forked parent in the child, but
    416  * prevent double use of connection fd.
    417  */
    418 pid_t
    419 fork()
    420 {
    421 	struct rumpclient_fork *rf;
    422 	pid_t rv;
    423 
    424 	DPRINTF(("fork\n"));
    425 
    426 	if ((rf = rumpclient_prefork()) == NULL)
    427 		return -1;
    428 
    429 	switch ((rv = host_fork())) {
    430 	case -1:
    431 		/* XXX: cancel rf */
    432 		break;
    433 	case 0:
    434 		if (rumpclient_fork_init(rf) == -1)
    435 			rv = -1;
    436 		break;
    437 	default:
    438 		break;
    439 	}
    440 
    441 	DPRINTF(("fork returns %d\n", rv));
    442 	return rv;
    443 }
    444 
    445 /*
    446  * select is done by calling poll.
    447  */
    448 int
    449 LIBCSELECT(int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
    450 	struct timeval *timeout)
    451 {
    452 	struct pollfd *pfds;
    453 	struct timespec ts, *tsp = NULL;
    454 	nfds_t realnfds;
    455 	int i, j;
    456 	int rv, incr;
    457 
    458 	DPRINTF(("select\n"));
    459 
    460 	/*
    461 	 * Well, first we must scan the fds to figure out how many
    462 	 * fds there really are.  This is because up to and including
    463 	 * nb5 poll() silently refuses nfds > process_maxopen_fds.
    464 	 * Seems to be fixed in current, thank the maker.
    465 	 * god damn cluster...bomb.
    466 	 */
    467 
    468 	for (i = 0, realnfds = 0; i < nfds; i++) {
    469 		if (readfds && FD_ISSET(i, readfds)) {
    470 			realnfds++;
    471 			continue;
    472 		}
    473 		if (writefds && FD_ISSET(i, writefds)) {
    474 			realnfds++;
    475 			continue;
    476 		}
    477 		if (exceptfds && FD_ISSET(i, exceptfds)) {
    478 			realnfds++;
    479 			continue;
    480 		}
    481 	}
    482 
    483 	if (realnfds) {
    484 		pfds = malloc(sizeof(*pfds) * realnfds);
    485 		if (!pfds)
    486 			return -1;
    487 	} else {
    488 		pfds = NULL;
    489 	}
    490 
    491 	for (i = 0, j = 0; i < nfds; i++) {
    492 		incr = 0;
    493 		pfds[j].events = pfds[j].revents = 0;
    494 		if (readfds && FD_ISSET(i, readfds)) {
    495 			pfds[j].fd = i;
    496 			pfds[j].events |= POLLIN;
    497 			incr=1;
    498 		}
    499 		if (writefds && FD_ISSET(i, writefds)) {
    500 			pfds[j].fd = i;
    501 			pfds[j].events |= POLLOUT;
    502 			incr=1;
    503 		}
    504 		if (exceptfds && FD_ISSET(i, exceptfds)) {
    505 			pfds[j].fd = i;
    506 			pfds[j].events |= POLLHUP|POLLERR;
    507 			incr=1;
    508 		}
    509 		if (incr)
    510 			j++;
    511 	}
    512 
    513 	if (timeout) {
    514 		TIMEVAL_TO_TIMESPEC(timeout, &ts);
    515 		tsp = &ts;
    516 	}
    517 	rv = pollts(pfds, realnfds, tsp, NULL);
    518 	if (rv <= 0)
    519 		goto out;
    520 
    521 	/*
    522 	 * ok, harvest results.  first zero out entries (can't use
    523 	 * FD_ZERO for the obvious select-me-not reason).  whee.
    524 	 */
    525 	for (i = 0; i < nfds; i++) {
    526 		if (readfds)
    527 			FD_CLR(i, readfds);
    528 		if (writefds)
    529 			FD_CLR(i, writefds);
    530 		if (exceptfds)
    531 			FD_CLR(i, exceptfds);
    532 	}
    533 
    534 	/* and then plug in the results */
    535 	for (i = 0; i < (int)realnfds; i++) {
    536 		if (readfds) {
    537 			if (pfds[i].revents & POLLIN) {
    538 				FD_SET(pfds[i].fd, readfds);
    539 			}
    540 		}
    541 		if (writefds) {
    542 			if (pfds[i].revents & POLLOUT) {
    543 				FD_SET(pfds[i].fd, writefds);
    544 			}
    545 		}
    546 		if (exceptfds) {
    547 			if (pfds[i].revents & (POLLHUP|POLLERR)) {
    548 				FD_SET(pfds[i].fd, exceptfds);
    549 			}
    550 		}
    551 	}
    552 
    553  out:
    554 	free(pfds);
    555 	return rv;
    556 }
    557 
    558 static void
    559 checkpoll(struct pollfd *fds, nfds_t nfds, int *hostcall, int *rumpcall)
    560 {
    561 	nfds_t i;
    562 
    563 	for (i = 0; i < nfds; i++) {
    564 		if (fds[i].fd == -1)
    565 			continue;
    566 
    567 		if (fd_isrump(fds[i].fd))
    568 			(*rumpcall)++;
    569 		else
    570 			(*hostcall)++;
    571 	}
    572 }
    573 
    574 static void
    575 adjustpoll(struct pollfd *fds, nfds_t nfds, int (*fdadj)(int))
    576 {
    577 	nfds_t i;
    578 
    579 	for (i = 0; i < nfds; i++) {
    580 		fds[i].fd = fdadj(fds[i].fd);
    581 	}
    582 }
    583 
    584 /*
    585  * poll is easy as long as the call comes in the fds only in one
    586  * kernel.  otherwise its quite tricky...
    587  */
    588 struct pollarg {
    589 	struct pollfd *pfds;
    590 	nfds_t nfds;
    591 	const struct timespec *ts;
    592 	const sigset_t *sigmask;
    593 	int pipefd;
    594 	int errnum;
    595 };
    596 
    597 static void *
    598 hostpoll(void *arg)
    599 {
    600 	int (*op_pollts)(struct pollfd *, nfds_t, const struct timespec *,
    601 			 const sigset_t *);
    602 	struct pollarg *parg = arg;
    603 	intptr_t rv;
    604 
    605 	op_pollts = syscalls[DUALCALL_POLLTS].bs_host;
    606 	rv = op_pollts(parg->pfds, parg->nfds, parg->ts, parg->sigmask);
    607 	if (rv == -1)
    608 		parg->errnum = errno;
    609 	rump_sys_write(parg->pipefd, &rv, sizeof(rv));
    610 
    611 	return (void *)(intptr_t)rv;
    612 }
    613 
    614 int
    615 LIBCPOLLTS(struct pollfd *fds, nfds_t nfds, const struct timespec *ts,
    616 	const sigset_t *sigmask)
    617 {
    618 	int (*op_pollts)(struct pollfd *, nfds_t, const struct timespec *,
    619 			 const sigset_t *);
    620 	int (*host_close)(int);
    621 	int hostcall = 0, rumpcall = 0;
    622 	pthread_t pt;
    623 	nfds_t i;
    624 	int rv;
    625 
    626 	DPRINTF(("poll\n"));
    627 	checkpoll(fds, nfds, &hostcall, &rumpcall);
    628 
    629 	if (hostcall && rumpcall) {
    630 		struct pollfd *pfd_host = NULL, *pfd_rump = NULL;
    631 		int rpipe[2] = {-1,-1}, hpipe[2] = {-1,-1};
    632 		struct pollarg parg;
    633 		uintptr_t lrv;
    634 		int sverrno = 0, trv;
    635 
    636 		/*
    637 		 * ok, this is where it gets tricky.  We must support
    638 		 * this since it's a very common operation in certain
    639 		 * types of software (telnet, netcat, etc).  We allocate
    640 		 * two vectors and run two poll commands in separate
    641 		 * threads.  Whichever returns first "wins" and the
    642 		 * other kernel's fds won't show activity.
    643 		 */
    644 		rv = -1;
    645 
    646 		/* allocate full vector for O(n) joining after call */
    647 		pfd_host = malloc(sizeof(*pfd_host)*(nfds+1));
    648 		if (!pfd_host)
    649 			goto out;
    650 		pfd_rump = malloc(sizeof(*pfd_rump)*(nfds+1));
    651 		if (!pfd_rump) {
    652 			goto out;
    653 		}
    654 
    655 		/* split vectors */
    656 		for (i = 0; i < nfds; i++) {
    657 			if (fds[i].fd == -1) {
    658 				pfd_host[i].fd = -1;
    659 				pfd_rump[i].fd = -1;
    660 			} else if (fd_isrump(fds[i].fd)) {
    661 				pfd_host[i].fd = -1;
    662 				pfd_rump[i].fd = fd_host2rump(fds[i].fd);
    663 				pfd_rump[i].events = fds[i].events;
    664 			} else {
    665 				pfd_rump[i].fd = -1;
    666 				pfd_host[i].fd = fds[i].fd;
    667 				pfd_host[i].events = fds[i].events;
    668 			}
    669 			fds[i].revents = 0;
    670 		}
    671 
    672 		/*
    673 		 * then, open two pipes, one for notifications
    674 		 * to each kernel.
    675 		 */
    676 		if (rump_sys_pipe(rpipe) == -1)
    677 			goto out;
    678 		if (pipe(hpipe) == -1)
    679 			goto out;
    680 
    681 		pfd_host[nfds].fd = hpipe[0];
    682 		pfd_host[nfds].events = POLLIN;
    683 		pfd_rump[nfds].fd = rpipe[0];
    684 		pfd_rump[nfds].events = POLLIN;
    685 
    686 		/*
    687 		 * then, create a thread to do host part and meanwhile
    688 		 * do rump kernel part right here
    689 		 */
    690 
    691 		parg.pfds = pfd_host;
    692 		parg.nfds = nfds+1;
    693 		parg.ts = ts;
    694 		parg.sigmask = sigmask;
    695 		parg.pipefd = rpipe[1];
    696 		pthread_create(&pt, NULL, hostpoll, &parg);
    697 
    698 		op_pollts = syscalls[DUALCALL_POLLTS].bs_rump;
    699 		lrv = op_pollts(pfd_rump, nfds+1, ts, NULL);
    700 		sverrno = errno;
    701 		write(hpipe[1], &rv, sizeof(rv));
    702 		pthread_join(pt, (void *)&trv);
    703 
    704 		/* check who "won" and merge results */
    705 		if (lrv != 0 && pfd_host[nfds].revents & POLLIN) {
    706 			rv = trv;
    707 
    708 			for (i = 0; i < nfds; i++) {
    709 				if (pfd_rump[i].fd != -1)
    710 					fds[i].revents = pfd_rump[i].revents;
    711 			}
    712 			sverrno = parg.errnum;
    713 		} else if (trv != 0 && pfd_rump[nfds].revents & POLLIN) {
    714 			rv = trv;
    715 
    716 			for (i = 0; i < nfds; i++) {
    717 				if (pfd_host[i].fd != -1)
    718 					fds[i].revents = pfd_host[i].revents;
    719 			}
    720 		} else {
    721 			rv = 0;
    722 		}
    723 
    724  out:
    725 		host_close = syscalls[DUALCALL_CLOSE].bs_host;
    726 		if (rpipe[0] != -1)
    727 			rump_sys_close(rpipe[0]);
    728 		if (rpipe[1] != -1)
    729 			rump_sys_close(rpipe[1]);
    730 		if (hpipe[0] != -1)
    731 			host_close(hpipe[0]);
    732 		if (hpipe[1] != -1)
    733 			host_close(hpipe[1]);
    734 		free(pfd_host);
    735 		free(pfd_rump);
    736 		errno = sverrno;
    737 	} else {
    738 		if (hostcall) {
    739 			op_pollts = syscalls[DUALCALL_POLLTS].bs_host;
    740 		} else {
    741 			op_pollts = syscalls[DUALCALL_POLLTS].bs_rump;
    742 			adjustpoll(fds, nfds, fd_host2rump);
    743 		}
    744 
    745 		rv = op_pollts(fds, nfds, ts, sigmask);
    746 		if (rumpcall)
    747 			adjustpoll(fds, nfds, fd_rump2host);
    748 	}
    749 
    750 	return rv;
    751 }
    752 
    753 int
    754 LIBCPOLL(struct pollfd *fds, nfds_t nfds, int timeout)
    755 {
    756 	struct timespec ts;
    757 	struct timespec *tsp = NULL;
    758 
    759 	if (timeout != INFTIM) {
    760 		ts.tv_sec = timeout / 1000;
    761 		ts.tv_nsec = (timeout % 1000) * 1000*1000;
    762 
    763 		tsp = &ts;
    764 	}
    765 
    766 	return pollts(fds, nfds, tsp, NULL);
    767 }
    768 
    769 int
    770 kqueue(void)
    771 {
    772 
    773 	fprintf(stderr, "kqueue unsupported");
    774 	abort();
    775 	/*NOTREACHED*/
    776 }
    777 
    778 /*ARGSUSED*/
    779 int
    780 kevent(int kq, const struct kevent *changelist, size_t nchanges,
    781 	struct kevent *eventlist, size_t nevents,
    782 	const struct timespec *timeout)
    783 {
    784 
    785 	fprintf(stderr, "kqueue unsupported");
    786 	abort();
    787 	/*NOTREACHED*/
    788 }
    789 
    790 /*
    791  * Rest are std type calls.
    792  */
    793 
    794 FDCALL(int, bind, DUALCALL_BIND,					\
    795 	(int fd, const struct sockaddr *name, socklen_t namelen),	\
    796 	(int, const struct sockaddr *, socklen_t),			\
    797 	(fd, name, namelen))
    798 
    799 FDCALL(int, connect, DUALCALL_CONNECT,					\
    800 	(int fd, const struct sockaddr *name, socklen_t namelen),	\
    801 	(int, const struct sockaddr *, socklen_t),			\
    802 	(fd, name, namelen))
    803 
    804 FDCALL(int, getpeername, DUALCALL_GETPEERNAME,				\
    805 	(int fd, struct sockaddr *name, socklen_t *namelen),		\
    806 	(int, struct sockaddr *, socklen_t *),				\
    807 	(fd, name, namelen))
    808 
    809 FDCALL(int, getsockname, DUALCALL_GETSOCKNAME, 				\
    810 	(int fd, struct sockaddr *name, socklen_t *namelen),		\
    811 	(int, struct sockaddr *, socklen_t *),				\
    812 	(fd, name, namelen))
    813 
    814 FDCALL(int, listen, DUALCALL_LISTEN,	 				\
    815 	(int fd, int backlog),						\
    816 	(int, int),							\
    817 	(fd, backlog))
    818 
    819 FDCALL(ssize_t, recvfrom, DUALCALL_RECVFROM, 				\
    820 	(int fd, void *buf, size_t len, int flags,			\
    821 	    struct sockaddr *from, socklen_t *fromlen),			\
    822 	(int, void *, size_t, int, struct sockaddr *, socklen_t *),	\
    823 	(fd, buf, len, flags, from, fromlen))
    824 
    825 FDCALL(ssize_t, sendto, DUALCALL_SENDTO, 				\
    826 	(int fd, const void *buf, size_t len, int flags,		\
    827 	    const struct sockaddr *to, socklen_t tolen),		\
    828 	(int, const void *, size_t, int,				\
    829 	    const struct sockaddr *, socklen_t),			\
    830 	(fd, buf, len, flags, to, tolen))
    831 
    832 FDCALL(ssize_t, recvmsg, DUALCALL_RECVMSG, 				\
    833 	(int fd, struct msghdr *msg, int flags),			\
    834 	(int, struct msghdr *, int),					\
    835 	(fd, msg, flags))
    836 
    837 FDCALL(ssize_t, sendmsg, DUALCALL_SENDMSG, 				\
    838 	(int fd, const struct msghdr *msg, int flags),			\
    839 	(int, const struct msghdr *, int),				\
    840 	(fd, msg, flags))
    841 
    842 FDCALL(int, getsockopt, DUALCALL_GETSOCKOPT, 				\
    843 	(int fd, int level, int optn, void *optval, socklen_t *optlen),	\
    844 	(int, int, int, void *, socklen_t *),				\
    845 	(fd, level, optn, optval, optlen))
    846 
    847 FDCALL(int, setsockopt, DUALCALL_SETSOCKOPT, 				\
    848 	(int fd, int level, int optn,					\
    849 	    const void *optval, socklen_t optlen),			\
    850 	(int, int, int, const void *, socklen_t),			\
    851 	(fd, level, optn, optval, optlen))
    852 
    853 FDCALL(int, shutdown, DUALCALL_SHUTDOWN, 				\
    854 	(int fd, int how),						\
    855 	(int, int),							\
    856 	(fd, how))
    857 
    858 #if _FORTIFY_SOURCE > 0
    859 #define STUB(fun) __ssp_weak_name(fun)
    860 ssize_t _sys_readlink(const char * __restrict, char * __restrict, size_t);
    861 ssize_t
    862 STUB(readlink)(const char * __restrict path, char * __restrict buf,
    863     size_t bufsiz)
    864 {
    865 	return _sys_readlink(path, buf, bufsiz);
    866 }
    867 
    868 char *_sys_getcwd(char *, size_t);
    869 char *
    870 STUB(getcwd)(char *buf, size_t size)
    871 {
    872 	return _sys_getcwd(buf, size);
    873 }
    874 #else
    875 #define STUB(fun) fun
    876 #endif
    877 
    878 FDCALL(ssize_t, STUB(read), DUALCALL_READ,				\
    879 	(int fd, void *buf, size_t buflen),				\
    880 	(int, void *, size_t),						\
    881 	(fd, buf, buflen))
    882 
    883 FDCALL(ssize_t, readv, DUALCALL_READV, 					\
    884 	(int fd, const struct iovec *iov, int iovcnt),			\
    885 	(int, const struct iovec *, int),				\
    886 	(fd, iov, iovcnt))
    887 
    888 FDCALL(ssize_t, writev, DUALCALL_WRITEV, 				\
    889 	(int fd, const struct iovec *iov, int iovcnt),			\
    890 	(int, const struct iovec *, int),				\
    891 	(fd, iov, iovcnt))
    892 
    893 FDCALL(int, close, DUALCALL_CLOSE,	 				\
    894 	(int fd),							\
    895 	(int),								\
    896 	(fd))
    897