Home | History | Annotate | Line # | Download | only in librumpuser
rumpuser_sp.c revision 1.22
      1 /*      $NetBSD: rumpuser_sp.c,v 1.22 2010/11/30 22:32:01 pooka Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 2010 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 /*
     29  * Sysproxy routines.  This provides system RPC support over host sockets.
     30  * The most notable limitation is that the client and server must share
     31  * the same ABI.  This does not mean that they have to be the same
     32  * machine or that they need to run the same version of the host OS,
     33  * just that they must agree on the data structures.  This even *might*
     34  * work correctly from one hardware architecture to another.
     35  *
     36  * Not finished yet, i.e. don't use in production.  Lacks locking plus
     37  * handling of multiple clients and unexpected connection closes.
     38  */
     39 
     40 #include <sys/cdefs.h>
     41 __RCSID("$NetBSD: rumpuser_sp.c,v 1.22 2010/11/30 22:32:01 pooka Exp $");
     42 
     43 #include <sys/types.h>
     44 #include <sys/atomic.h>
     45 #include <sys/mman.h>
     46 #include <sys/socket.h>
     47 
     48 #include <arpa/inet.h>
     49 #include <netinet/in.h>
     50 #include <netinet/tcp.h>
     51 
     52 #include <assert.h>
     53 #include <errno.h>
     54 #include <fcntl.h>
     55 #include <poll.h>
     56 #include <pthread.h>
     57 #include <stdarg.h>
     58 #include <stdio.h>
     59 #include <stdlib.h>
     60 #include <string.h>
     61 #include <unistd.h>
     62 
     63 #include <rump/rumpuser.h>
     64 #include "rumpuser_int.h"
     65 
     66 #include "sp_common.c"
     67 
     68 #ifndef MAXCLI
     69 #define MAXCLI 256
     70 #endif
     71 #ifndef MAXWORKER
     72 #define MAXWORKER 128
     73 #endif
     74 #ifndef IDLEWORKER
     75 #define IDLEWORKER 16
     76 #endif
     77 int rumpsp_maxworker = MAXWORKER;
     78 int rumpsp_idleworker = IDLEWORKER;
     79 
     80 static struct pollfd pfdlist[MAXCLI];
     81 static struct spclient spclist[MAXCLI];
     82 static unsigned int disco;
     83 
     84 static struct rumpuser_sp_ops spops;
     85 
     86 /*
     87  * Manual wrappers, since librump does not have access to the
     88  * user namespace wrapped interfaces.
     89  */
     90 
     91 static void
     92 lwproc_switch(struct lwp *l)
     93 {
     94 
     95 	spops.spop_schedule();
     96 	spops.spop_lwproc_switch(l);
     97 	spops.spop_unschedule();
     98 }
     99 
    100 static void
    101 lwproc_release(void)
    102 {
    103 
    104 	spops.spop_schedule();
    105 	spops.spop_lwproc_release();
    106 	spops.spop_unschedule();
    107 }
    108 
    109 static int
    110 lwproc_newproc(struct spclient *spc)
    111 {
    112 	int rv;
    113 
    114 	spops.spop_schedule();
    115 	rv = spops.spop_lwproc_newproc(spc);
    116 	spops.spop_unschedule();
    117 
    118 	return rv;
    119 }
    120 
    121 static int
    122 lwproc_newlwp(pid_t pid)
    123 {
    124 	int rv;
    125 
    126 	spops.spop_schedule();
    127 	rv = spops.spop_lwproc_newlwp(pid);
    128 	spops.spop_unschedule();
    129 
    130 	return rv;
    131 }
    132 
    133 static struct lwp *
    134 lwproc_curlwp(void)
    135 {
    136 	struct lwp *l;
    137 
    138 	spops.spop_schedule();
    139 	l = spops.spop_lwproc_curlwp();
    140 	spops.spop_unschedule();
    141 
    142 	return l;
    143 }
    144 
    145 static pid_t
    146 lwproc_getpid(void)
    147 {
    148 	pid_t p;
    149 
    150 	spops.spop_schedule();
    151 	p = spops.spop_getpid();
    152 	spops.spop_unschedule();
    153 
    154 	return p;
    155 }
    156 
    157 static int
    158 rumpsyscall(int sysnum, void *data, register_t *retval)
    159 {
    160 	int rv;
    161 
    162 	spops.spop_schedule();
    163 	rv = spops.spop_syscall(sysnum, data, retval);
    164 	spops.spop_unschedule();
    165 
    166 	return rv;
    167 }
    168 
    169 static uint64_t
    170 nextreq(struct spclient *spc)
    171 {
    172 	uint64_t nw;
    173 
    174 	pthread_mutex_lock(&spc->spc_mtx);
    175 	nw = spc->spc_nextreq++;
    176 	pthread_mutex_unlock(&spc->spc_mtx);
    177 
    178 	return nw;
    179 }
    180 
    181 static void
    182 send_error_resp(struct spclient *spc, uint64_t reqno, int error)
    183 {
    184 	struct rsp_hdr rhdr;
    185 
    186 	rhdr.rsp_len = sizeof(rhdr);
    187 	rhdr.rsp_reqno = reqno;
    188 	rhdr.rsp_class = RUMPSP_ERROR;
    189 	rhdr.rsp_type = 0;
    190 	rhdr.rsp_error = error;
    191 
    192 	sendlock(spc);
    193 	(void)dosend(spc, &rhdr, sizeof(rhdr));
    194 	sendunlock(spc);
    195 }
    196 
    197 static int
    198 send_syscall_resp(struct spclient *spc, uint64_t reqno, int error,
    199 	register_t *retval)
    200 {
    201 	struct rsp_hdr rhdr;
    202 	struct rsp_sysresp sysresp;
    203 	int rv;
    204 
    205 	rhdr.rsp_len = sizeof(rhdr) + sizeof(sysresp);
    206 	rhdr.rsp_reqno = reqno;
    207 	rhdr.rsp_class = RUMPSP_RESP;
    208 	rhdr.rsp_type = RUMPSP_SYSCALL;
    209 	rhdr.rsp_sysnum = 0;
    210 
    211 	sysresp.rsys_error = error;
    212 	memcpy(sysresp.rsys_retval, retval, sizeof(sysresp.rsys_retval));
    213 
    214 	sendlock(spc);
    215 	rv = dosend(spc, &rhdr, sizeof(rhdr));
    216 	rv = dosend(spc, &sysresp, sizeof(sysresp));
    217 	sendunlock(spc);
    218 
    219 	return rv;
    220 }
    221 
    222 static int
    223 copyin_req(struct spclient *spc, const void *remaddr, size_t *dlen,
    224 	int wantstr, void **resp)
    225 {
    226 	struct rsp_hdr rhdr;
    227 	struct rsp_copydata copydata;
    228 	struct respwait rw;
    229 	int rv;
    230 
    231 	DPRINTF(("copyin_req: %zu bytes from %p\n", *dlen, remaddr));
    232 
    233 	rhdr.rsp_len = sizeof(rhdr) + sizeof(copydata);
    234 	rhdr.rsp_class = RUMPSP_REQ;
    235 	if (wantstr)
    236 		rhdr.rsp_type = RUMPSP_COPYINSTR;
    237 	else
    238 		rhdr.rsp_type = RUMPSP_COPYIN;
    239 	rhdr.rsp_sysnum = 0;
    240 
    241 	copydata.rcp_addr = __UNCONST(remaddr);
    242 	copydata.rcp_len = *dlen;
    243 
    244 	putwait(spc, &rw, &rhdr);
    245 	rv = dosend(spc, &rhdr, sizeof(rhdr));
    246 	rv = dosend(spc, &copydata, sizeof(copydata));
    247 	if (rv) {
    248 		unputwait(spc, &rw);
    249 		return rv;
    250 	}
    251 
    252 	rv = waitresp(spc, &rw);
    253 
    254 	DPRINTF(("copyin: response %d\n", rv));
    255 
    256 	*resp = rw.rw_data;
    257 	if (wantstr)
    258 		*dlen = rw.rw_dlen;
    259 
    260 	return rv;
    261 
    262 }
    263 
    264 static int
    265 send_copyout_req(struct spclient *spc, const void *remaddr,
    266 	const void *data, size_t dlen)
    267 {
    268 	struct rsp_hdr rhdr;
    269 	struct rsp_copydata copydata;
    270 	int rv;
    271 
    272 	DPRINTF(("copyout_req (async): %zu bytes to %p\n", dlen, remaddr));
    273 
    274 	rhdr.rsp_len = sizeof(rhdr) + sizeof(copydata) + dlen;
    275 	rhdr.rsp_reqno = nextreq(spc);
    276 	rhdr.rsp_class = RUMPSP_REQ;
    277 	rhdr.rsp_type = RUMPSP_COPYOUT;
    278 	rhdr.rsp_sysnum = 0;
    279 
    280 	copydata.rcp_addr = __UNCONST(remaddr);
    281 	copydata.rcp_len = dlen;
    282 
    283 	sendlock(spc);
    284 	rv = dosend(spc, &rhdr, sizeof(rhdr));
    285 	rv = dosend(spc, &copydata, sizeof(copydata));
    286 	rv = dosend(spc, data, dlen);
    287 	sendunlock(spc);
    288 
    289 	return rv;
    290 }
    291 
    292 static int
    293 anonmmap_req(struct spclient *spc, size_t howmuch, void **resp)
    294 {
    295 	struct rsp_hdr rhdr;
    296 	struct respwait rw;
    297 	int rv;
    298 
    299 	DPRINTF(("anonmmap_req: %zu bytes\n", howmuch));
    300 
    301 	rhdr.rsp_len = sizeof(rhdr) + sizeof(howmuch);
    302 	rhdr.rsp_class = RUMPSP_REQ;
    303 	rhdr.rsp_type = RUMPSP_ANONMMAP;
    304 	rhdr.rsp_sysnum = 0;
    305 
    306 	putwait(spc, &rw, &rhdr);
    307 	rv = dosend(spc, &rhdr, sizeof(rhdr));
    308 	rv = dosend(spc, &howmuch, sizeof(howmuch));
    309 	if (rv) {
    310 		unputwait(spc, &rw);
    311 		return rv;
    312 	}
    313 
    314 	rv = waitresp(spc, &rw);
    315 
    316 	*resp = rw.rw_data;
    317 
    318 	DPRINTF(("anonmmap: mapped at %p\n", **(void ***)resp));
    319 
    320 	return rv;
    321 }
    322 
    323 static void
    324 spcref(struct spclient *spc)
    325 {
    326 
    327 	pthread_mutex_lock(&spc->spc_mtx);
    328 	spc->spc_refcnt++;
    329 	pthread_mutex_unlock(&spc->spc_mtx);
    330 }
    331 
    332 static void
    333 spcrelease(struct spclient *spc)
    334 {
    335 	int ref;
    336 
    337 	pthread_mutex_lock(&spc->spc_mtx);
    338 	ref = --spc->spc_refcnt;
    339 	pthread_mutex_unlock(&spc->spc_mtx);
    340 
    341 	if (ref > 0)
    342 		return;
    343 
    344 	DPRINTF(("spcrelease: spc %p fd %d\n", spc, spc->spc_fd));
    345 
    346 	_DIAGASSERT(TAILQ_EMPTY(&spc->spc_respwait));
    347 	_DIAGASSERT(spc->spc_buf == NULL);
    348 
    349 	lwproc_switch(spc->spc_mainlwp);
    350 	lwproc_release();
    351 	spc->spc_mainlwp = NULL;
    352 
    353 	close(spc->spc_fd);
    354 	spc->spc_fd = -1;
    355 	spc->spc_dying = 0;
    356 
    357 	atomic_inc_uint(&disco);
    358 }
    359 
    360 static void
    361 serv_handledisco(unsigned int idx)
    362 {
    363 	struct spclient *spc = &spclist[idx];
    364 
    365 	DPRINTF(("rump_sp: disconnecting [%u]\n", idx));
    366 
    367 	pfdlist[idx].fd = -1;
    368 	pfdlist[idx].revents = 0;
    369 	pthread_mutex_lock(&spc->spc_mtx);
    370 	spc->spc_dying = 1;
    371 	kickall(spc);
    372 	pthread_mutex_unlock(&spc->spc_mtx);
    373 
    374 	/*
    375 	 * Nobody's going to attempt to send/receive anymore,
    376 	 * so reinit info relevant to that.
    377 	 */
    378 	/*LINTED:pointer casts may be ok*/
    379 	memset((char *)spc + SPC_ZEROFF, 0, sizeof(*spc) - SPC_ZEROFF);
    380 
    381 	spcrelease(spc);
    382 }
    383 
    384 static unsigned
    385 serv_handleconn(int fd, connecthook_fn connhook, int busy)
    386 {
    387 	struct sockaddr_storage ss;
    388 	socklen_t sl = sizeof(ss);
    389 	int newfd, flags;
    390 	unsigned i;
    391 
    392 	/*LINTED: cast ok */
    393 	newfd = accept(fd, (struct sockaddr *)&ss, &sl);
    394 	if (newfd == -1)
    395 		return 0;
    396 
    397 	if (busy) {
    398 		close(newfd); /* EBUSY */
    399 		return 0;
    400 	}
    401 
    402 	/* XXX: should do some sort of handshake too */
    403 
    404 	flags = fcntl(newfd, F_GETFL, 0);
    405 	if (fcntl(newfd, F_SETFL, flags | O_NONBLOCK) == -1) {
    406 		close(newfd);
    407 		return 0;
    408 	}
    409 
    410 	if (connhook(newfd) != 0) {
    411 		close(newfd);
    412 		return 0;
    413 	}
    414 
    415 	/* find empty slot the simple way */
    416 	for (i = 0; i < MAXCLI; i++) {
    417 		if (pfdlist[i].fd == -1 && spclist[i].spc_dying == 0)
    418 			break;
    419 	}
    420 
    421 	if (lwproc_newproc(&spclist[i]) != 0) {
    422 		close(newfd);
    423 		return 0;
    424 	}
    425 
    426 	assert(i < MAXCLI);
    427 
    428 	pfdlist[i].fd = newfd;
    429 	spclist[i].spc_fd = newfd;
    430 	spclist[i].spc_mainlwp = lwproc_curlwp();
    431 	spclist[i].spc_istatus = SPCSTATUS_BUSY; /* dedicated receiver */
    432 	spclist[i].spc_pid = lwproc_getpid();
    433 	spclist[i].spc_refcnt = 1;
    434 
    435 	TAILQ_INIT(&spclist[i].spc_respwait);
    436 
    437 	DPRINTF(("rump_sp: added new connection fd %d at idx %u, pid %d\n",
    438 	    newfd, i, lwproc_getpid()));
    439 
    440 	lwproc_switch(NULL);
    441 
    442 	return i;
    443 }
    444 
    445 static void
    446 serv_handlesyscall(struct spclient *spc, struct rsp_hdr *rhdr, uint8_t *data)
    447 {
    448 	register_t retval[2] = {0, 0};
    449 	int rv, sysnum;
    450 
    451 	sysnum = (int)rhdr->rsp_sysnum;
    452 	DPRINTF(("rump_sp: handling syscall %d from client %d\n",
    453 	    sysnum, 0));
    454 
    455 	lwproc_newlwp(spc->spc_pid);
    456 	rv = rumpsyscall(sysnum, data, retval);
    457 	lwproc_release();
    458 
    459 	DPRINTF(("rump_sp: got return value %d & %d/%d\n",
    460 	    rv, retval[0], retval[1]));
    461 
    462 	send_syscall_resp(spc, rhdr->rsp_reqno, rv, retval);
    463 }
    464 
    465 struct sysbouncearg {
    466 	struct spclient *sba_spc;
    467 	struct rsp_hdr sba_hdr;
    468 	uint8_t *sba_data;
    469 
    470 	TAILQ_ENTRY(sysbouncearg) sba_entries;
    471 };
    472 static pthread_mutex_t sbamtx;
    473 static pthread_cond_t sbacv;
    474 static int nworker, idleworker;
    475 static TAILQ_HEAD(, sysbouncearg) syslist = TAILQ_HEAD_INITIALIZER(syslist);
    476 
    477 /*ARGSUSED*/
    478 static void *
    479 serv_syscallbouncer(void *arg)
    480 {
    481 	struct sysbouncearg *sba;
    482 
    483 	for (;;) {
    484 		pthread_mutex_lock(&sbamtx);
    485 		if (idleworker >= rumpsp_idleworker) {
    486 			nworker--;
    487 			pthread_mutex_unlock(&sbamtx);
    488 			break;
    489 		}
    490 		idleworker++;
    491 		while (TAILQ_EMPTY(&syslist)) {
    492 			pthread_cond_wait(&sbacv, &sbamtx);
    493 		}
    494 
    495 		sba = TAILQ_FIRST(&syslist);
    496 		TAILQ_REMOVE(&syslist, sba, sba_entries);
    497 		idleworker--;
    498 		pthread_mutex_unlock(&sbamtx);
    499 
    500 		serv_handlesyscall(sba->sba_spc,
    501 		    &sba->sba_hdr, sba->sba_data);
    502 		spcrelease(sba->sba_spc);
    503 		free(sba->sba_data);
    504 		free(sba);
    505 	}
    506 
    507 	return NULL;
    508 }
    509 
    510 static int
    511 sp_copyin(void *arg, const void *raddr, void *laddr, size_t *len, int wantstr)
    512 {
    513 	struct spclient *spc = arg;
    514 	void *rdata = NULL; /* XXXuninit */
    515 	int rv, nlocks;
    516 
    517 	rumpuser__kunlock(0, &nlocks, NULL);
    518 
    519 	rv = copyin_req(spc, raddr, len, wantstr, &rdata);
    520 	if (rv)
    521 		goto out;
    522 
    523 	memcpy(laddr, rdata, *len);
    524 	free(rdata);
    525 
    526  out:
    527 	rumpuser__klock(nlocks, NULL);
    528 	if (rv)
    529 		return EFAULT;
    530 	return 0;
    531 }
    532 
    533 int
    534 rumpuser_sp_copyin(void *arg, const void *raddr, void *laddr, size_t len)
    535 {
    536 
    537 	return sp_copyin(arg, raddr, laddr, &len, 0);
    538 }
    539 
    540 int
    541 rumpuser_sp_copyinstr(void *arg, const void *raddr, void *laddr, size_t *len)
    542 {
    543 
    544 	return sp_copyin(arg, raddr, laddr, len, 1);
    545 }
    546 
    547 static int
    548 sp_copyout(void *arg, const void *laddr, void *raddr, size_t dlen)
    549 {
    550 	struct spclient *spc = arg;
    551 	int nlocks, rv;
    552 
    553 	rumpuser__kunlock(0, &nlocks, NULL);
    554 	rv = send_copyout_req(spc, raddr, laddr, dlen);
    555 	rumpuser__klock(nlocks, NULL);
    556 
    557 	if (rv)
    558 		return EFAULT;
    559 	return 0;
    560 }
    561 
    562 int
    563 rumpuser_sp_copyout(void *arg, const void *laddr, void *raddr, size_t dlen)
    564 {
    565 
    566 	return sp_copyout(arg, laddr, raddr, dlen);
    567 }
    568 
    569 int
    570 rumpuser_sp_copyoutstr(void *arg, const void *laddr, void *raddr, size_t *dlen)
    571 {
    572 
    573 	return sp_copyout(arg, laddr, raddr, *dlen);
    574 }
    575 
    576 int
    577 rumpuser_sp_anonmmap(void *arg, size_t howmuch, void **addr)
    578 {
    579 	struct spclient *spc = arg;
    580 	void *resp, *rdata;
    581 	int nlocks, rv;
    582 
    583 	rumpuser__kunlock(0, &nlocks, NULL);
    584 
    585 	rv = anonmmap_req(spc, howmuch, &rdata);
    586 	if (rv) {
    587 		rv = EFAULT;
    588 		goto out;
    589 	}
    590 
    591 	resp = *(void **)rdata;
    592 	free(rdata);
    593 
    594 	if (resp == NULL) {
    595 		rv = ENOMEM;
    596 	}
    597 
    598 	*addr = resp;
    599 
    600  out:
    601 	rumpuser__klock(nlocks, NULL);
    602 
    603 	if (rv)
    604 		return rv;
    605 	return 0;
    606 }
    607 
    608 /*
    609  *
    610  * Startup routines and mainloop for server.
    611  *
    612  */
    613 
    614 struct spservarg {
    615 	int sps_sock;
    616 	connecthook_fn sps_connhook;
    617 };
    618 
    619 static pthread_attr_t pattr_detached;
    620 static void
    621 handlereq(struct spclient *spc)
    622 {
    623 	struct sysbouncearg *sba;
    624 	pthread_t pt;
    625 	int retries;
    626 
    627 	if (__predict_false(spc->spc_hdr.rsp_type != RUMPSP_SYSCALL)) {
    628 		send_error_resp(spc, spc->spc_hdr.rsp_reqno, EINVAL);
    629 		spcfreebuf(spc);
    630 		return;
    631 	}
    632 
    633 	retries = 0;
    634 	while ((sba = malloc(sizeof(*sba))) == NULL) {
    635 		if (nworker == 0 || retries > 10) {
    636 			send_error_resp(spc, spc->spc_hdr.rsp_reqno, EAGAIN);
    637 			spcfreebuf(spc);
    638 			return;
    639 		}
    640 		/* slim chance of more memory? */
    641 		usleep(10000);
    642 	}
    643 
    644 	sba->sba_spc = spc;
    645 	sba->sba_hdr = spc->spc_hdr;
    646 	sba->sba_data = spc->spc_buf;
    647 	spcresetbuf(spc);
    648 
    649 	spcref(spc);
    650 
    651 	pthread_mutex_lock(&sbamtx);
    652 	TAILQ_INSERT_TAIL(&syslist, sba, sba_entries);
    653 	if (idleworker > 0) {
    654 		/* do we have a daemon's tool (i.e. idle threads)? */
    655 		pthread_cond_signal(&sbacv);
    656 	} else if (nworker < rumpsp_maxworker) {
    657 		/*
    658 		 * Else, need to create one
    659 		 * (if we can, otherwise just expect another
    660 		 * worker to pick up the syscall)
    661 		 */
    662 		if (pthread_create(&pt, &pattr_detached,
    663 		    serv_syscallbouncer, NULL) == 0)
    664 			nworker++;
    665 	}
    666 	pthread_mutex_unlock(&sbamtx);
    667 }
    668 
    669 static void *
    670 spserver(void *arg)
    671 {
    672 	struct spservarg *sarg = arg;
    673 	struct spclient *spc;
    674 	unsigned idx;
    675 	int seen;
    676 	int rv;
    677 	unsigned int nfds, maxidx;
    678 
    679 	for (idx = 0; idx < MAXCLI; idx++) {
    680 		pfdlist[idx].fd = -1;
    681 		pfdlist[idx].events = POLLIN;
    682 
    683 		spc = &spclist[idx];
    684 		pthread_mutex_init(&spc->spc_mtx, NULL);
    685 		pthread_cond_init(&spc->spc_cv, NULL);
    686 	}
    687 	pfdlist[0].fd = sarg->sps_sock;
    688 	pfdlist[0].events = POLLIN;
    689 	nfds = 1;
    690 	maxidx = 0;
    691 
    692 	pthread_attr_init(&pattr_detached);
    693 	pthread_attr_setdetachstate(&pattr_detached, PTHREAD_CREATE_DETACHED);
    694 	/* XXX: doesn't stacksize currently work on NetBSD */
    695 	pthread_attr_setstacksize(&pattr_detached, 32*1024);
    696 
    697 	pthread_mutex_init(&sbamtx, NULL);
    698 	pthread_cond_init(&sbacv, NULL);
    699 
    700 	DPRINTF(("rump_sp: server mainloop\n"));
    701 
    702 	for (;;) {
    703 		int discoed;
    704 
    705 		/* g/c hangarounds (eventually) */
    706 		discoed = atomic_swap_uint(&disco, 0);
    707 		while (discoed--) {
    708 			nfds--;
    709 			idx = maxidx;
    710 			while (idx) {
    711 				if (pfdlist[idx].fd != -1) {
    712 					maxidx = idx;
    713 					break;
    714 				}
    715 				idx--;
    716 			}
    717 			DPRINTF(("rump_sp: set maxidx to [%u]\n",
    718 			    maxidx));
    719 		}
    720 
    721 		DPRINTF(("rump_sp: loop nfd %d\n", maxidx+1));
    722 		seen = 0;
    723 		rv = poll(pfdlist, maxidx+1, INFTIM);
    724 		assert(maxidx+1 <= MAXCLI);
    725 		assert(rv != 0);
    726 		if (rv == -1) {
    727 			if (errno == EINTR)
    728 				continue;
    729 			fprintf(stderr, "rump_spserver: poll returned %d\n",
    730 			    errno);
    731 			break;
    732 		}
    733 
    734 		for (idx = 0; seen < rv && idx < MAXCLI; idx++) {
    735 			if ((pfdlist[idx].revents & POLLIN) == 0)
    736 				continue;
    737 
    738 			seen++;
    739 			DPRINTF(("rump_sp: activity at [%u] %d/%d\n",
    740 			    idx, seen, rv));
    741 			if (idx > 0) {
    742 				spc = &spclist[idx];
    743 				DPRINTF(("rump_sp: mainloop read [%u]\n", idx));
    744 				switch (readframe(spc)) {
    745 				case 0:
    746 					break;
    747 				case -1:
    748 					serv_handledisco(idx);
    749 					break;
    750 				default:
    751 					switch (spc->spc_hdr.rsp_class) {
    752 					case RUMPSP_RESP:
    753 						kickwaiter(spc);
    754 						break;
    755 					case RUMPSP_REQ:
    756 						handlereq(spc);
    757 						break;
    758 					default:
    759 						send_error_resp(spc,
    760 						    spc->spc_hdr.rsp_reqno,
    761 						    ENOENT);
    762 						spcfreebuf(spc);
    763 						break;
    764 					}
    765 					break;
    766 				}
    767 
    768 			} else {
    769 				DPRINTF(("rump_sp: mainloop new connection\n"));
    770 
    771 				idx = serv_handleconn(pfdlist[0].fd,
    772 				    sarg->sps_connhook, nfds == MAXCLI);
    773 				if (idx)
    774 					nfds++;
    775 				if (idx > maxidx)
    776 					maxidx = idx;
    777 				DPRINTF(("rump_sp: maxid now %d\n", maxidx));
    778 			}
    779 		}
    780 	}
    781 
    782 	return NULL;
    783 }
    784 
    785 int
    786 rumpuser_sp_init(const struct rumpuser_sp_ops *spopsp, const char *url)
    787 {
    788 	pthread_t pt;
    789 	struct spservarg *sarg;
    790 	struct sockaddr *sap;
    791 	char *p;
    792 	unsigned idx;
    793 	int error, s;
    794 
    795 	p = strdup(url);
    796 	if (p == NULL)
    797 		return ENOMEM;
    798 	error = parseurl(p, &sap, &idx, 1);
    799 	free(p);
    800 	if (error)
    801 		return error;
    802 
    803 	s = socket(parsetab[idx].domain, SOCK_STREAM, 0);
    804 	if (s == -1)
    805 		return errno;
    806 
    807 	spops = *spopsp;
    808 	sarg = malloc(sizeof(*sarg));
    809 	if (sarg == NULL) {
    810 		close(s);
    811 		return ENOMEM;
    812 	}
    813 
    814 	sarg->sps_sock = s;
    815 	sarg->sps_connhook = parsetab[idx].connhook;
    816 
    817 	/* sloppy error recovery */
    818 
    819 	/*LINTED*/
    820 	if (bind(s, sap, sap->sa_len) == -1) {
    821 		fprintf(stderr, "rump_sp: server bind failed\n");
    822 		return errno;
    823 	}
    824 	if (listen(s, MAXCLI) == -1) {
    825 		fprintf(stderr, "rump_sp: server listen failed\n");
    826 		return errno;
    827 	}
    828 
    829 	if ((error = pthread_create(&pt, NULL, spserver, sarg)) != 0) {
    830 		fprintf(stderr, "rump_sp: cannot create wrkr thread\n");
    831 		return errno;
    832 	}
    833 	pthread_detach(pt);
    834 
    835 	return 0;
    836 }
    837