Home | History | Annotate | Line # | Download | only in libpuffs
puffs.c revision 1.87 
      1 /*	$NetBSD: puffs.c,v 1.87 2008/01/16 00:29:42 pooka Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 2005, 2006, 2007  Antti Kantee.  All Rights Reserved.
      5  *
      6  * Development of this software was supported by the
      7  * Google Summer of Code program and the Ulla Tuominen Foundation.
      8  * The Google SoC project was mentored by Bill Studenmund.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
     20  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     21  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
     22  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     25  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     29  * SUCH DAMAGE.
     30  */
     31 
     32 #include <sys/cdefs.h>
     33 #if !defined(lint)
     34 __RCSID("$NetBSD: puffs.c,v 1.87 2008/01/16 00:29:42 pooka Exp $");
     35 #endif /* !lint */
     36 
     37 #include <sys/param.h>
     38 #include <sys/mount.h>
     39 
     40 #include <assert.h>
     41 #include <err.h>
     42 #include <errno.h>
     43 #include <fcntl.h>
     44 #include <mntopts.h>
     45 #include <paths.h>
     46 #include <puffs.h>
     47 #include <stdio.h>
     48 #include <stdlib.h>
     49 #include <string.h>
     50 #include <syslog.h>
     51 #include <unistd.h>
     52 
     53 #include "puffs_priv.h"
     54 
     55 /* Most file systems want this for opts, so just give it to them */
     56 const struct mntopt puffsmopts[] = {
     57 	MOPT_STDOPTS,
     58 	PUFFSMOPT_STD,
     59 	MOPT_NULL,
     60 };
     61 
     62 #ifdef PUFFS_WITH_THREADS
     63 #include <pthread.h>
     64 pthread_mutex_t pu_lock = PTHREAD_MUTEX_INITIALIZER;
     65 #endif
     66 
     67 #define FILLOP(lower, upper)						\
     68 do {									\
     69 	if (pops->puffs_node_##lower)					\
     70 		opmask[PUFFS_VN_##upper] = 1;				\
     71 } while (/*CONSTCOND*/0)
     72 static void
     73 fillvnopmask(struct puffs_ops *pops, uint8_t *opmask)
     74 {
     75 
     76 	memset(opmask, 0, PUFFS_VN_MAX);
     77 
     78 	FILLOP(create,   CREATE);
     79 	FILLOP(mknod,    MKNOD);
     80 	FILLOP(open,     OPEN);
     81 	FILLOP(close,    CLOSE);
     82 	FILLOP(access,   ACCESS);
     83 	FILLOP(getattr,  GETATTR);
     84 	FILLOP(setattr,  SETATTR);
     85 	FILLOP(poll,     POLL); /* XXX: not ready in kernel */
     86 	FILLOP(mmap,     MMAP);
     87 	FILLOP(fsync,    FSYNC);
     88 	FILLOP(seek,     SEEK);
     89 	FILLOP(remove,   REMOVE);
     90 	FILLOP(link,     LINK);
     91 	FILLOP(rename,   RENAME);
     92 	FILLOP(mkdir,    MKDIR);
     93 	FILLOP(rmdir,    RMDIR);
     94 	FILLOP(symlink,  SYMLINK);
     95 	FILLOP(readdir,  READDIR);
     96 	FILLOP(readlink, READLINK);
     97 	FILLOP(reclaim,  RECLAIM);
     98 	FILLOP(inactive, INACTIVE);
     99 	FILLOP(print,    PRINT);
    100 	FILLOP(read,     READ);
    101 	FILLOP(write,    WRITE);
    102 }
    103 #undef FILLOP
    104 
    105 /*
    106  * Go over all framev entries and write everything we can.  This is
    107  * mostly for the benefit of delivering "unmount" to the kernel.
    108  */
    109 static void
    110 finalpush(struct puffs_usermount *pu)
    111 {
    112 	struct puffs_fctrl_io *fio;
    113 
    114 	LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    115 		if (fio->stat & FIO_WRGONE)
    116 			continue;
    117 
    118 		puffs_framev_output(pu, fio->fctrl, fio);
    119 	}
    120 }
    121 
    122 /*ARGSUSED*/
    123 static void
    124 puffs_defaulterror(struct puffs_usermount *pu, uint8_t type,
    125 	int error, const char *str, void *cookie)
    126 {
    127 
    128 	fprintf(stderr, "abort: type %d, error %d, cookie %p (%s)\n",
    129 	    type, error, cookie, str);
    130 	abort();
    131 }
    132 
    133 int
    134 puffs_getselectable(struct puffs_usermount *pu)
    135 {
    136 
    137 	return pu->pu_fd;
    138 }
    139 
    140 uint64_t
    141 puffs__nextreq(struct puffs_usermount *pu)
    142 {
    143 	uint64_t rv;
    144 
    145 	PU_LOCK();
    146 	rv = pu->pu_nextreq++;
    147 	PU_UNLOCK();
    148 
    149 	return rv;
    150 }
    151 
    152 int
    153 puffs_setblockingmode(struct puffs_usermount *pu, int mode)
    154 {
    155 	int rv, x;
    156 
    157 	assert(puffs_getstate(pu) == PUFFS_STATE_RUNNING);
    158 
    159 	if (mode != PUFFSDEV_BLOCK && mode != PUFFSDEV_NONBLOCK) {
    160 		errno = EINVAL;
    161 		return -1;
    162 	}
    163 
    164 	x = mode;
    165 	rv = ioctl(pu->pu_fd, FIONBIO, &x);
    166 
    167 	if (rv == 0) {
    168 		if (mode == PUFFSDEV_BLOCK)
    169 			pu->pu_state &= ~PU_ASYNCFD;
    170 		else
    171 			pu->pu_state |= PU_ASYNCFD;
    172 	}
    173 
    174 	return rv;
    175 }
    176 
    177 int
    178 puffs_getstate(struct puffs_usermount *pu)
    179 {
    180 
    181 	return pu->pu_state & PU_STATEMASK;
    182 }
    183 
    184 void
    185 puffs_setstacksize(struct puffs_usermount *pu, size_t ss)
    186 {
    187 	long psize, minsize;
    188 	int stackshift;
    189 	int bonus;
    190 
    191 	assert(puffs_getstate(pu) == PUFFS_STATE_BEFOREMOUNT);
    192 
    193 	psize = sysconf(_SC_PAGESIZE);
    194 	minsize = 4*psize;
    195 	if (ss < minsize || ss == PUFFS_STACKSIZE_MIN) {
    196 		if (ss != PUFFS_STACKSIZE_MIN)
    197 			fprintf(stderr, "puffs_setstacksize: adjusting "
    198 			    "stacksize to minimum %ld\n", minsize);
    199 		ss = 4*psize;
    200 	}
    201 
    202 	stackshift = -1;
    203 	bonus = 0;
    204 	while (ss) {
    205 		if (ss & 0x1)
    206 			bonus++;
    207 		ss >>= 1;
    208 		stackshift++;
    209 	}
    210 	if (bonus > 1) {
    211 		stackshift++;
    212 		fprintf(stderr, "puffs_setstacksize: using next power of two: "
    213 		    "%zu\n", 1<<stackshift);
    214 	}
    215 
    216 	pu->pu_cc_stackshift = stackshift;
    217 }
    218 
    219 struct puffs_pathobj *
    220 puffs_getrootpathobj(struct puffs_usermount *pu)
    221 {
    222 	struct puffs_node *pnr;
    223 
    224 	pnr = pu->pu_pn_root;
    225 	if (pnr == NULL) {
    226 		errno = ENOENT;
    227 		return NULL;
    228 	}
    229 
    230 	return &pnr->pn_po;
    231 }
    232 
    233 void
    234 puffs_setroot(struct puffs_usermount *pu, struct puffs_node *pn)
    235 {
    236 
    237 	pu->pu_pn_root = pn;
    238 }
    239 
    240 struct puffs_node *
    241 puffs_getroot(struct puffs_usermount *pu)
    242 {
    243 
    244 	return pu->pu_pn_root;
    245 }
    246 
    247 void
    248 puffs_setrootinfo(struct puffs_usermount *pu, enum vtype vt,
    249 	vsize_t vsize, dev_t rdev)
    250 {
    251 	struct puffs_kargs *pargs = pu->pu_kargp;
    252 
    253 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT) {
    254 		warnx("puffs_setrootinfo: call has effect only "
    255 		    "before mount\n");
    256 		return;
    257 	}
    258 
    259 	pargs->pa_root_vtype = vt;
    260 	pargs->pa_root_vsize = vsize;
    261 	pargs->pa_root_rdev = rdev;
    262 }
    263 
    264 void *
    265 puffs_getspecific(struct puffs_usermount *pu)
    266 {
    267 
    268 	return pu->pu_privdata;
    269 }
    270 
    271 size_t
    272 puffs_getmaxreqlen(struct puffs_usermount *pu)
    273 {
    274 
    275 	return pu->pu_maxreqlen;
    276 }
    277 
    278 void
    279 puffs_setmaxreqlen(struct puffs_usermount *pu, size_t reqlen)
    280 {
    281 
    282 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    283 		warnx("puffs_setmaxreqlen: call has effect only "
    284 		    "before mount\n");
    285 
    286 	pu->pu_kargp->pa_maxmsglen = reqlen;
    287 }
    288 
    289 void
    290 puffs_setfhsize(struct puffs_usermount *pu, size_t fhsize, int flags)
    291 {
    292 
    293 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    294 		warnx("puffs_setfhsize: call has effect only before mount\n");
    295 
    296 	pu->pu_kargp->pa_fhsize = fhsize;
    297 	pu->pu_kargp->pa_fhflags = flags;
    298 }
    299 
    300 void
    301 puffs_setncookiehash(struct puffs_usermount *pu, int nhash)
    302 {
    303 
    304 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    305 		warnx("puffs_setfhsize: call has effect only before mount\n");
    306 
    307 	pu->pu_kargp->pa_nhashbuckets = nhash;
    308 }
    309 
    310 void
    311 puffs_set_pathbuild(struct puffs_usermount *pu, pu_pathbuild_fn fn)
    312 {
    313 
    314 	pu->pu_pathbuild = fn;
    315 }
    316 
    317 void
    318 puffs_set_pathtransform(struct puffs_usermount *pu, pu_pathtransform_fn fn)
    319 {
    320 
    321 	pu->pu_pathtransform = fn;
    322 }
    323 
    324 void
    325 puffs_set_pathcmp(struct puffs_usermount *pu, pu_pathcmp_fn fn)
    326 {
    327 
    328 	pu->pu_pathcmp = fn;
    329 }
    330 
    331 void
    332 puffs_set_pathfree(struct puffs_usermount *pu, pu_pathfree_fn fn)
    333 {
    334 
    335 	pu->pu_pathfree = fn;
    336 }
    337 
    338 void
    339 puffs_set_namemod(struct puffs_usermount *pu, pu_namemod_fn fn)
    340 {
    341 
    342 	pu->pu_namemod = fn;
    343 }
    344 
    345 void
    346 puffs_set_errnotify(struct puffs_usermount *pu, pu_errnotify_fn fn)
    347 {
    348 
    349 	pu->pu_errnotify = fn;
    350 }
    351 
    352 void
    353 puffs_set_cmap(struct puffs_usermount *pu, pu_cmap_fn fn)
    354 {
    355 
    356 	pu->pu_cmap = fn;
    357 }
    358 
    359 void
    360 puffs_ml_setloopfn(struct puffs_usermount *pu, puffs_ml_loop_fn lfn)
    361 {
    362 
    363 	pu->pu_ml_lfn = lfn;
    364 }
    365 
    366 void
    367 puffs_ml_settimeout(struct puffs_usermount *pu, struct timespec *ts)
    368 {
    369 
    370 	if (ts == NULL) {
    371 		pu->pu_ml_timep = NULL;
    372 	} else {
    373 		pu->pu_ml_timeout = *ts;
    374 		pu->pu_ml_timep = &pu->pu_ml_timeout;
    375 	}
    376 }
    377 
    378 void
    379 puffs_set_prepost(struct puffs_usermount *pu,
    380 	pu_prepost_fn pre, pu_prepost_fn pst)
    381 {
    382 
    383 	pu->pu_oppre = pre;
    384 	pu->pu_oppost = pst;
    385 }
    386 
    387 void
    388 puffs_setback(struct puffs_cc *pcc, int whatback)
    389 {
    390 	struct puffs_req *preq = puffs__framebuf_getdataptr(pcc->pcc_pb);
    391 
    392 	assert(PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN && (
    393 	    preq->preq_optype == PUFFS_VN_OPEN ||
    394 	    preq->preq_optype == PUFFS_VN_MMAP ||
    395 	    preq->preq_optype == PUFFS_VN_REMOVE ||
    396 	    preq->preq_optype == PUFFS_VN_RMDIR ||
    397 	    preq->preq_optype == PUFFS_VN_INACTIVE));
    398 
    399 	preq->preq_setbacks |= whatback & PUFFS_SETBACK_MASK;
    400 }
    401 
    402 int
    403 puffs_daemon(struct puffs_usermount *pu, int nochdir, int noclose)
    404 {
    405 	ssize_t n;
    406 	int parent, value, fd;
    407 
    408 	if (pipe(pu->pu_dpipe) == -1)
    409 		return -1;
    410 
    411 	switch (fork()) {
    412 	case -1:
    413 		return -1;
    414 	case 0:
    415 		parent = 0;
    416 		break;
    417 	default:
    418 		parent = 1;
    419 		break;
    420 	}
    421 	pu->pu_state |= PU_PUFFSDAEMON;
    422 
    423 	if (parent) {
    424 		n = read(pu->pu_dpipe[0], &value, sizeof(int));
    425 		if (n == -1)
    426 			err(1, "puffs_daemon");
    427 		assert(n == sizeof(value));
    428 		if (value) {
    429 			errno = value;
    430 			err(1, "puffs_daemon");
    431 		}
    432 		exit(0);
    433 	} else {
    434 		if (setsid() == -1)
    435 			goto fail;
    436 
    437 		if (!nochdir)
    438 			chdir("/");
    439 
    440 		if (!noclose) {
    441 			fd = open(_PATH_DEVNULL, O_RDWR, 0);
    442 			if (fd == -1)
    443 				goto fail;
    444 			dup2(fd, STDIN_FILENO);
    445 			dup2(fd, STDOUT_FILENO);
    446 			dup2(fd, STDERR_FILENO);
    447 			if (fd > STDERR_FILENO)
    448 				close(fd);
    449 		}
    450 		return 0;
    451 	}
    452 
    453  fail:
    454 	n = write(pu->pu_dpipe[1], &errno, sizeof(int));
    455 	assert(n == 4);
    456 	return -1;
    457 }
    458 
    459 int
    460 puffs_mount(struct puffs_usermount *pu, const char *dir, int mntflags,
    461 	void *cookie)
    462 {
    463 	char rp[MAXPATHLEN];
    464 	ssize_t n;
    465 	int rv, fd, sverrno;
    466 	char *comfd;
    467 
    468 	pu->pu_kargp->pa_root_cookie = cookie;
    469 
    470 	/* XXXkludgehere */
    471 	/* kauth doesn't provide this service any longer */
    472 	if (geteuid() != 0)
    473 		mntflags |= MNT_NOSUID | MNT_NODEV;
    474 
    475 	if (realpath(dir, rp) == NULL) {
    476 		rv = -1;
    477 		goto out;
    478 	}
    479 
    480 	if (strcmp(dir, rp) != 0) {
    481 		warnx("puffs_mount: \"%s\" is a relative path.", dir);
    482 		warnx("puffs_mount: using \"%s\" instead.", rp);
    483 	}
    484 
    485 	/*
    486 	 * Undocumented...  Well, documented only here.
    487 	 *
    488 	 * This is used for imaginative purposes.  If the env variable is
    489 	 * set, puffs_mount() doesn't do the regular mount procedure.
    490 	 * Rather, it crams the mount data down the comfd and sets comfd as
    491 	 * the puffs descriptor.
    492 	 *
    493 	 * This shouldn't be used unless you can read my mind ( ... or write
    494 	 * it, not to mention execute it, but that's starting to get silly).
    495 	 */
    496 	if ((comfd = getenv("PUFFS_COMFD")) != NULL) {
    497 		size_t len;
    498 
    499 		if (sscanf(comfd, "%d", &pu->pu_fd) != 1) {
    500 			errno = EINVAL;
    501 			rv = -1;
    502 			goto out;
    503 		}
    504 		/* check that what we got at least resembles an fd */
    505 		if (fcntl(pu->pu_fd, F_GETFL) == -1) {
    506 			rv = -1;
    507 			goto out;
    508 		}
    509 
    510 		len = strlen(dir)+1;
    511 
    512 #define allwrite(buf, len)						\
    513 do {									\
    514 	ssize_t al_rv;							\
    515 	al_rv = write(pu->pu_fd, buf, len);				\
    516 	if (al_rv != len) {						\
    517 		if (al_rv != -1)					\
    518 			errno = EIO;					\
    519 		rv = -1;						\
    520 		abort();\
    521 		goto out;						\
    522 	}								\
    523 } while (/*CONSTCOND*/0)
    524 		allwrite(&len, sizeof(len));
    525 		allwrite(dir, len);
    526 		len = strlen(pu->pu_kargp->pa_mntfromname)+1;
    527 		allwrite(&len, sizeof(len));
    528 		allwrite(pu->pu_kargp->pa_mntfromname, len);
    529 		allwrite(&mntflags, sizeof(mntflags));
    530 		allwrite(pu->pu_kargp, sizeof(*pu->pu_kargp));
    531 		allwrite(&pu->pu_flags, sizeof(pu->pu_flags));
    532 #undef allwrite
    533 
    534 		rv = 0;
    535 	} else {
    536 		fd = open(_PATH_PUFFS, O_RDWR);
    537 		if (fd == -1) {
    538 			warnx("puffs_mount: cannot open %s", _PATH_PUFFS);
    539 			rv = -1;
    540 			goto out;
    541 		}
    542 		if (fd <= 2)
    543 			warnx("puffs_mount: device fd %d (<= 2), sure this is "
    544 			    "what you want?", fd);
    545 
    546 		pu->pu_kargp->pa_fd = pu->pu_fd = fd;
    547 		if ((rv = mount(MOUNT_PUFFS, rp, mntflags,
    548 		    pu->pu_kargp, sizeof(struct puffs_kargs))) == -1)
    549 			goto out;
    550 	}
    551 
    552 	PU_SETSTATE(pu, PUFFS_STATE_RUNNING);
    553 
    554  out:
    555 	if (rv != 0)
    556 		sverrno = errno;
    557 	else
    558 		sverrno = 0;
    559 	free(pu->pu_kargp);
    560 	pu->pu_kargp = NULL;
    561 
    562 	if (pu->pu_state & PU_PUFFSDAEMON) {
    563 		n = write(pu->pu_dpipe[1], &sverrno, sizeof(int));
    564 		assert(n == 4);
    565 		close(pu->pu_dpipe[0]);
    566 		close(pu->pu_dpipe[1]);
    567 	}
    568 
    569 	errno = sverrno;
    570 	return rv;
    571 }
    572 
    573 struct puffs_usermount *
    574 _puffs_init(int develv, struct puffs_ops *pops, const char *mntfromname,
    575 	const char *puffsname, void *priv, uint32_t pflags)
    576 {
    577 	struct puffs_usermount *pu;
    578 	struct puffs_kargs *pargs;
    579 	int sverrno;
    580 
    581 	if (develv != PUFFS_DEVEL_LIBVERSION) {
    582 		warnx("puffs_init: mounting with lib version %d, need %d",
    583 		    develv, PUFFS_DEVEL_LIBVERSION);
    584 		errno = EINVAL;
    585 		return NULL;
    586 	}
    587 
    588 	pu = malloc(sizeof(struct puffs_usermount));
    589 	if (pu == NULL)
    590 		goto failfree;
    591 	memset(pu, 0, sizeof(struct puffs_usermount));
    592 
    593 	pargs = pu->pu_kargp = malloc(sizeof(struct puffs_kargs));
    594 	if (pargs == NULL)
    595 		goto failfree;
    596 	memset(pargs, 0, sizeof(struct puffs_kargs));
    597 
    598 	pargs->pa_vers = PUFFSDEVELVERS | PUFFSVERSION;
    599 	pargs->pa_flags = PUFFS_FLAG_KERN(pflags);
    600 	fillvnopmask(pops, pargs->pa_vnopmask);
    601 	(void)strlcpy(pargs->pa_typename, puffsname,
    602 	    sizeof(pargs->pa_typename));
    603 	(void)strlcpy(pargs->pa_mntfromname, mntfromname,
    604 	    sizeof(pargs->pa_mntfromname));
    605 
    606 	puffs_zerostatvfs(&pargs->pa_svfsb);
    607 	pargs->pa_root_cookie = NULL;
    608 	pargs->pa_root_vtype = VDIR;
    609 	pargs->pa_root_vsize = 0;
    610 	pargs->pa_root_rdev = 0;
    611 	pargs->pa_maxmsglen = 0;
    612 
    613 	pu->pu_flags = pflags;
    614 	pu->pu_ops = *pops;
    615 	free(pops); /* XXX */
    616 
    617 	pu->pu_privdata = priv;
    618 	pu->pu_cc_stackshift = PUFFS_CC_STACKSHIFT_DEFAULT;
    619 	LIST_INIT(&pu->pu_pnodelst);
    620 	LIST_INIT(&pu->pu_ios);
    621 	LIST_INIT(&pu->pu_ios_rmlist);
    622 	LIST_INIT(&pu->pu_ccnukelst);
    623 	TAILQ_INIT(&pu->pu_sched);
    624 	TAILQ_INIT(&pu->pu_exq);
    625 
    626 	pu->pu_framectrl[PU_FRAMECTRL_FS].rfb = puffs_fsframe_read;
    627 	pu->pu_framectrl[PU_FRAMECTRL_FS].wfb = puffs_fsframe_write;
    628 	pu->pu_framectrl[PU_FRAMECTRL_FS].cmpfb = puffs_fsframe_cmp;
    629 	pu->pu_framectrl[PU_FRAMECTRL_FS].gotfb = puffs_fsframe_gotframe;
    630 	pu->pu_framectrl[PU_FRAMECTRL_FS].fdnotfn = puffs_framev_unmountonclose;
    631 
    632 	/* defaults for some user-settable translation functions */
    633 	pu->pu_cmap = NULL; /* identity translation */
    634 
    635 	pu->pu_pathbuild = puffs_stdpath_buildpath;
    636 	pu->pu_pathfree = puffs_stdpath_freepath;
    637 	pu->pu_pathcmp = puffs_stdpath_cmppath;
    638 	pu->pu_pathtransform = NULL;
    639 	pu->pu_namemod = NULL;
    640 
    641 	pu->pu_errnotify = puffs_defaulterror;
    642 
    643 	PU_SETSTATE(pu, PUFFS_STATE_BEFOREMOUNT);
    644 
    645 	return pu;
    646 
    647  failfree:
    648 	/* can't unmount() from here for obvious reasons */
    649 	sverrno = errno;
    650 	free(pu);
    651 	errno = sverrno;
    652 	return NULL;
    653 }
    654 
    655 /*
    656  * XXX: there's currently no clean way to request unmount from
    657  * within the user server, so be very brutal about it.
    658  */
    659 /*ARGSUSED1*/
    660 int
    661 puffs_exit(struct puffs_usermount *pu, int force)
    662 {
    663 	struct puffs_node *pn;
    664 
    665 	force = 1; /* currently */
    666 
    667 	if (pu->pu_fd)
    668 		close(pu->pu_fd);
    669 
    670 	while ((pn = LIST_FIRST(&pu->pu_pnodelst)) != NULL)
    671 		puffs_pn_put(pn);
    672 
    673 	finalpush(pu);
    674 	puffs_framev_exit(pu);
    675 	if (pu->pu_state & PU_HASKQ)
    676 		close(pu->pu_kq);
    677 	free(pu);
    678 
    679 	return 0; /* always succesful for now, WILL CHANGE */
    680 }
    681 
    682 int
    683 puffs_mainloop(struct puffs_usermount *pu)
    684 {
    685 	struct puffs_framectrl *pfctrl;
    686 	struct puffs_fctrl_io *fio;
    687 	struct puffs_cc *pcc;
    688 	struct kevent *curev;
    689 	size_t nchanges;
    690 	int ndone, sverrno;
    691 
    692 	assert(puffs_getstate(pu) >= PUFFS_STATE_RUNNING);
    693 
    694 	pu->pu_kq = kqueue();
    695 	if (pu->pu_kq == -1)
    696 		goto out;
    697 	pu->pu_state |= PU_HASKQ;
    698 
    699 	puffs_setblockingmode(pu, PUFFSDEV_NONBLOCK);
    700 	if (puffs__framev_addfd_ctrl(pu, puffs_getselectable(pu),
    701 	    PUFFS_FBIO_READ | PUFFS_FBIO_WRITE,
    702 	    &pu->pu_framectrl[PU_FRAMECTRL_FS]) == -1)
    703 		goto out;
    704 
    705 	curev = realloc(pu->pu_evs, (2*pu->pu_nfds)*sizeof(struct kevent));
    706 	if (curev == NULL)
    707 		goto out;
    708 	pu->pu_evs = curev;
    709 
    710 	LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    711 		EV_SET(curev, fio->io_fd, EVFILT_READ, EV_ADD,
    712 		    0, 0, (uintptr_t)fio);
    713 		curev++;
    714 		EV_SET(curev, fio->io_fd, EVFILT_WRITE, EV_ADD | EV_DISABLE,
    715 		    0, 0, (uintptr_t)fio);
    716 		curev++;
    717 	}
    718 	if (kevent(pu->pu_kq, pu->pu_evs, 2*pu->pu_nfds, NULL, 0, NULL) == -1)
    719 		goto out;
    720 
    721 	pu->pu_state |= PU_INLOOP;
    722 
    723 	while (puffs_getstate(pu) != PUFFS_STATE_UNMOUNTED) {
    724 		if (pu->pu_ml_lfn)
    725 			pu->pu_ml_lfn(pu);
    726 
    727 		/* XXX: can we still do these optimizations? */
    728 #if 0
    729 		/*
    730 		 * Do this here, because:
    731 		 *  a) loopfunc might generate some results
    732 		 *  b) it's still "after" event handling (except for round 1)
    733 		 */
    734 		if (puffs_req_putput(ppr) == -1)
    735 			goto out;
    736 		puffs_req_resetput(ppr);
    737 
    738 		/* micro optimization: skip kevent syscall if possible */
    739 		if (pu->pu_nfds == 1 && pu->pu_ml_timep == NULL
    740 		    && (pu->pu_state & PU_ASYNCFD) == 0) {
    741 			pfctrl = XXX->fctrl;
    742 			puffs_framev_input(pu, pfctrl, XXX);
    743 			continue;
    744 		}
    745 #endif
    746 
    747 		/* else: do full processing */
    748 		/* Don't bother worrying about O(n) for now */
    749 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    750 			if (fio->stat & FIO_WRGONE)
    751 				continue;
    752 
    753 			pfctrl = fio->fctrl;
    754 
    755 			/*
    756 			 * Try to write out everything to avoid the
    757 			 * need for enabling EVFILT_WRITE.  The likely
    758 			 * case is that we can fit everything into the
    759 			 * socket buffer.
    760 			 */
    761 			puffs_framev_output(pu, pfctrl, fio);
    762 		}
    763 
    764 		/*
    765 		 * Build list of which to enable/disable in writecheck.
    766 		 */
    767 		nchanges = 0;
    768 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    769 			if (fio->stat & FIO_WRGONE)
    770 				continue;
    771 
    772 			/* en/disable write checks for kqueue as needed */
    773 			assert((FIO_EN_WRITE(fio) && FIO_RM_WRITE(fio)) == 0);
    774 			if (FIO_EN_WRITE(fio)) {
    775 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    776 				    EVFILT_WRITE, EV_ENABLE, 0, 0,
    777 				    (uintptr_t)fio);
    778 				fio->stat |= FIO_WR;
    779 				nchanges++;
    780 			}
    781 			if (FIO_RM_WRITE(fio)) {
    782 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    783 				    EVFILT_WRITE, EV_DISABLE, 0, 0,
    784 				    (uintptr_t)fio);
    785 				fio->stat &= ~FIO_WR;
    786 				nchanges++;
    787 			}
    788 			assert(nchanges <= pu->pu_nfds);
    789 		}
    790 
    791 		ndone = kevent(pu->pu_kq, pu->pu_evs, nchanges,
    792 		    pu->pu_evs, 2*pu->pu_nfds, pu->pu_ml_timep);
    793 
    794 		if (ndone == -1) {
    795 			if (errno != EINTR)
    796 				goto out;
    797 			else
    798 				continue;
    799 		}
    800 
    801 		/* uoptimize */
    802 		if (ndone == 0)
    803 			continue;
    804 
    805 		/* iterate over the results */
    806 		for (curev = pu->pu_evs; ndone--; curev++) {
    807 			int what;
    808 
    809 #if 0
    810 			/* get & possibly dispatch events from kernel */
    811 			if (curev->ident == puffsfd) {
    812 				if (puffs_req_handle(pgr, ppr, 0) == -1)
    813 					goto out;
    814 				continue;
    815 			}
    816 #endif
    817 
    818 			fio = (void *)curev->udata;
    819 			pfctrl = fio->fctrl;
    820 			if (curev->flags & EV_ERROR) {
    821 				assert(curev->filter == EVFILT_WRITE);
    822 				fio->stat &= ~FIO_WR;
    823 
    824 				/* XXX: how to know if it's a transient error */
    825 				puffs_framev_writeclose(pu, fio,
    826 				    (int)curev->data);
    827 				puffs_framev_notify(fio, PUFFS_FBIO_ERROR);
    828 				continue;
    829 			}
    830 
    831 			what = 0;
    832 			if (curev->filter == EVFILT_READ) {
    833 				puffs_framev_input(pu, pfctrl, fio);
    834 				what |= PUFFS_FBIO_READ;
    835 			}
    836 
    837 			else if (curev->filter == EVFILT_WRITE) {
    838 				puffs_framev_output(pu, pfctrl, fio);
    839 				what |= PUFFS_FBIO_WRITE;
    840 			}
    841 			if (what)
    842 				puffs_framev_notify(fio, what);
    843 		}
    844 
    845 		/*
    846 		 * Schedule continuations.
    847 		 */
    848 		while ((pcc = TAILQ_FIRST(&pu->pu_sched)) != NULL) {
    849 			TAILQ_REMOVE(&pu->pu_sched, pcc, entries);
    850 			puffs_goto(pcc);
    851 		}
    852 
    853 		/*
    854 		 * Really free fd's now that we don't have references
    855 		 * to them.
    856 		 */
    857 		while ((fio = LIST_FIRST(&pu->pu_ios_rmlist)) != NULL) {
    858 			LIST_REMOVE(fio, fio_entries);
    859 			free(fio);
    860 		}
    861 	}
    862 	finalpush(pu);
    863 	errno = 0;
    864 
    865  out:
    866 	/* store the real error for a while */
    867 	sverrno = errno;
    868 
    869 	errno = sverrno;
    870 	if (errno)
    871 		return -1;
    872 	else
    873 		return 0;
    874 }
    875