Home | History | Annotate | Line # | Download | only in libpuffs
puffs.c revision 1.125
      1 /*	$NetBSD: puffs.c,v 1.125 2021/10/30 10:34:18 nia 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 1.125 2021/10/30 10:34:18 nia 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 <pthread.h>
     47 #include <puffs.h>
     48 #include <stdio.h>
     49 #include <stdlib.h>
     50 #include <string.h>
     51 #include <syslog.h>
     52 #include <unistd.h>
     53 
     54 #include "puffs_priv.h"
     55 
     56 /* Most file systems want this for opts, so just give it to them */
     57 const struct mntopt puffsmopts[] = {
     58 	MOPT_STDOPTS,
     59 	PUFFSMOPT_STD,
     60 	MOPT_NULL,
     61 };
     62 
     63 pthread_mutex_t pu_lock = PTHREAD_MUTEX_INITIALIZER;
     64 
     65 #define FILLOP(lower, upper)						\
     66 do {									\
     67 	if (pops->puffs_node_##lower)					\
     68 		opmask[PUFFS_VN_##upper] = 1;				\
     69 } while (/*CONSTCOND*/0)
     70 static void
     71 fillvnopmask(struct puffs_ops *pops, struct puffs_kargs *pa)
     72 {
     73 	uint8_t *opmask = pa->pa_vnopmask;
     74 
     75 	memset(opmask, 0, sizeof(pa->pa_vnopmask));
     76 
     77 	FILLOP(create,   CREATE);
     78 	FILLOP(mknod,    MKNOD);
     79 	FILLOP(open,     OPEN);
     80 	FILLOP(close,    CLOSE);
     81 	FILLOP(access,   ACCESS);
     82 	FILLOP(getattr,  GETATTR);
     83 	FILLOP(setattr,  SETATTR);
     84 	FILLOP(poll,     POLL);
     85 	FILLOP(mmap,     MMAP);
     86 	FILLOP(fsync,    FSYNC);
     87 	FILLOP(seek,     SEEK);
     88 	FILLOP(remove,   REMOVE);
     89 	FILLOP(link,     LINK);
     90 	FILLOP(rename,   RENAME);
     91 	FILLOP(mkdir,    MKDIR);
     92 	FILLOP(rmdir,    RMDIR);
     93 	FILLOP(symlink,  SYMLINK);
     94 	FILLOP(readdir,  READDIR);
     95 	FILLOP(readlink, READLINK);
     96 	FILLOP(reclaim,  RECLAIM);
     97 	FILLOP(inactive, INACTIVE);
     98 	FILLOP(print,    PRINT);
     99 	FILLOP(read,     READ);
    100 	FILLOP(write,    WRITE);
    101 	FILLOP(advlock,  ADVLOCK);
    102 	FILLOP(abortop,  ABORTOP);
    103 	FILLOP(pathconf, PATHCONF);
    104 
    105 	FILLOP(getextattr,  GETEXTATTR);
    106 	FILLOP(setextattr,  SETEXTATTR);
    107 	FILLOP(listextattr, LISTEXTATTR);
    108 	FILLOP(deleteextattr, DELETEEXTATTR);
    109 	FILLOP(fallocate, FALLOCATE);
    110 	FILLOP(fdiscard, FDISCARD);
    111 }
    112 #undef FILLOP
    113 
    114 /*
    115  * Go over all framev entries and write everything we can.  This is
    116  * mostly for the benefit of delivering "unmount" to the kernel.
    117  */
    118 static void
    119 finalpush(struct puffs_usermount *pu)
    120 {
    121 	struct puffs_fctrl_io *fio;
    122 
    123 	LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    124 		if (fio->stat & FIO_WRGONE)
    125 			continue;
    126 
    127 		puffs__framev_output(pu, fio->fctrl, fio);
    128 	}
    129 }
    130 
    131 /*ARGSUSED*/
    132 void
    133 puffs_kernerr_abort(struct puffs_usermount *pu, uint8_t type,
    134 	int error, const char *str, puffs_cookie_t cookie)
    135 {
    136 
    137 	warnx("%s: type %d, error %d, cookie %p (%s)", __func__,
    138 	    type, error, cookie, str);
    139 	abort();
    140 }
    141 
    142 /*ARGSUSED*/
    143 void
    144 puffs_kernerr_log(struct puffs_usermount *pu, uint8_t type,
    145 	int error, const char *str, puffs_cookie_t cookie)
    146 {
    147 
    148 	syslog(LOG_WARNING, "%s: type %d, error %d, cookie %p (%s)", __func__,
    149 	    type, error, cookie, str);
    150 }
    151 
    152 int
    153 puffs_getselectable(struct puffs_usermount *pu)
    154 {
    155 
    156 	return pu->pu_fd;
    157 }
    158 
    159 uint64_t
    160 puffs__nextreq(struct puffs_usermount *pu)
    161 {
    162 	uint64_t rv;
    163 
    164 	PU_LOCK();
    165 	rv = pu->pu_nextreq++ | (uint64_t)1<<63;
    166 	PU_UNLOCK();
    167 
    168 	return rv;
    169 }
    170 
    171 int
    172 puffs_setblockingmode(struct puffs_usermount *pu, int mode)
    173 {
    174 	int rv, x;
    175 
    176 	assert(puffs_getstate(pu) == PUFFS_STATE_RUNNING);
    177 
    178 	if (mode != PUFFSDEV_BLOCK && mode != PUFFSDEV_NONBLOCK) {
    179 		errno = EINVAL;
    180 		return -1;
    181 	}
    182 
    183 	x = mode;
    184 	rv = ioctl(pu->pu_fd, FIONBIO, &x);
    185 
    186 	if (rv == 0) {
    187 		if (mode == PUFFSDEV_BLOCK)
    188 			pu->pu_state &= ~PU_ASYNCFD;
    189 		else
    190 			pu->pu_state |= PU_ASYNCFD;
    191 	}
    192 
    193 	return rv;
    194 }
    195 
    196 int
    197 puffs_getstate(struct puffs_usermount *pu)
    198 {
    199 
    200 	return pu->pu_state & PU_STATEMASK;
    201 }
    202 
    203 void
    204 puffs_setstacksize(struct puffs_usermount *pu, size_t ss)
    205 {
    206 	long psize, minsize;
    207 	int stackshift;
    208 	int bonus;
    209 
    210 	assert(puffs_getstate(pu) == PUFFS_STATE_BEFOREMOUNT);
    211 
    212 	psize = sysconf(_SC_PAGESIZE);
    213 	minsize = 4*psize;
    214 	if (ss < (size_t)minsize || ss == PUFFS_STACKSIZE_MIN) {
    215 		if (ss != PUFFS_STACKSIZE_MIN)
    216 			warnx("%s: adjusting " "stacksize to minimum %ld",
    217 			    __func__, minsize);
    218 		ss = 4*psize;
    219 	}
    220 
    221 	stackshift = -1;
    222 	bonus = 0;
    223 	while (ss) {
    224 		if (ss & 0x1)
    225 			bonus++;
    226 		ss >>= 1;
    227 		stackshift++;
    228 	}
    229 	if (bonus > 1) {
    230 		stackshift++;
    231 		warnx("%s: using next power of two: %d", __func__,
    232 		    1 << stackshift);
    233 	}
    234 
    235 	pu->pu_cc_stackshift = stackshift;
    236 }
    237 
    238 struct puffs_pathobj *
    239 puffs_getrootpathobj(struct puffs_usermount *pu)
    240 {
    241 	struct puffs_node *pnr;
    242 
    243 	pnr = pu->pu_pn_root;
    244 	if (pnr == NULL) {
    245 		errno = ENOENT;
    246 		return NULL;
    247 	}
    248 
    249 	return &pnr->pn_po;
    250 }
    251 
    252 void
    253 puffs_setroot(struct puffs_usermount *pu, struct puffs_node *pn)
    254 {
    255 
    256 	pu->pu_pn_root = pn;
    257 }
    258 
    259 struct puffs_node *
    260 puffs_getroot(struct puffs_usermount *pu)
    261 {
    262 
    263 	return pu->pu_pn_root;
    264 }
    265 
    266 void
    267 puffs_setrootinfo(struct puffs_usermount *pu, enum vtype vt,
    268 	vsize_t vsize, dev_t rdev)
    269 {
    270 	struct puffs_kargs *pargs = pu->pu_kargp;
    271 
    272 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT) {
    273 		warnx("%s: call has effect only before mount", __func__);
    274 		return;
    275 	}
    276 
    277 	pargs->pa_root_vtype = vt;
    278 	pargs->pa_root_vsize = vsize;
    279 	pargs->pa_root_rdev = rdev;
    280 }
    281 
    282 void *
    283 puffs_getspecific(struct puffs_usermount *pu)
    284 {
    285 
    286 	return pu->pu_privdata;
    287 }
    288 
    289 void
    290 puffs_setspecific(struct puffs_usermount *pu, void *privdata)
    291 {
    292 
    293 	pu->pu_privdata = privdata;
    294 }
    295 
    296 void
    297 puffs_setmntinfo(struct puffs_usermount *pu,
    298 	const char *mntfromname, const char *puffsname)
    299 {
    300 	struct puffs_kargs *pargs = pu->pu_kargp;
    301 
    302 	(void)strlcpy(pargs->pa_mntfromname, mntfromname,
    303 	    sizeof(pargs->pa_mntfromname));
    304 	(void)strlcpy(pargs->pa_typename, puffsname,
    305 	    sizeof(pargs->pa_typename));
    306 }
    307 
    308 size_t
    309 puffs_getmaxreqlen(struct puffs_usermount *pu)
    310 {
    311 
    312 	return pu->pu_maxreqlen;
    313 }
    314 
    315 void
    316 puffs_setmaxreqlen(struct puffs_usermount *pu, size_t reqlen)
    317 {
    318 
    319 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    320 		warnx("%s: call has effect only before mount", __func__);
    321 
    322 	pu->pu_kargp->pa_maxmsglen = reqlen;
    323 }
    324 
    325 void
    326 puffs_setfhsize(struct puffs_usermount *pu, size_t fhsize, int flags)
    327 {
    328 
    329 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    330 		warnx("%s: call has effect only before mount", __func__);
    331 
    332 	pu->pu_kargp->pa_fhsize = fhsize;
    333 	pu->pu_kargp->pa_fhflags = flags;
    334 }
    335 
    336 void
    337 puffs_setncookiehash(struct puffs_usermount *pu, int nhash)
    338 {
    339 
    340 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    341 		warnx("%s: call has effect only before mount", __func__);
    342 
    343 	pu->pu_kargp->pa_nhashbuckets = nhash;
    344 }
    345 
    346 void
    347 puffs_set_pathbuild(struct puffs_usermount *pu, pu_pathbuild_fn fn)
    348 {
    349 
    350 	pu->pu_pathbuild = fn;
    351 }
    352 
    353 void
    354 puffs_set_pathtransform(struct puffs_usermount *pu, pu_pathtransform_fn fn)
    355 {
    356 
    357 	pu->pu_pathtransform = fn;
    358 }
    359 
    360 void
    361 puffs_set_pathcmp(struct puffs_usermount *pu, pu_pathcmp_fn fn)
    362 {
    363 
    364 	pu->pu_pathcmp = fn;
    365 }
    366 
    367 void
    368 puffs_set_pathfree(struct puffs_usermount *pu, pu_pathfree_fn fn)
    369 {
    370 
    371 	pu->pu_pathfree = fn;
    372 }
    373 
    374 void
    375 puffs_set_namemod(struct puffs_usermount *pu, pu_namemod_fn fn)
    376 {
    377 
    378 	pu->pu_namemod = fn;
    379 }
    380 
    381 void
    382 puffs_set_errnotify(struct puffs_usermount *pu, pu_errnotify_fn fn)
    383 {
    384 
    385 	pu->pu_errnotify = fn;
    386 }
    387 
    388 void
    389 puffs_set_cmap(struct puffs_usermount *pu, pu_cmap_fn fn)
    390 {
    391 
    392 	pu->pu_cmap = fn;
    393 }
    394 
    395 void
    396 puffs_ml_setloopfn(struct puffs_usermount *pu, puffs_ml_loop_fn lfn)
    397 {
    398 
    399 	pu->pu_ml_lfn = lfn;
    400 }
    401 
    402 void
    403 puffs_ml_settimeout(struct puffs_usermount *pu, struct timespec *ts)
    404 {
    405 
    406 	if (ts == NULL) {
    407 		pu->pu_ml_timep = NULL;
    408 	} else {
    409 		pu->pu_ml_timeout = *ts;
    410 		pu->pu_ml_timep = &pu->pu_ml_timeout;
    411 	}
    412 }
    413 
    414 void
    415 puffs_set_prepost(struct puffs_usermount *pu,
    416 	pu_prepost_fn pre, pu_prepost_fn pst)
    417 {
    418 
    419 	pu->pu_oppre = pre;
    420 	pu->pu_oppost = pst;
    421 }
    422 
    423 void
    424 puffs_setback(struct puffs_cc *pcc, int whatback)
    425 {
    426 	struct puffs_req *preq = puffs__framebuf_getdataptr(pcc->pcc_pb);
    427 
    428 	assert(PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN && (
    429 	    preq->preq_optype == PUFFS_VN_OPEN ||
    430 	    preq->preq_optype == PUFFS_VN_MMAP ||
    431 	    preq->preq_optype == PUFFS_VN_REMOVE ||
    432 	    preq->preq_optype == PUFFS_VN_RMDIR ||
    433 	    preq->preq_optype == PUFFS_VN_INACTIVE));
    434 
    435 	preq->preq_setbacks |= whatback & PUFFS_SETBACK_MASK;
    436 }
    437 
    438 int
    439 puffs_daemon(struct puffs_usermount *pu, int nochdir, int noclose)
    440 {
    441 	long int n;
    442 	int parent, value, fd;
    443 
    444 	if (pipe(pu->pu_dpipe) == -1)
    445 		return -1;
    446 
    447 	switch (fork()) {
    448 	case -1:
    449 		return -1;
    450 	case 0:
    451 		parent = 0;
    452 		break;
    453 	default:
    454 		parent = 1;
    455 		break;
    456 	}
    457 	pu->pu_state |= PU_PUFFSDAEMON;
    458 
    459 	if (parent) {
    460 		close(pu->pu_dpipe[1]);
    461 		n = read(pu->pu_dpipe[0], &value, sizeof(int));
    462 		if (n == -1)
    463 			err(1, "puffs_daemon");
    464 		if (n != sizeof(value))
    465 			errx(1, "puffs_daemon got %ld bytes", n);
    466 		if (value) {
    467 			errno = value;
    468 			err(1, "puffs_daemon");
    469 		}
    470 		exit(0);
    471 	} else {
    472 		if (setsid() == -1)
    473 			goto fail;
    474 
    475 		if (!nochdir)
    476 			chdir("/");
    477 
    478 		if (!noclose) {
    479 			fd = open(_PATH_DEVNULL, O_RDWR, 0);
    480 			if (fd == -1)
    481 				goto fail;
    482 			dup2(fd, STDIN_FILENO);
    483 			dup2(fd, STDOUT_FILENO);
    484 			dup2(fd, STDERR_FILENO);
    485 			if (fd > STDERR_FILENO)
    486 				close(fd);
    487 		}
    488 		return 0;
    489 	}
    490 
    491  fail:
    492 	n = write(pu->pu_dpipe[1], &errno, sizeof(int));
    493 	assert(n == 4);
    494 	return -1;
    495 }
    496 
    497 static void
    498 shutdaemon(struct puffs_usermount *pu, int error)
    499 {
    500 	ssize_t n;
    501 
    502 	n = write(pu->pu_dpipe[1], &error, sizeof(int));
    503 	assert(n == 4);
    504 	close(pu->pu_dpipe[0]);
    505 	close(pu->pu_dpipe[1]);
    506 	pu->pu_state &= ~PU_PUFFSDAEMON;
    507 }
    508 
    509 int
    510 puffs_mount(struct puffs_usermount *pu, const char *dir, int mntflags,
    511 	puffs_cookie_t cookie)
    512 {
    513 	int rv, fd, sverrno;
    514 	char *comfd;
    515 
    516 	pu->pu_kargp->pa_root_cookie = cookie;
    517 
    518 	/* XXXkludgehere */
    519 	/* kauth doesn't provide this service any longer */
    520 	if (geteuid() != 0)
    521 		mntflags |= MNT_NOSUID | MNT_NODEV;
    522 
    523 	/*
    524 	 * Undocumented...  Well, documented only here.
    525 	 *
    526 	 * This is used for imaginative purposes.  If the env variable is
    527 	 * set, puffs_mount() doesn't do the regular mount procedure.
    528 	 * Rather, it crams the mount data down the comfd and sets comfd as
    529 	 * the puffs descriptor.
    530 	 *
    531 	 * This shouldn't be used unless you can read my mind ( ... or write
    532 	 * it, not to mention execute it, but that's starting to get silly).
    533 	 */
    534 	if ((comfd = getenv("PUFFS_COMFD")) != NULL) {
    535 		size_t len;
    536 
    537 		if (sscanf(comfd, "%d", &pu->pu_fd) != 1) {
    538 			errno = EINVAL;
    539 			rv = -1;
    540 			goto out;
    541 		}
    542 		/* check that what we got at least resembles an fd */
    543 		if (fcntl(pu->pu_fd, F_GETFL) == -1) {
    544 			rv = -1;
    545 			goto out;
    546 		}
    547 
    548 #define allwrite(buf, len)						\
    549 do {									\
    550 	ssize_t al_rv;							\
    551 	al_rv = write(pu->pu_fd, buf, len);				\
    552 	if ((size_t)al_rv != len) {					\
    553 		if (al_rv != -1)					\
    554 			errno = EIO;					\
    555 		rv = -1;						\
    556 		goto out;						\
    557 	}								\
    558 } while (/*CONSTCOND*/0)
    559 		len = strlen(dir)+1;
    560 		allwrite(&len, sizeof(len));
    561 		allwrite(dir, len);
    562 		len = strlen(pu->pu_kargp->pa_mntfromname)+1;
    563 		allwrite(&len, sizeof(len));
    564 		allwrite(pu->pu_kargp->pa_mntfromname, len);
    565 		allwrite(&mntflags, sizeof(mntflags));
    566 		len = sizeof(*pu->pu_kargp);
    567 		allwrite(&len, sizeof(len));
    568 		allwrite(pu->pu_kargp, sizeof(*pu->pu_kargp));
    569 		allwrite(&pu->pu_flags, sizeof(pu->pu_flags));
    570 #undef allwrite
    571 
    572 		rv = 0;
    573 	} else {
    574 		char rp[MAXPATHLEN];
    575 		size_t rplen,dirlen;
    576 
    577 		if (realpath(dir, rp) == NULL) {
    578 			rv = -1;
    579 			goto out;
    580 		}
    581 
    582 		rplen = strlen(rp);
    583 		dirlen = strlen(dir);
    584 		if (strncmp(dir, rp, rplen) != 0 ||
    585 		    strspn(dir + rplen, "/") != dirlen - rplen) {
    586 			warnx("%s: `%s' is a %s path.", __func__, dir,
    587 			    dir[0] != '/' ? "relative" : "non canonical");
    588 			warnx("%s: using `%s' instead.", __func__, rp);
    589 		}
    590 
    591 		fd = open(_PATH_PUFFS, O_RDWR);
    592 		if (fd == -1) {
    593 			warnx("%s: cannot open `%s'", __func__, _PATH_PUFFS);
    594 			rv = -1;
    595 			goto out;
    596 		}
    597 		if (fd <= 2)
    598 			warnx("%s: device fd %d (<= 2), sure this is "
    599 			    "what you want?", __func__, fd);
    600 
    601 		pu->pu_kargp->pa_fd = pu->pu_fd = fd;
    602 		if ((rv = mount(MOUNT_PUFFS, rp, mntflags,
    603 		    pu->pu_kargp, sizeof(struct puffs_kargs))) == -1)
    604 			goto out;
    605 	}
    606 
    607 	PU_SETSTATE(pu, PUFFS_STATE_RUNNING);
    608 
    609  out:
    610 	if (rv != 0)
    611 		sverrno = errno;
    612 	else
    613 		sverrno = 0;
    614 	free(pu->pu_kargp);
    615 	pu->pu_kargp = NULL;
    616 
    617 	if (pu->pu_state & PU_PUFFSDAEMON)
    618 		shutdaemon(pu, sverrno);
    619 
    620 	errno = sverrno;
    621 	return rv;
    622 }
    623 
    624 struct puffs_usermount *
    625 puffs_init(struct puffs_ops *pops, const char *mntfromname,
    626 	const char *puffsname, void *priv, uint32_t pflags)
    627 {
    628 	struct puffs_usermount *pu;
    629 	struct puffs_kargs *pargs;
    630 	int sverrno;
    631 
    632 	if (puffsname == PUFFS_DEFER)
    633 		puffsname = "n/a";
    634 	if (mntfromname == PUFFS_DEFER)
    635 		mntfromname = "n/a";
    636 	if (priv == PUFFS_DEFER)
    637 		priv = NULL;
    638 
    639 	pu = malloc(sizeof(struct puffs_usermount));
    640 	if (pu == NULL)
    641 		goto failfree;
    642 	memset(pu, 0, sizeof(struct puffs_usermount));
    643 
    644 	pargs = pu->pu_kargp = malloc(sizeof(struct puffs_kargs));
    645 	if (pargs == NULL)
    646 		goto failfree;
    647 	memset(pargs, 0, sizeof(struct puffs_kargs));
    648 
    649 	pargs->pa_vers = PUFFSVERSION;
    650 	pargs->pa_flags = PUFFS_FLAG_KERN(pflags);
    651 	fillvnopmask(pops, pargs);
    652 	puffs_setmntinfo(pu, mntfromname, puffsname);
    653 
    654 	puffs_zerostatvfs(&pargs->pa_svfsb);
    655 	pargs->pa_root_cookie = NULL;
    656 	pargs->pa_root_vtype = VDIR;
    657 	pargs->pa_root_vsize = 0;
    658 	pargs->pa_root_rdev = 0;
    659 	pargs->pa_maxmsglen = 0;
    660 	if (/*CONSTCOND*/ sizeof(time_t) == 4)
    661 		pargs->pa_time32 = 1;
    662 	else
    663 		pargs->pa_time32 = 0;
    664 
    665 	pu->pu_flags = pflags;
    666 	pu->pu_ops = *pops;
    667 	free(pops); /* XXX */
    668 
    669 	pu->pu_privdata = priv;
    670 	pu->pu_cc_stackshift = PUFFS_CC_STACKSHIFT_DEFAULT;
    671 	LIST_INIT(&pu->pu_pnodelst);
    672 	LIST_INIT(&pu->pu_ios);
    673 	LIST_INIT(&pu->pu_ios_rmlist);
    674 	LIST_INIT(&pu->pu_ccmagazin);
    675 	TAILQ_INIT(&pu->pu_sched);
    676 
    677 	pu->pu_framectrl[PU_FRAMECTRL_FS].rfb = puffs__fsframe_read;
    678 	pu->pu_framectrl[PU_FRAMECTRL_FS].wfb = puffs__fsframe_write;
    679 	pu->pu_framectrl[PU_FRAMECTRL_FS].cmpfb = puffs__fsframe_cmp;
    680 	pu->pu_framectrl[PU_FRAMECTRL_FS].gotfb = puffs__fsframe_gotframe;
    681 	pu->pu_framectrl[PU_FRAMECTRL_FS].fdnotfn = puffs_framev_unmountonclose;
    682 
    683 	/* defaults for some user-settable translation functions */
    684 	pu->pu_cmap = NULL; /* identity translation */
    685 
    686 	pu->pu_pathbuild = puffs_stdpath_buildpath;
    687 	pu->pu_pathfree = puffs_stdpath_freepath;
    688 	pu->pu_pathcmp = puffs_stdpath_cmppath;
    689 	pu->pu_pathtransform = NULL;
    690 	pu->pu_namemod = NULL;
    691 
    692 	pu->pu_errnotify = puffs_kernerr_log;
    693 
    694 	PU_SETSTATE(pu, PUFFS_STATE_BEFOREMOUNT);
    695 
    696 	return pu;
    697 
    698  failfree:
    699 	/* can't unmount() from here for obvious reasons */
    700 	sverrno = errno;
    701 	free(pu);
    702 	errno = sverrno;
    703 	return NULL;
    704 }
    705 
    706 void
    707 puffs_cancel(struct puffs_usermount *pu, int error)
    708 {
    709 
    710 	assert(puffs_getstate(pu) < PUFFS_STATE_RUNNING);
    711 	shutdaemon(pu, error);
    712 	free(pu);
    713 }
    714 
    715 /*ARGSUSED1*/
    716 int
    717 puffs_exit(struct puffs_usermount *pu, int unused /* strict compat */)
    718 {
    719 	struct puffs_framebuf *pb;
    720 	struct puffs_req *preq;
    721 	void *winp;
    722 	size_t winlen;
    723 	int sverrno;
    724 
    725 	pb = puffs_framebuf_make();
    726 	if (pb == NULL) {
    727 		errno = ENOMEM;
    728 		return -1;
    729 	}
    730 
    731 	winlen = sizeof(struct puffs_req);
    732 	if (puffs_framebuf_getwindow(pb, 0, &winp, &winlen) == -1) {
    733 		sverrno = errno;
    734 		puffs_framebuf_destroy(pb);
    735 		errno = sverrno;
    736 		return -1;
    737 	}
    738 	preq = winp;
    739 
    740 	preq->preq_buflen = sizeof(struct puffs_req);
    741 	preq->preq_opclass = PUFFSOP_UNMOUNT;
    742 	preq->preq_id = puffs__nextreq(pu);
    743 
    744 	puffs_framev_enqueue_justsend(pu, puffs_getselectable(pu), pb, 1, 0);
    745 
    746 	return 0;
    747 }
    748 
    749 /* no sigset_t static initializer */
    750 static int sigs[NSIG] = { 0, };
    751 static int sigcatch = 0;
    752 
    753 int
    754 puffs_unmountonsignal(int sig, bool sigignore)
    755 {
    756 
    757 	if (sig < 0 || sig >= (int)NSIG) {
    758 		errno = EINVAL;
    759 		return -1;
    760 	}
    761 	if (sigignore)
    762 		if (signal(sig, SIG_IGN) == SIG_ERR)
    763 			return -1;
    764 
    765 	if (!sigs[sig])
    766 		sigcatch++;
    767 	sigs[sig] = 1;
    768 
    769 	return 0;
    770 }
    771 
    772 /*
    773  * Actual mainloop.  This is called from a context which can block.
    774  * It is called either from puffs_mainloop (indirectly, via
    775  * puffs_cc_continue() or from puffs_cc_yield()).
    776  */
    777 void
    778 puffs__theloop(struct puffs_cc *pcc)
    779 {
    780 	struct puffs_usermount *pu = pcc->pcc_pu;
    781 	struct puffs_framectrl *pfctrl;
    782 	struct puffs_fctrl_io *fio;
    783 	struct kevent *curev;
    784 	size_t nchanges;
    785 	int ndone;
    786 
    787 	while (puffs_getstate(pu) != PUFFS_STATE_UNMOUNTED) {
    788 
    789 		/*
    790 		 * Schedule existing requests.
    791 		 */
    792 		while ((pcc = TAILQ_FIRST(&pu->pu_sched)) != NULL) {
    793 			TAILQ_REMOVE(&pu->pu_sched, pcc, pcc_schedent);
    794 			puffs__goto(pcc);
    795 		}
    796 
    797 		if (pu->pu_ml_lfn)
    798 			pu->pu_ml_lfn(pu);
    799 
    800 		/* XXX: can we still do these optimizations? */
    801 #if 0
    802 		/*
    803 		 * Do this here, because:
    804 		 *  a) loopfunc might generate some results
    805 		 *  b) it's still "after" event handling (except for round 1)
    806 		 */
    807 		if (puffs_req_putput(ppr) == -1)
    808 			goto out;
    809 		puffs_req_resetput(ppr);
    810 
    811 		/* micro optimization: skip kevent syscall if possible */
    812 		if (pu->pu_nfds == 1 && pu->pu_ml_timep == NULL
    813 		    && (pu->pu_state & PU_ASYNCFD) == 0) {
    814 			pfctrl = XXX->fctrl;
    815 			puffs_framev_input(pu, pfctrl, XXX);
    816 			continue;
    817 		}
    818 #endif
    819 
    820 		/* else: do full processing */
    821 		/* Don't bother worrying about O(n) for now */
    822 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    823 			if (fio->stat & FIO_WRGONE)
    824 				continue;
    825 
    826 			pfctrl = fio->fctrl;
    827 
    828 			/*
    829 			 * Try to write out everything to avoid the
    830 			 * need for enabling EVFILT_WRITE.  The likely
    831 			 * case is that we can fit everything into the
    832 			 * socket buffer.
    833 			 */
    834 			puffs__framev_output(pu, pfctrl, fio);
    835 		}
    836 
    837 		/*
    838 		 * Build list of which to enable/disable in writecheck.
    839 		 */
    840 		nchanges = 0;
    841 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    842 			if (fio->stat & FIO_WRGONE)
    843 				continue;
    844 
    845 			/* en/disable write checks for kqueue as needed */
    846 			assert((FIO_EN_WRITE(fio) && FIO_RM_WRITE(fio)) == 0);
    847 			if (FIO_EN_WRITE(fio)) {
    848 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    849 				    EVFILT_WRITE, EV_ENABLE, 0, 0,
    850 				    (intptr_t)fio);
    851 				fio->stat |= FIO_WR;
    852 				nchanges++;
    853 			}
    854 			if (FIO_RM_WRITE(fio)) {
    855 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    856 				    EVFILT_WRITE, EV_DISABLE, 0, 0,
    857 				    (intptr_t)fio);
    858 				fio->stat &= ~FIO_WR;
    859 				nchanges++;
    860 			}
    861 		}
    862 
    863 		ndone = kevent(pu->pu_kq, pu->pu_evs, nchanges,
    864 		    pu->pu_evs, pu->pu_nevs, pu->pu_ml_timep);
    865 
    866 		if (ndone == -1) {
    867 			if (errno != EINTR)
    868 				break;
    869 			else
    870 				continue;
    871 		}
    872 
    873 		/* uoptimize */
    874 		if (ndone == 0)
    875 			continue;
    876 
    877 		/* iterate over the results */
    878 		for (curev = pu->pu_evs; ndone--; curev++) {
    879 			int what;
    880 
    881 #if 0
    882 			/* get & possibly dispatch events from kernel */
    883 			if (curev->ident == puffsfd) {
    884 				if (puffs_req_handle(pgr, ppr, 0) == -1)
    885 					goto out;
    886 				continue;
    887 			}
    888 #endif
    889 
    890 			fio = (void *)curev->udata;
    891 			if (__predict_true(fio))
    892 				pfctrl = fio->fctrl;
    893 			else
    894 				pfctrl = NULL;
    895 			if (curev->flags & EV_ERROR) {
    896 				assert(curev->filter == EVFILT_WRITE);
    897 				fio->stat &= ~FIO_WR;
    898 
    899 				/* XXX: how to know if it's a transient error */
    900 				puffs__framev_writeclose(pu, fio,
    901 				    (int)curev->data);
    902 				puffs__framev_notify(fio, PUFFS_FBIO_ERROR);
    903 				continue;
    904 			}
    905 
    906 			what = 0;
    907 			switch (curev->filter) {
    908 			case EVFILT_READ:
    909 				puffs__framev_input(pu, pfctrl, fio);
    910 				what |= PUFFS_FBIO_READ;
    911 				break;
    912 			case EVFILT_WRITE:
    913 				puffs__framev_output(pu, pfctrl, fio);
    914 				what |= PUFFS_FBIO_WRITE;
    915 				break;
    916 			case EVFILT_SIGNAL:
    917 				if ((pu->pu_state & PU_DONEXIT) == 0) {
    918 					PU_SETSFLAG(pu, PU_DONEXIT);
    919 					puffs_exit(pu, 0);
    920 				}
    921 				break;
    922 			default:
    923 				warn("unhandled filter %d", curev->filter);
    924 			}
    925 			if (what)
    926 				puffs__framev_notify(fio, what);
    927 		}
    928 
    929 		/*
    930 		 * Really free fd's now that we don't have references
    931 		 * to them.
    932 		 */
    933 		while ((fio = LIST_FIRST(&pu->pu_ios_rmlist)) != NULL) {
    934 			LIST_REMOVE(fio, fio_entries);
    935 			free(fio);
    936 		}
    937 	}
    938 
    939 	if (puffs__cc_restoremain(pu) == -1)
    940 		warn("cannot restore main context.  impending doom");
    941 }
    942 int
    943 puffs_mainloop(struct puffs_usermount *pu)
    944 {
    945 	struct puffs_fctrl_io *fio;
    946 	struct puffs_cc *pcc;
    947 	struct kevent *curev;
    948 	size_t nevs;
    949 	int sverrno, i;
    950 
    951 	assert(puffs_getstate(pu) >= PUFFS_STATE_RUNNING);
    952 
    953 	pu->pu_kq = kqueue();
    954 	if (pu->pu_kq == -1)
    955 		goto out;
    956 	pu->pu_state |= PU_HASKQ;
    957 
    958 	puffs_setblockingmode(pu, PUFFSDEV_NONBLOCK);
    959 	if (puffs__framev_addfd_ctrl(pu, puffs_getselectable(pu),
    960 	    PUFFS_FBIO_READ | PUFFS_FBIO_WRITE,
    961 	    &pu->pu_framectrl[PU_FRAMECTRL_FS]) == -1)
    962 		goto out;
    963 
    964 	nevs = pu->pu_nevs + sigcatch;
    965 	if (reallocarr(&pu->pu_evs, nevs, sizeof(struct kevent)) != 0) {
    966 		errno = ENOMEM;
    967 		goto out;
    968 	}
    969 	pu->pu_nevs = nevs;
    970 
    971 	curev = pu->pu_evs;
    972 
    973 	LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    974 		EV_SET(curev, fio->io_fd, EVFILT_READ, EV_ADD,
    975 		    0, 0, (intptr_t)fio);
    976 		curev++;
    977 		EV_SET(curev, fio->io_fd, EVFILT_WRITE, EV_ADD | EV_DISABLE,
    978 		    0, 0, (intptr_t)fio);
    979 		curev++;
    980 	}
    981 	for (i = 0; i < NSIG; i++) {
    982 		if (sigs[i]) {
    983 			EV_SET(curev, i, EVFILT_SIGNAL, EV_ADD | EV_ENABLE,
    984 			    0, 0, 0);
    985 			curev++;
    986 		}
    987 	}
    988 	assert(curev - pu->pu_evs == (ssize_t)pu->pu_nevs);
    989 	if (kevent(pu->pu_kq, pu->pu_evs, pu->pu_nevs, NULL, 0, NULL) == -1)
    990 		goto out;
    991 
    992 	pu->pu_state |= PU_INLOOP;
    993 
    994 	/*
    995 	 * Create alternate execution context and jump to it.  Note
    996 	 * that we come "out" of savemain twice.  Where we come out
    997 	 * of it depends on the architecture.  If the return address is
    998 	 * stored on the stack, we jump out from puffs_cc_continue(),
    999 	 * for a register return address from puffs__cc_savemain().
   1000 	 * PU_MAINRESTORE makes sure we DTRT in both cases.
   1001 	 */
   1002 	if (puffs__cc_create(pu, puffs__theloop, &pcc) == -1) {
   1003 		goto out;
   1004 	}
   1005 
   1006 #if 0
   1007 	if (puffs__cc_savemain(pu) == -1) {
   1008 		goto out;
   1009 	}
   1010 #else
   1011 	/*
   1012 	 * XXX
   1013 	 * puffs__cc_savemain() uses getcontext() and then returns.
   1014 	 * the caller (this function) may overwrite the stack frame
   1015 	 * of puffs__cc_savemain(), so when we call setcontext() later and
   1016 	 * return from puffs__cc_savemain() again, the return address or
   1017 	 * saved stack pointer can be garbage.
   1018 	 * avoid this by calling getcontext() directly here.
   1019 	 */
   1020 	extern int puffs_fakecc;
   1021 	if (!puffs_fakecc) {
   1022 		PU_CLRSFLAG(pu, PU_MAINRESTORE);
   1023 		if (getcontext(&pu->pu_mainctx) == -1) {
   1024 			goto out;
   1025 		}
   1026 	}
   1027 #endif
   1028 
   1029 	if ((pu->pu_state & PU_MAINRESTORE) == 0)
   1030 		puffs_cc_continue(pcc);
   1031 
   1032 	finalpush(pu);
   1033 	errno = 0;
   1034 
   1035  out:
   1036 	/* store the real error for a while */
   1037 	sverrno = errno;
   1038 
   1039 	errno = sverrno;
   1040 	if (errno)
   1041 		return -1;
   1042 	else
   1043 		return 0;
   1044 }
   1045