Home | History | Annotate | Line # | Download | only in libpuffs
puffs.c revision 1.80 
      1 /*	$NetBSD: puffs.c,v 1.80 2007/12/05 12:11:56 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 1.80 2007/12/05 12:11:56 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;
    188 	int stackshift;
    189 
    190 	psize = sysconf(_SC_PAGESIZE);
    191 	if (ss < 2*psize) {
    192 		ss = 2*psize;
    193 		fprintf(stderr, "puffs_setstacksize: adjusting stacksize "
    194 		    "to minimum %zu\n", ss);
    195 	}
    196 
    197 	assert(puffs_getstate(pu) == PUFFS_STATE_BEFOREMOUNT);
    198 	stackshift = -1;
    199 	while (ss) {
    200 		ss >>= 1;
    201 		stackshift++;
    202 	}
    203 	pu->pu_cc_stackshift = stackshift;
    204 	assert(1<<stackshift == ss);
    205 }
    206 
    207 struct puffs_pathobj *
    208 puffs_getrootpathobj(struct puffs_usermount *pu)
    209 {
    210 	struct puffs_node *pnr;
    211 
    212 	pnr = pu->pu_pn_root;
    213 	if (pnr == NULL) {
    214 		errno = ENOENT;
    215 		return NULL;
    216 	}
    217 
    218 	return &pnr->pn_po;
    219 }
    220 
    221 void
    222 puffs_setroot(struct puffs_usermount *pu, struct puffs_node *pn)
    223 {
    224 
    225 	pu->pu_pn_root = pn;
    226 }
    227 
    228 struct puffs_node *
    229 puffs_getroot(struct puffs_usermount *pu)
    230 {
    231 
    232 	return pu->pu_pn_root;
    233 }
    234 
    235 void
    236 puffs_setrootinfo(struct puffs_usermount *pu, enum vtype vt,
    237 	vsize_t vsize, dev_t rdev)
    238 {
    239 	struct puffs_kargs *pargs = pu->pu_kargp;
    240 
    241 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT) {
    242 		warnx("puffs_setrootinfo: call has effect only "
    243 		    "before mount\n");
    244 		return;
    245 	}
    246 
    247 	pargs->pa_root_vtype = vt;
    248 	pargs->pa_root_vsize = vsize;
    249 	pargs->pa_root_rdev = rdev;
    250 }
    251 
    252 void *
    253 puffs_getspecific(struct puffs_usermount *pu)
    254 {
    255 
    256 	return pu->pu_privdata;
    257 }
    258 
    259 size_t
    260 puffs_getmaxreqlen(struct puffs_usermount *pu)
    261 {
    262 
    263 	return pu->pu_maxreqlen;
    264 }
    265 
    266 void
    267 puffs_setmaxreqlen(struct puffs_usermount *pu, size_t reqlen)
    268 {
    269 
    270 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    271 		warnx("puffs_setmaxreqlen: call has effect only "
    272 		    "before mount\n");
    273 
    274 	pu->pu_kargp->pa_maxmsglen = reqlen;
    275 }
    276 
    277 void
    278 puffs_setfhsize(struct puffs_usermount *pu, size_t fhsize, int flags)
    279 {
    280 
    281 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    282 		warnx("puffs_setfhsize: call has effect only before mount\n");
    283 
    284 	pu->pu_kargp->pa_fhsize = fhsize;
    285 	pu->pu_kargp->pa_fhflags = flags;
    286 }
    287 
    288 void
    289 puffs_setncookiehash(struct puffs_usermount *pu, int nhash)
    290 {
    291 
    292 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    293 		warnx("puffs_setfhsize: call has effect only before mount\n");
    294 
    295 	pu->pu_kargp->pa_nhashbuckets = nhash;
    296 }
    297 
    298 void
    299 puffs_set_pathbuild(struct puffs_usermount *pu, pu_pathbuild_fn fn)
    300 {
    301 
    302 	pu->pu_pathbuild = fn;
    303 }
    304 
    305 void
    306 puffs_set_pathtransform(struct puffs_usermount *pu, pu_pathtransform_fn fn)
    307 {
    308 
    309 	pu->pu_pathtransform = fn;
    310 }
    311 
    312 void
    313 puffs_set_pathcmp(struct puffs_usermount *pu, pu_pathcmp_fn fn)
    314 {
    315 
    316 	pu->pu_pathcmp = fn;
    317 }
    318 
    319 void
    320 puffs_set_pathfree(struct puffs_usermount *pu, pu_pathfree_fn fn)
    321 {
    322 
    323 	pu->pu_pathfree = fn;
    324 }
    325 
    326 void
    327 puffs_set_namemod(struct puffs_usermount *pu, pu_namemod_fn fn)
    328 {
    329 
    330 	pu->pu_namemod = fn;
    331 }
    332 
    333 void
    334 puffs_set_errnotify(struct puffs_usermount *pu, pu_errnotify_fn fn)
    335 {
    336 
    337 	pu->pu_errnotify = fn;
    338 }
    339 
    340 void
    341 puffs_ml_setloopfn(struct puffs_usermount *pu, puffs_ml_loop_fn lfn)
    342 {
    343 
    344 	pu->pu_ml_lfn = lfn;
    345 }
    346 
    347 void
    348 puffs_ml_settimeout(struct puffs_usermount *pu, struct timespec *ts)
    349 {
    350 
    351 	if (ts == NULL) {
    352 		pu->pu_ml_timep = NULL;
    353 	} else {
    354 		pu->pu_ml_timeout = *ts;
    355 		pu->pu_ml_timep = &pu->pu_ml_timeout;
    356 	}
    357 }
    358 
    359 void
    360 puffs_set_prepost(struct puffs_usermount *pu,
    361 	pu_prepost_fn pre, pu_prepost_fn pst)
    362 {
    363 
    364 	pu->pu_oppre = pre;
    365 	pu->pu_oppost = pst;
    366 }
    367 
    368 void
    369 puffs_setback(struct puffs_cc *pcc, int whatback)
    370 {
    371 	struct puffs_req *preq = puffs__framebuf_getdataptr(pcc->pcc_pb);
    372 
    373 	assert(PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN && (
    374 	    preq->preq_optype == PUFFS_VN_OPEN ||
    375 	    preq->preq_optype == PUFFS_VN_MMAP ||
    376 	    preq->preq_optype == PUFFS_VN_REMOVE ||
    377 	    preq->preq_optype == PUFFS_VN_RMDIR ||
    378 	    preq->preq_optype == PUFFS_VN_INACTIVE));
    379 
    380 	preq->preq_setbacks |= whatback & PUFFS_SETBACK_MASK;
    381 }
    382 
    383 int
    384 puffs_daemon(struct puffs_usermount *pu, int nochdir, int noclose)
    385 {
    386 	ssize_t n;
    387 	int parent, value, fd;
    388 
    389 	if (pipe(pu->pu_dpipe) == -1)
    390 		return -1;
    391 
    392 	switch (fork()) {
    393 	case -1:
    394 		return -1;
    395 	case 0:
    396 		parent = 0;
    397 		break;
    398 	default:
    399 		parent = 1;
    400 		break;
    401 	}
    402 	pu->pu_state |= PU_PUFFSDAEMON;
    403 
    404 	if (parent) {
    405 		n = read(pu->pu_dpipe[0], &value, sizeof(int));
    406 		if (n == -1)
    407 			err(1, "puffs_daemon");
    408 		assert(n == sizeof(value));
    409 		if (value) {
    410 			errno = value;
    411 			err(1, "puffs_daemon");
    412 		}
    413 		exit(0);
    414 	} else {
    415 		if (setsid() == -1)
    416 			goto fail;
    417 
    418 		if (!nochdir)
    419 			chdir("/");
    420 
    421 		if (!noclose) {
    422 			fd = open(_PATH_DEVNULL, O_RDWR, 0);
    423 			if (fd == -1)
    424 				goto fail;
    425 			dup2(fd, STDIN_FILENO);
    426 			dup2(fd, STDOUT_FILENO);
    427 			dup2(fd, STDERR_FILENO);
    428 			if (fd > STDERR_FILENO)
    429 				close(fd);
    430 		}
    431 		return 0;
    432 	}
    433 
    434  fail:
    435 	n = write(pu->pu_dpipe[1], &errno, sizeof(int));
    436 	assert(n == 4);
    437 	return -1;
    438 }
    439 
    440 int
    441 puffs_mount(struct puffs_usermount *pu, const char *dir, int mntflags,
    442 	void *cookie)
    443 {
    444 	char rp[MAXPATHLEN];
    445 	ssize_t n;
    446 	int rv, fd, sverrno;
    447 
    448 #if 1
    449 	/* XXXkludgehere */
    450 	/* kauth doesn't provide this service any longer */
    451 	if (geteuid() != 0)
    452 		mntflags |= MNT_NOSUID | MNT_NODEV;
    453 #endif
    454 
    455 	if (realpath(dir, rp) == NULL) {
    456 		rv = -1;
    457 		goto out;
    458 	}
    459 
    460 	if (strcmp(dir, rp) != 0) {
    461 		warnx("puffs_mount: \"%s\" is a relative path.", dir);
    462 		warnx("puffs_mount: using \"%s\" instead.", rp);
    463 	}
    464 
    465 	fd = open(_PATH_PUFFS, O_RDWR);
    466 	if (fd == -1) {
    467 		warnx("puffs_mount: cannot open %s", _PATH_PUFFS);
    468 		rv = -1;
    469 		goto out;
    470 	}
    471 	if (fd <= 2)
    472 		warnx("puffs_init: device fd %d (<= 2), sure this is "
    473 		    "what you want?", fd);
    474 
    475 	pu->pu_kargp->pa_fd = pu->pu_fd = fd;
    476 	pu->pu_kargp->pa_root_cookie = cookie;
    477 	if ((rv = mount(MOUNT_PUFFS, rp, mntflags,
    478 	    pu->pu_kargp, sizeof(struct puffs_kargs))) == -1)
    479 		goto out;
    480 #if 0
    481 	if ((rv = ioctl(pu->pu_fd, PUFFSREQSIZEOP, &pu->pu_maxreqlen)) == -1)
    482 		goto out;
    483 #endif
    484 	PU_SETSTATE(pu, PUFFS_STATE_RUNNING);
    485 
    486  out:
    487 	if (rv != 0)
    488 		sverrno = errno;
    489 	else
    490 		sverrno = 0;
    491 	free(pu->pu_kargp);
    492 	pu->pu_kargp = NULL;
    493 
    494 	if (pu->pu_state & PU_PUFFSDAEMON) {
    495 		n = write(pu->pu_dpipe[1], &sverrno, sizeof(int));
    496 		assert(n == 4);
    497 		close(pu->pu_dpipe[0]);
    498 		close(pu->pu_dpipe[1]);
    499 	}
    500 
    501 	errno = sverrno;
    502 	return rv;
    503 }
    504 
    505 struct puffs_usermount *
    506 _puffs_init(int develv, struct puffs_ops *pops, const char *mntfromname,
    507 	const char *puffsname, void *priv, uint32_t pflags)
    508 {
    509 	struct puffs_usermount *pu;
    510 	struct puffs_kargs *pargs;
    511 	int sverrno;
    512 
    513 	if (develv != PUFFS_DEVEL_LIBVERSION) {
    514 		warnx("puffs_init: mounting with lib version %d, need %d",
    515 		    develv, PUFFS_DEVEL_LIBVERSION);
    516 		errno = EINVAL;
    517 		return NULL;
    518 	}
    519 
    520 	pu = malloc(sizeof(struct puffs_usermount));
    521 	if (pu == NULL)
    522 		goto failfree;
    523 	memset(pu, 0, sizeof(struct puffs_usermount));
    524 
    525 	pargs = pu->pu_kargp = malloc(sizeof(struct puffs_kargs));
    526 	if (pargs == NULL)
    527 		goto failfree;
    528 	memset(pargs, 0, sizeof(struct puffs_kargs));
    529 
    530 	pargs->pa_vers = PUFFSDEVELVERS | PUFFSVERSION;
    531 	pargs->pa_flags = PUFFS_FLAG_KERN(pflags);
    532 	fillvnopmask(pops, pargs->pa_vnopmask);
    533 	(void)strlcpy(pargs->pa_typename, puffsname,
    534 	    sizeof(pargs->pa_typename));
    535 	(void)strlcpy(pargs->pa_mntfromname, mntfromname,
    536 	    sizeof(pargs->pa_mntfromname));
    537 
    538 	puffs_zerostatvfs(&pargs->pa_svfsb);
    539 	pargs->pa_root_cookie = NULL;
    540 	pargs->pa_root_vtype = VDIR;
    541 	pargs->pa_root_vsize = 0;
    542 	pargs->pa_root_rdev = 0;
    543 	pargs->pa_maxmsglen = 0;
    544 
    545 	pu->pu_flags = pflags;
    546 	pu->pu_ops = *pops;
    547 	free(pops); /* XXX */
    548 
    549 	pu->pu_privdata = priv;
    550 	pu->pu_cc_stackshift = PUFFS_CC_STACKSHIFT_DEFAULT;
    551 	LIST_INIT(&pu->pu_pnodelst);
    552 	LIST_INIT(&pu->pu_ios);
    553 	LIST_INIT(&pu->pu_ios_rmlist);
    554 	LIST_INIT(&pu->pu_ccnukelst);
    555 	TAILQ_INIT(&pu->pu_sched);
    556 	TAILQ_INIT(&pu->pu_exq);
    557 
    558 	pu->pu_framectrl[PU_FRAMECTRL_FS].rfb = puffs_fsframe_read;
    559 	pu->pu_framectrl[PU_FRAMECTRL_FS].wfb = puffs_fsframe_write;
    560 	pu->pu_framectrl[PU_FRAMECTRL_FS].cmpfb = puffs_fsframe_cmp;
    561 	pu->pu_framectrl[PU_FRAMECTRL_FS].gotfb = puffs_fsframe_gotframe;
    562 	pu->pu_framectrl[PU_FRAMECTRL_FS].fdnotfn = puffs_framev_unmountonclose;
    563 
    564 	/* defaults for some user-settable translation functions */
    565 	pu->pu_cmap = NULL; /* identity translation */
    566 
    567 	pu->pu_pathbuild = puffs_stdpath_buildpath;
    568 	pu->pu_pathfree = puffs_stdpath_freepath;
    569 	pu->pu_pathcmp = puffs_stdpath_cmppath;
    570 	pu->pu_pathtransform = NULL;
    571 	pu->pu_namemod = NULL;
    572 
    573 	pu->pu_errnotify = puffs_defaulterror;
    574 
    575 	PU_SETSTATE(pu, PUFFS_STATE_BEFOREMOUNT);
    576 
    577 	return pu;
    578 
    579  failfree:
    580 	/* can't unmount() from here for obvious reasons */
    581 	sverrno = errno;
    582 	free(pu);
    583 	errno = sverrno;
    584 	return NULL;
    585 }
    586 
    587 /*
    588  * XXX: there's currently no clean way to request unmount from
    589  * within the user server, so be very brutal about it.
    590  */
    591 /*ARGSUSED1*/
    592 int
    593 puffs_exit(struct puffs_usermount *pu, int force)
    594 {
    595 	struct puffs_node *pn;
    596 
    597 	force = 1; /* currently */
    598 
    599 	if (pu->pu_fd)
    600 		close(pu->pu_fd);
    601 
    602 	while ((pn = LIST_FIRST(&pu->pu_pnodelst)) != NULL)
    603 		puffs_pn_put(pn);
    604 
    605 	finalpush(pu);
    606 	puffs_framev_exit(pu);
    607 	if (pu->pu_state & PU_HASKQ)
    608 		close(pu->pu_kq);
    609 	free(pu);
    610 
    611 	return 0; /* always succesful for now, WILL CHANGE */
    612 }
    613 
    614 int
    615 puffs_mainloop(struct puffs_usermount *pu)
    616 {
    617 	struct puffs_framectrl *pfctrl;
    618 	struct puffs_fctrl_io *fio;
    619 	struct puffs_cc *pcc;
    620 	struct kevent *curev;
    621 	size_t nchanges;
    622 	int ndone, sverrno;
    623 
    624 	assert(puffs_getstate(pu) >= PUFFS_STATE_RUNNING);
    625 
    626 	pu->pu_kq = kqueue();
    627 	if (pu->pu_kq == -1)
    628 		goto out;
    629 	pu->pu_state |= PU_HASKQ;
    630 
    631 	puffs_setblockingmode(pu, PUFFSDEV_NONBLOCK);
    632 	if (puffs__framev_addfd_ctrl(pu, puffs_getselectable(pu),
    633 	    PUFFS_FBIO_READ | PUFFS_FBIO_WRITE,
    634 	    &pu->pu_framectrl[PU_FRAMECTRL_FS]) == -1)
    635 		goto out;
    636 
    637 	curev = realloc(pu->pu_evs, (2*pu->pu_nfds)*sizeof(struct kevent));
    638 	if (curev == NULL)
    639 		goto out;
    640 	pu->pu_evs = curev;
    641 
    642 	LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    643 		EV_SET(curev, fio->io_fd, EVFILT_READ, EV_ADD,
    644 		    0, 0, (uintptr_t)fio);
    645 		curev++;
    646 		EV_SET(curev, fio->io_fd, EVFILT_WRITE, EV_ADD | EV_DISABLE,
    647 		    0, 0, (uintptr_t)fio);
    648 		curev++;
    649 	}
    650 	if (kevent(pu->pu_kq, pu->pu_evs, 2*pu->pu_nfds, NULL, 0, NULL) == -1)
    651 		goto out;
    652 
    653 	pu->pu_state |= PU_INLOOP;
    654 
    655 	while (puffs_getstate(pu) != PUFFS_STATE_UNMOUNTED) {
    656 		if (pu->pu_ml_lfn)
    657 			pu->pu_ml_lfn(pu);
    658 
    659 		/* XXX: can we still do these optimizations? */
    660 #if 0
    661 		/*
    662 		 * Do this here, because:
    663 		 *  a) loopfunc might generate some results
    664 		 *  b) it's still "after" event handling (except for round 1)
    665 		 */
    666 		if (puffs_req_putput(ppr) == -1)
    667 			goto out;
    668 		puffs_req_resetput(ppr);
    669 
    670 		/* micro optimization: skip kevent syscall if possible */
    671 		if (pu->pu_nfds == 1 && pu->pu_ml_timep == NULL
    672 		    && (pu->pu_state & PU_ASYNCFD) == 0) {
    673 			pfctrl = XXX->fctrl;
    674 			puffs_framev_input(pu, pfctrl, XXX);
    675 			continue;
    676 		}
    677 #endif
    678 
    679 		/* else: do full processing */
    680 		/* Don't bother worrying about O(n) for now */
    681 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    682 			if (fio->stat & FIO_WRGONE)
    683 				continue;
    684 
    685 			pfctrl = fio->fctrl;
    686 
    687 			/*
    688 			 * Try to write out everything to avoid the
    689 			 * need for enabling EVFILT_WRITE.  The likely
    690 			 * case is that we can fit everything into the
    691 			 * socket buffer.
    692 			 */
    693 			puffs_framev_output(pu, pfctrl, fio);
    694 		}
    695 
    696 		/*
    697 		 * Build list of which to enable/disable in writecheck.
    698 		 */
    699 		nchanges = 0;
    700 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    701 			if (fio->stat & FIO_WRGONE)
    702 				continue;
    703 
    704 			/* en/disable write checks for kqueue as needed */
    705 			assert((FIO_EN_WRITE(fio) && FIO_RM_WRITE(fio)) == 0);
    706 			if (FIO_EN_WRITE(fio)) {
    707 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    708 				    EVFILT_WRITE, EV_ENABLE, 0, 0,
    709 				    (uintptr_t)fio);
    710 				fio->stat |= FIO_WR;
    711 				nchanges++;
    712 			}
    713 			if (FIO_RM_WRITE(fio)) {
    714 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    715 				    EVFILT_WRITE, EV_DISABLE, 0, 0,
    716 				    (uintptr_t)fio);
    717 				fio->stat &= ~FIO_WR;
    718 				nchanges++;
    719 			}
    720 			assert(nchanges <= pu->pu_nfds);
    721 		}
    722 
    723 		ndone = kevent(pu->pu_kq, pu->pu_evs, nchanges,
    724 		    pu->pu_evs, 2*pu->pu_nfds, pu->pu_ml_timep);
    725 
    726 		if (ndone == -1) {
    727 			if (errno != EINTR)
    728 				goto out;
    729 			else
    730 				continue;
    731 		}
    732 
    733 		/* uoptimize */
    734 		if (ndone == 0)
    735 			continue;
    736 
    737 		/* iterate over the results */
    738 		for (curev = pu->pu_evs; ndone--; curev++) {
    739 			int what;
    740 
    741 #if 0
    742 			/* get & possibly dispatch events from kernel */
    743 			if (curev->ident == puffsfd) {
    744 				if (puffs_req_handle(pgr, ppr, 0) == -1)
    745 					goto out;
    746 				continue;
    747 			}
    748 #endif
    749 
    750 			fio = (void *)curev->udata;
    751 			pfctrl = fio->fctrl;
    752 			if (curev->flags & EV_ERROR) {
    753 				assert(curev->filter == EVFILT_WRITE);
    754 				fio->stat &= ~FIO_WR;
    755 
    756 				/* XXX: how to know if it's a transient error */
    757 				puffs_framev_writeclose(pu, fio,
    758 				    (int)curev->data);
    759 				puffs_framev_notify(fio, PUFFS_FBIO_ERROR);
    760 				continue;
    761 			}
    762 
    763 			what = 0;
    764 			if (curev->filter == EVFILT_READ) {
    765 				puffs_framev_input(pu, pfctrl, fio);
    766 				what |= PUFFS_FBIO_READ;
    767 			}
    768 
    769 			else if (curev->filter == EVFILT_WRITE) {
    770 				puffs_framev_output(pu, pfctrl, fio);
    771 				what |= PUFFS_FBIO_WRITE;
    772 			}
    773 			if (what)
    774 				puffs_framev_notify(fio, what);
    775 		}
    776 
    777 		/*
    778 		 * Schedule continuations.
    779 		 */
    780 		while ((pcc = TAILQ_FIRST(&pu->pu_sched)) != NULL) {
    781 			TAILQ_REMOVE(&pu->pu_sched, pcc, entries);
    782 			puffs_goto(pcc);
    783 		}
    784 
    785 		/*
    786 		 * Really free fd's now that we don't have references
    787 		 * to them.
    788 		 */
    789 		while ((fio = LIST_FIRST(&pu->pu_ios_rmlist)) != NULL) {
    790 			LIST_REMOVE(fio, fio_entries);
    791 			free(fio);
    792 		}
    793 	}
    794 	finalpush(pu);
    795 	errno = 0;
    796 
    797  out:
    798 	/* store the real error for a while */
    799 	sverrno = errno;
    800 
    801 	errno = sverrno;
    802 	if (errno)
    803 		return -1;
    804 	else
    805 		return 0;
    806 }
    807