Home | History | Annotate | Line # | Download | only in libpuffs
puffs.c revision 1.79
      1 /*	$NetBSD: puffs.c,v 1.79 2007/12/04 21:24:11 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 1.79 2007/12/04 21:24:11 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 int
    141 puffs_setblockingmode(struct puffs_usermount *pu, int mode)
    142 {
    143 	int rv, x;
    144 
    145 	assert(puffs_getstate(pu) == PUFFS_STATE_RUNNING);
    146 
    147 	if (mode != PUFFSDEV_BLOCK && mode != PUFFSDEV_NONBLOCK) {
    148 		errno = EINVAL;
    149 		return -1;
    150 	}
    151 
    152 	x = mode;
    153 	rv = ioctl(pu->pu_fd, FIONBIO, &x);
    154 
    155 	if (rv == 0) {
    156 		if (mode == PUFFSDEV_BLOCK)
    157 			pu->pu_state &= ~PU_ASYNCFD;
    158 		else
    159 			pu->pu_state |= PU_ASYNCFD;
    160 	}
    161 
    162 	return rv;
    163 }
    164 
    165 int
    166 puffs_getstate(struct puffs_usermount *pu)
    167 {
    168 
    169 	return pu->pu_state & PU_STATEMASK;
    170 }
    171 
    172 void
    173 puffs_setstacksize(struct puffs_usermount *pu, size_t ss)
    174 {
    175 	long psize;
    176 	int stackshift;
    177 
    178 	psize = sysconf(_SC_PAGESIZE);
    179 	if (ss < 2*psize) {
    180 		ss = 2*psize;
    181 		fprintf(stderr, "puffs_setstacksize: adjusting stacksize "
    182 		    "to minimum %zu\n", ss);
    183 	}
    184 
    185 	assert(puffs_getstate(pu) == PUFFS_STATE_BEFOREMOUNT);
    186 	stackshift = -1;
    187 	while (ss) {
    188 		ss >>= 1;
    189 		stackshift++;
    190 	}
    191 	pu->pu_cc_stackshift = stackshift;
    192 	assert(1<<stackshift == ss);
    193 }
    194 
    195 struct puffs_pathobj *
    196 puffs_getrootpathobj(struct puffs_usermount *pu)
    197 {
    198 	struct puffs_node *pnr;
    199 
    200 	pnr = pu->pu_pn_root;
    201 	if (pnr == NULL) {
    202 		errno = ENOENT;
    203 		return NULL;
    204 	}
    205 
    206 	return &pnr->pn_po;
    207 }
    208 
    209 void
    210 puffs_setroot(struct puffs_usermount *pu, struct puffs_node *pn)
    211 {
    212 
    213 	pu->pu_pn_root = pn;
    214 }
    215 
    216 struct puffs_node *
    217 puffs_getroot(struct puffs_usermount *pu)
    218 {
    219 
    220 	return pu->pu_pn_root;
    221 }
    222 
    223 void
    224 puffs_setrootinfo(struct puffs_usermount *pu, enum vtype vt,
    225 	vsize_t vsize, dev_t rdev)
    226 {
    227 	struct puffs_kargs *pargs = pu->pu_kargp;
    228 
    229 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT) {
    230 		warnx("puffs_setrootinfo: call has effect only "
    231 		    "before mount\n");
    232 		return;
    233 	}
    234 
    235 	pargs->pa_root_vtype = vt;
    236 	pargs->pa_root_vsize = vsize;
    237 	pargs->pa_root_rdev = rdev;
    238 }
    239 
    240 void *
    241 puffs_getspecific(struct puffs_usermount *pu)
    242 {
    243 
    244 	return pu->pu_privdata;
    245 }
    246 
    247 size_t
    248 puffs_getmaxreqlen(struct puffs_usermount *pu)
    249 {
    250 
    251 	return pu->pu_maxreqlen;
    252 }
    253 
    254 void
    255 puffs_setmaxreqlen(struct puffs_usermount *pu, size_t reqlen)
    256 {
    257 
    258 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    259 		warnx("puffs_setmaxreqlen: call has effect only "
    260 		    "before mount\n");
    261 
    262 	pu->pu_kargp->pa_maxmsglen = reqlen;
    263 }
    264 
    265 void
    266 puffs_setfhsize(struct puffs_usermount *pu, size_t fhsize, int flags)
    267 {
    268 
    269 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    270 		warnx("puffs_setfhsize: call has effect only before mount\n");
    271 
    272 	pu->pu_kargp->pa_fhsize = fhsize;
    273 	pu->pu_kargp->pa_fhflags = flags;
    274 }
    275 
    276 void
    277 puffs_setncookiehash(struct puffs_usermount *pu, int nhash)
    278 {
    279 
    280 	if (puffs_getstate(pu) != PUFFS_STATE_BEFOREMOUNT)
    281 		warnx("puffs_setfhsize: call has effect only before mount\n");
    282 
    283 	pu->pu_kargp->pa_nhashbuckets = nhash;
    284 }
    285 
    286 void
    287 puffs_set_pathbuild(struct puffs_usermount *pu, pu_pathbuild_fn fn)
    288 {
    289 
    290 	pu->pu_pathbuild = fn;
    291 }
    292 
    293 void
    294 puffs_set_pathtransform(struct puffs_usermount *pu, pu_pathtransform_fn fn)
    295 {
    296 
    297 	pu->pu_pathtransform = fn;
    298 }
    299 
    300 void
    301 puffs_set_pathcmp(struct puffs_usermount *pu, pu_pathcmp_fn fn)
    302 {
    303 
    304 	pu->pu_pathcmp = fn;
    305 }
    306 
    307 void
    308 puffs_set_pathfree(struct puffs_usermount *pu, pu_pathfree_fn fn)
    309 {
    310 
    311 	pu->pu_pathfree = fn;
    312 }
    313 
    314 void
    315 puffs_set_namemod(struct puffs_usermount *pu, pu_namemod_fn fn)
    316 {
    317 
    318 	pu->pu_namemod = fn;
    319 }
    320 
    321 void
    322 puffs_set_errnotify(struct puffs_usermount *pu, pu_errnotify_fn fn)
    323 {
    324 
    325 	pu->pu_errnotify = fn;
    326 }
    327 
    328 void
    329 puffs_ml_setloopfn(struct puffs_usermount *pu, puffs_ml_loop_fn lfn)
    330 {
    331 
    332 	pu->pu_ml_lfn = lfn;
    333 }
    334 
    335 void
    336 puffs_ml_settimeout(struct puffs_usermount *pu, struct timespec *ts)
    337 {
    338 
    339 	if (ts == NULL) {
    340 		pu->pu_ml_timep = NULL;
    341 	} else {
    342 		pu->pu_ml_timeout = *ts;
    343 		pu->pu_ml_timep = &pu->pu_ml_timeout;
    344 	}
    345 }
    346 
    347 void
    348 puffs_set_prepost(struct puffs_usermount *pu,
    349 	pu_prepost_fn pre, pu_prepost_fn pst)
    350 {
    351 
    352 	pu->pu_oppre = pre;
    353 	pu->pu_oppost = pst;
    354 }
    355 
    356 void
    357 puffs_setback(struct puffs_cc *pcc, int whatback)
    358 {
    359 	struct puffs_req *preq = puffs__framebuf_getdataptr(pcc->pcc_pb);
    360 
    361 	assert(PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN && (
    362 	    preq->preq_optype == PUFFS_VN_OPEN ||
    363 	    preq->preq_optype == PUFFS_VN_MMAP ||
    364 	    preq->preq_optype == PUFFS_VN_REMOVE ||
    365 	    preq->preq_optype == PUFFS_VN_RMDIR ||
    366 	    preq->preq_optype == PUFFS_VN_INACTIVE));
    367 
    368 	preq->preq_setbacks |= whatback & PUFFS_SETBACK_MASK;
    369 }
    370 
    371 int
    372 puffs_daemon(struct puffs_usermount *pu, int nochdir, int noclose)
    373 {
    374 	ssize_t n;
    375 	int parent, value, fd;
    376 
    377 	if (pipe(pu->pu_dpipe) == -1)
    378 		return -1;
    379 
    380 	switch (fork()) {
    381 	case -1:
    382 		return -1;
    383 	case 0:
    384 		parent = 0;
    385 		break;
    386 	default:
    387 		parent = 1;
    388 		break;
    389 	}
    390 	pu->pu_state |= PU_PUFFSDAEMON;
    391 
    392 	if (parent) {
    393 		n = read(pu->pu_dpipe[0], &value, sizeof(int));
    394 		if (n == -1)
    395 			err(1, "puffs_daemon");
    396 		assert(n == sizeof(value));
    397 		if (value) {
    398 			errno = value;
    399 			err(1, "puffs_daemon");
    400 		}
    401 		exit(0);
    402 	} else {
    403 		if (setsid() == -1)
    404 			goto fail;
    405 
    406 		if (!nochdir)
    407 			chdir("/");
    408 
    409 		if (!noclose) {
    410 			fd = open(_PATH_DEVNULL, O_RDWR, 0);
    411 			if (fd == -1)
    412 				goto fail;
    413 			dup2(fd, STDIN_FILENO);
    414 			dup2(fd, STDOUT_FILENO);
    415 			dup2(fd, STDERR_FILENO);
    416 			if (fd > STDERR_FILENO)
    417 				close(fd);
    418 		}
    419 		return 0;
    420 	}
    421 
    422  fail:
    423 	n = write(pu->pu_dpipe[1], &errno, sizeof(int));
    424 	assert(n == 4);
    425 	return -1;
    426 }
    427 
    428 int
    429 puffs_mount(struct puffs_usermount *pu, const char *dir, int mntflags,
    430 	void *cookie)
    431 {
    432 	char rp[MAXPATHLEN];
    433 	ssize_t n;
    434 	int rv, fd, sverrno;
    435 
    436 #if 1
    437 	/* XXXkludgehere */
    438 	/* kauth doesn't provide this service any longer */
    439 	if (geteuid() != 0)
    440 		mntflags |= MNT_NOSUID | MNT_NODEV;
    441 #endif
    442 
    443 	if (realpath(dir, rp) == NULL) {
    444 		rv = -1;
    445 		goto out;
    446 	}
    447 
    448 	if (strcmp(dir, rp) != 0) {
    449 		warnx("puffs_mount: \"%s\" is a relative path.", dir);
    450 		warnx("puffs_mount: using \"%s\" instead.", rp);
    451 	}
    452 
    453 	fd = open(_PATH_PUFFS, O_RDWR);
    454 	if (fd == -1) {
    455 		warnx("puffs_mount: cannot open %s", _PATH_PUFFS);
    456 		rv = -1;
    457 		goto out;
    458 	}
    459 	if (fd <= 2)
    460 		warnx("puffs_init: device fd %d (<= 2), sure this is "
    461 		    "what you want?", fd);
    462 
    463 	pu->pu_kargp->pa_fd = pu->pu_fd = fd;
    464 	pu->pu_kargp->pa_root_cookie = cookie;
    465 	if ((rv = mount(MOUNT_PUFFS, rp, mntflags,
    466 	    pu->pu_kargp, sizeof(struct puffs_kargs))) == -1)
    467 		goto out;
    468 #if 0
    469 	if ((rv = ioctl(pu->pu_fd, PUFFSREQSIZEOP, &pu->pu_maxreqlen)) == -1)
    470 		goto out;
    471 #endif
    472 	PU_SETSTATE(pu, PUFFS_STATE_RUNNING);
    473 
    474  out:
    475 	if (rv != 0)
    476 		sverrno = errno;
    477 	else
    478 		sverrno = 0;
    479 	free(pu->pu_kargp);
    480 	pu->pu_kargp = NULL;
    481 
    482 	if (pu->pu_state & PU_PUFFSDAEMON) {
    483 		n = write(pu->pu_dpipe[1], &sverrno, sizeof(int));
    484 		assert(n == 4);
    485 		close(pu->pu_dpipe[0]);
    486 		close(pu->pu_dpipe[1]);
    487 	}
    488 
    489 	errno = sverrno;
    490 	return rv;
    491 }
    492 
    493 struct puffs_usermount *
    494 _puffs_init(int develv, struct puffs_ops *pops, const char *mntfromname,
    495 	const char *puffsname, void *priv, uint32_t pflags)
    496 {
    497 	struct puffs_usermount *pu;
    498 	struct puffs_kargs *pargs;
    499 	int sverrno;
    500 
    501 	if (develv != PUFFS_DEVEL_LIBVERSION) {
    502 		warnx("puffs_init: mounting with lib version %d, need %d",
    503 		    develv, PUFFS_DEVEL_LIBVERSION);
    504 		errno = EINVAL;
    505 		return NULL;
    506 	}
    507 
    508 	pu = malloc(sizeof(struct puffs_usermount));
    509 	if (pu == NULL)
    510 		goto failfree;
    511 	memset(pu, 0, sizeof(struct puffs_usermount));
    512 
    513 	pargs = pu->pu_kargp = malloc(sizeof(struct puffs_kargs));
    514 	if (pargs == NULL)
    515 		goto failfree;
    516 	memset(pargs, 0, sizeof(struct puffs_kargs));
    517 
    518 	pargs->pa_vers = PUFFSDEVELVERS | PUFFSVERSION;
    519 	pargs->pa_flags = PUFFS_FLAG_KERN(pflags);
    520 	fillvnopmask(pops, pargs->pa_vnopmask);
    521 	(void)strlcpy(pargs->pa_typename, puffsname,
    522 	    sizeof(pargs->pa_typename));
    523 	(void)strlcpy(pargs->pa_mntfromname, mntfromname,
    524 	    sizeof(pargs->pa_mntfromname));
    525 
    526 	puffs_zerostatvfs(&pargs->pa_svfsb);
    527 	pargs->pa_root_cookie = NULL;
    528 	pargs->pa_root_vtype = VDIR;
    529 	pargs->pa_root_vsize = 0;
    530 	pargs->pa_root_rdev = 0;
    531 	pargs->pa_maxmsglen = 0;
    532 
    533 	pu->pu_flags = pflags;
    534 	pu->pu_ops = *pops;
    535 	free(pops); /* XXX */
    536 
    537 	pu->pu_privdata = priv;
    538 	pu->pu_cc_stackshift = PUFFS_CC_STACKSHIFT_DEFAULT;
    539 	LIST_INIT(&pu->pu_pnodelst);
    540 	LIST_INIT(&pu->pu_ios);
    541 	LIST_INIT(&pu->pu_ios_rmlist);
    542 	LIST_INIT(&pu->pu_ccnukelst);
    543 	TAILQ_INIT(&pu->pu_sched);
    544 	TAILQ_INIT(&pu->pu_exq);
    545 
    546 	pu->pu_framectrl[PU_FRAMECTRL_FS].rfb = puffs_fsframe_read;
    547 	pu->pu_framectrl[PU_FRAMECTRL_FS].wfb = puffs_fsframe_write;
    548 	pu->pu_framectrl[PU_FRAMECTRL_FS].cmpfb = puffs_fsframe_cmp;
    549 	pu->pu_framectrl[PU_FRAMECTRL_FS].gotfb = puffs_fsframe_gotframe;
    550 	pu->pu_framectrl[PU_FRAMECTRL_FS].fdnotfn = puffs_framev_unmountonclose;
    551 
    552 	/* defaults for some user-settable translation functions */
    553 	pu->pu_cmap = NULL; /* identity translation */
    554 
    555 	pu->pu_pathbuild = puffs_stdpath_buildpath;
    556 	pu->pu_pathfree = puffs_stdpath_freepath;
    557 	pu->pu_pathcmp = puffs_stdpath_cmppath;
    558 	pu->pu_pathtransform = NULL;
    559 	pu->pu_namemod = NULL;
    560 
    561 	pu->pu_errnotify = puffs_defaulterror;
    562 
    563 	PU_SETSTATE(pu, PUFFS_STATE_BEFOREMOUNT);
    564 
    565 	return pu;
    566 
    567  failfree:
    568 	/* can't unmount() from here for obvious reasons */
    569 	sverrno = errno;
    570 	free(pu);
    571 	errno = sverrno;
    572 	return NULL;
    573 }
    574 
    575 /*
    576  * XXX: there's currently no clean way to request unmount from
    577  * within the user server, so be very brutal about it.
    578  */
    579 /*ARGSUSED1*/
    580 int
    581 puffs_exit(struct puffs_usermount *pu, int force)
    582 {
    583 	struct puffs_node *pn;
    584 
    585 	force = 1; /* currently */
    586 
    587 	if (pu->pu_fd)
    588 		close(pu->pu_fd);
    589 
    590 	while ((pn = LIST_FIRST(&pu->pu_pnodelst)) != NULL)
    591 		puffs_pn_put(pn);
    592 
    593 	finalpush(pu);
    594 	puffs_framev_exit(pu);
    595 	if (pu->pu_state & PU_HASKQ)
    596 		close(pu->pu_kq);
    597 	free(pu);
    598 
    599 	return 0; /* always succesful for now, WILL CHANGE */
    600 }
    601 
    602 int
    603 puffs_mainloop(struct puffs_usermount *pu)
    604 {
    605 	struct puffs_framectrl *pfctrl;
    606 	struct puffs_fctrl_io *fio;
    607 	struct puffs_cc *pcc;
    608 	struct kevent *curev;
    609 	size_t nchanges;
    610 	int ndone, sverrno;
    611 
    612 	assert(puffs_getstate(pu) >= PUFFS_STATE_RUNNING);
    613 
    614 	pu->pu_kq = kqueue();
    615 	if (pu->pu_kq == -1)
    616 		goto out;
    617 	pu->pu_state |= PU_HASKQ;
    618 
    619 	puffs_setblockingmode(pu, PUFFSDEV_NONBLOCK);
    620 	if (puffs__framev_addfd_ctrl(pu, puffs_getselectable(pu),
    621 	    PUFFS_FBIO_READ | PUFFS_FBIO_WRITE,
    622 	    &pu->pu_framectrl[PU_FRAMECTRL_FS]) == -1)
    623 		goto out;
    624 
    625 	curev = realloc(pu->pu_evs, (2*pu->pu_nfds)*sizeof(struct kevent));
    626 	if (curev == NULL)
    627 		goto out;
    628 	pu->pu_evs = curev;
    629 
    630 	LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    631 		EV_SET(curev, fio->io_fd, EVFILT_READ, EV_ADD,
    632 		    0, 0, (uintptr_t)fio);
    633 		curev++;
    634 		EV_SET(curev, fio->io_fd, EVFILT_WRITE, EV_ADD | EV_DISABLE,
    635 		    0, 0, (uintptr_t)fio);
    636 		curev++;
    637 	}
    638 	if (kevent(pu->pu_kq, pu->pu_evs, 2*pu->pu_nfds, NULL, 0, NULL) == -1)
    639 		goto out;
    640 
    641 	pu->pu_state |= PU_INLOOP;
    642 
    643 	while (puffs_getstate(pu) != PUFFS_STATE_UNMOUNTED) {
    644 		if (pu->pu_ml_lfn)
    645 			pu->pu_ml_lfn(pu);
    646 
    647 		/* XXX: can we still do these optimizations? */
    648 #if 0
    649 		/*
    650 		 * Do this here, because:
    651 		 *  a) loopfunc might generate some results
    652 		 *  b) it's still "after" event handling (except for round 1)
    653 		 */
    654 		if (puffs_req_putput(ppr) == -1)
    655 			goto out;
    656 		puffs_req_resetput(ppr);
    657 
    658 		/* micro optimization: skip kevent syscall if possible */
    659 		if (pu->pu_nfds == 1 && pu->pu_ml_timep == NULL
    660 		    && (pu->pu_state & PU_ASYNCFD) == 0) {
    661 			pfctrl = XXX->fctrl;
    662 			puffs_framev_input(pu, pfctrl, XXX);
    663 			continue;
    664 		}
    665 #endif
    666 
    667 		/* else: do full processing */
    668 		/* Don't bother worrying about O(n) for now */
    669 		nchanges = 0;
    670 
    671 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    672 			if (fio->stat & FIO_WRGONE)
    673 				continue;
    674 
    675 			pfctrl = fio->fctrl;
    676 
    677 			/*
    678 			 * Try to write out everything to avoid the
    679 			 * need for enabling EVFILT_WRITE.  The likely
    680 			 * case is that we can fit everything into the
    681 			 * socket buffer.
    682 			 */
    683 			puffs_framev_output(pu, pfctrl, fio);
    684 		}
    685 
    686 		/*
    687 		 * Build list of which to enable/disable in writecheck.
    688 		 */
    689 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    690 			if (fio->stat & FIO_WRGONE)
    691 				continue;
    692 
    693 			/* en/disable write checks for kqueue as needed */
    694 			assert((FIO_EN_WRITE(fio) && FIO_RM_WRITE(fio)) == 0);
    695 			if (FIO_EN_WRITE(fio)) {
    696 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    697 				    EVFILT_WRITE, EV_ENABLE, 0, 0,
    698 				    (uintptr_t)fio);
    699 				fio->stat |= FIO_WR;
    700 				nchanges++;
    701 			}
    702 			if (FIO_RM_WRITE(fio)) {
    703 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    704 				    EVFILT_WRITE, EV_DISABLE, 0, 0,
    705 				    (uintptr_t)fio);
    706 				fio->stat &= ~FIO_WR;
    707 				nchanges++;
    708 			}
    709 			assert(nchanges <= pu->pu_nfds);
    710 		}
    711 
    712 		ndone = kevent(pu->pu_kq, pu->pu_evs, nchanges,
    713 		    pu->pu_evs, 2*pu->pu_nfds, pu->pu_ml_timep);
    714 
    715 		if (ndone == -1) {
    716 			if (errno != EINTR)
    717 				goto out;
    718 			else
    719 				continue;
    720 		}
    721 
    722 		/* uoptimize */
    723 		if (ndone == 0)
    724 			continue;
    725 
    726 		/* iterate over the results */
    727 		for (curev = pu->pu_evs; ndone--; curev++) {
    728 			int what;
    729 
    730 #if 0
    731 			/* get & possibly dispatch events from kernel */
    732 			if (curev->ident == puffsfd) {
    733 				if (puffs_req_handle(pgr, ppr, 0) == -1)
    734 					goto out;
    735 				continue;
    736 			}
    737 #endif
    738 
    739 			fio = (void *)curev->udata;
    740 			pfctrl = fio->fctrl;
    741 			if (curev->flags & EV_ERROR) {
    742 				assert(curev->filter == EVFILT_WRITE);
    743 				fio->stat &= ~FIO_WR;
    744 
    745 				/* XXX: how to know if it's a transient error */
    746 				puffs_framev_writeclose(pu, fio,
    747 				    (int)curev->data);
    748 				puffs_framev_notify(fio, PUFFS_FBIO_ERROR);
    749 				continue;
    750 			}
    751 
    752 			what = 0;
    753 			if (curev->filter == EVFILT_READ) {
    754 				puffs_framev_input(pu, pfctrl, fio);
    755 				what |= PUFFS_FBIO_READ;
    756 			}
    757 
    758 			else if (curev->filter == EVFILT_WRITE) {
    759 				puffs_framev_output(pu, pfctrl, fio);
    760 				what |= PUFFS_FBIO_WRITE;
    761 			}
    762 			if (what)
    763 				puffs_framev_notify(fio, what);
    764 		}
    765 
    766 		/*
    767 		 * Schedule continuations.
    768 		 */
    769 		while ((pcc = TAILQ_FIRST(&pu->pu_sched)) != NULL) {
    770 			TAILQ_REMOVE(&pu->pu_sched, pcc, entries);
    771 			puffs_goto(pcc);
    772 		}
    773 
    774 		/*
    775 		 * Really free fd's now that we don't have references
    776 		 * to them.
    777 		 */
    778 		while ((fio = LIST_FIRST(&pu->pu_ios_rmlist)) != NULL) {
    779 			LIST_REMOVE(fio, fio_entries);
    780 			free(fio);
    781 		}
    782 	}
    783 	finalpush(pu);
    784 	errno = 0;
    785 
    786  out:
    787 	/* store the real error for a while */
    788 	sverrno = errno;
    789 
    790 	errno = sverrno;
    791 	if (errno)
    792 		return -1;
    793 	else
    794 		return 0;
    795 }
    796