Home | History | Annotate | Line # | Download | only in libpuffs
puffs.c revision 1.83 
      1 /*	$NetBSD: puffs.c,v 1.83 2008/01/02 17:57:51 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 1.83 2008/01/02 17:57:51 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_set_cmap(struct puffs_usermount *pu, pu_cmap_fn fn)
    342 {
    343 
    344 	pu->pu_cmap = fn;
    345 }
    346 
    347 void
    348 puffs_ml_setloopfn(struct puffs_usermount *pu, puffs_ml_loop_fn lfn)
    349 {
    350 
    351 	pu->pu_ml_lfn = lfn;
    352 }
    353 
    354 void
    355 puffs_ml_settimeout(struct puffs_usermount *pu, struct timespec *ts)
    356 {
    357 
    358 	if (ts == NULL) {
    359 		pu->pu_ml_timep = NULL;
    360 	} else {
    361 		pu->pu_ml_timeout = *ts;
    362 		pu->pu_ml_timep = &pu->pu_ml_timeout;
    363 	}
    364 }
    365 
    366 void
    367 puffs_set_prepost(struct puffs_usermount *pu,
    368 	pu_prepost_fn pre, pu_prepost_fn pst)
    369 {
    370 
    371 	pu->pu_oppre = pre;
    372 	pu->pu_oppost = pst;
    373 }
    374 
    375 void
    376 puffs_setback(struct puffs_cc *pcc, int whatback)
    377 {
    378 	struct puffs_req *preq = puffs__framebuf_getdataptr(pcc->pcc_pb);
    379 
    380 	assert(PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN && (
    381 	    preq->preq_optype == PUFFS_VN_OPEN ||
    382 	    preq->preq_optype == PUFFS_VN_MMAP ||
    383 	    preq->preq_optype == PUFFS_VN_REMOVE ||
    384 	    preq->preq_optype == PUFFS_VN_RMDIR ||
    385 	    preq->preq_optype == PUFFS_VN_INACTIVE));
    386 
    387 	preq->preq_setbacks |= whatback & PUFFS_SETBACK_MASK;
    388 }
    389 
    390 int
    391 puffs_daemon(struct puffs_usermount *pu, int nochdir, int noclose)
    392 {
    393 	ssize_t n;
    394 	int parent, value, fd;
    395 
    396 	if (pipe(pu->pu_dpipe) == -1)
    397 		return -1;
    398 
    399 	switch (fork()) {
    400 	case -1:
    401 		return -1;
    402 	case 0:
    403 		parent = 0;
    404 		break;
    405 	default:
    406 		parent = 1;
    407 		break;
    408 	}
    409 	pu->pu_state |= PU_PUFFSDAEMON;
    410 
    411 	if (parent) {
    412 		n = read(pu->pu_dpipe[0], &value, sizeof(int));
    413 		if (n == -1)
    414 			err(1, "puffs_daemon");
    415 		assert(n == sizeof(value));
    416 		if (value) {
    417 			errno = value;
    418 			err(1, "puffs_daemon");
    419 		}
    420 		exit(0);
    421 	} else {
    422 		if (setsid() == -1)
    423 			goto fail;
    424 
    425 		if (!nochdir)
    426 			chdir("/");
    427 
    428 		if (!noclose) {
    429 			fd = open(_PATH_DEVNULL, O_RDWR, 0);
    430 			if (fd == -1)
    431 				goto fail;
    432 			dup2(fd, STDIN_FILENO);
    433 			dup2(fd, STDOUT_FILENO);
    434 			dup2(fd, STDERR_FILENO);
    435 			if (fd > STDERR_FILENO)
    436 				close(fd);
    437 		}
    438 		return 0;
    439 	}
    440 
    441  fail:
    442 	n = write(pu->pu_dpipe[1], &errno, sizeof(int));
    443 	assert(n == 4);
    444 	return -1;
    445 }
    446 
    447 int
    448 puffs_mount(struct puffs_usermount *pu, const char *dir, int mntflags,
    449 	void *cookie)
    450 {
    451 	char rp[MAXPATHLEN];
    452 	ssize_t n;
    453 	int rv, fd, sverrno;
    454 	char *comfd;
    455 
    456 	pu->pu_kargp->pa_root_cookie = cookie;
    457 
    458 	/* XXXkludgehere */
    459 	/* kauth doesn't provide this service any longer */
    460 	if (geteuid() != 0)
    461 		mntflags |= MNT_NOSUID | MNT_NODEV;
    462 
    463 	if (realpath(dir, rp) == NULL) {
    464 		rv = -1;
    465 		goto out;
    466 	}
    467 
    468 	if (strcmp(dir, rp) != 0) {
    469 		warnx("puffs_mount: \"%s\" is a relative path.", dir);
    470 		warnx("puffs_mount: using \"%s\" instead.", rp);
    471 	}
    472 
    473 	/*
    474 	 * Undocumented...  Well, documented only here.
    475 	 *
    476 	 * This is used for imaginative purposes.  If the env variable is
    477 	 * set, puffs_mount() doesn't do the regular mount procedure.
    478 	 * Rather, it crams the mount data down the comfd and sets comfd as
    479 	 * the puffs descriptor.
    480 	 *
    481 	 * This shouldn't be used unless you can read my mind ( ... or write
    482 	 * it, not to mention execute it, but that's starting to get silly).
    483 	 */
    484 	if ((comfd = getenv("PUFFS_COMFD")) != NULL) {
    485 		size_t len;
    486 
    487 		if (sscanf(comfd, "%d", &pu->pu_fd) != 1) {
    488 			errno = EINVAL;
    489 			rv = -1;
    490 			goto out;
    491 		}
    492 		/* check that what we got at least resembles an fd */
    493 		if (fcntl(pu->pu_fd, F_GETFL) == -1) {
    494 			rv = -1;
    495 			goto out;
    496 		}
    497 
    498 		len = strlen(dir)+1;
    499 
    500 #define allwrite(buf, len)						\
    501 do {									\
    502 	ssize_t al_rv;							\
    503 	al_rv = write(pu->pu_fd, buf, len);				\
    504 	if (al_rv != len) {						\
    505 		if (al_rv != -1)					\
    506 			errno = EIO;					\
    507 		rv = -1;						\
    508 		abort();\
    509 		goto out;						\
    510 	}								\
    511 } while (/*CONSTCOND*/0)
    512 		allwrite(&len, sizeof(len));
    513 		allwrite(dir, len);
    514 		len = strlen(dir)+1;
    515 		allwrite(&len, sizeof(len));
    516 		allwrite(pu->pu_kargp->pa_mntfromname, len);
    517 		allwrite(&mntflags, sizeof(mntflags));
    518 		allwrite(pu->pu_kargp, sizeof(*pu->pu_kargp));
    519 		allwrite(&pu->pu_flags, sizeof(pu->pu_flags));
    520 #undef allwrite
    521 
    522 		rv = 0;
    523 	} else {
    524 		fd = open(_PATH_PUFFS, O_RDWR);
    525 		if (fd == -1) {
    526 			warnx("puffs_mount: cannot open %s", _PATH_PUFFS);
    527 			rv = -1;
    528 			goto out;
    529 		}
    530 		if (fd <= 2)
    531 			warnx("puffs_mount: device fd %d (<= 2), sure this is "
    532 			    "what you want?", fd);
    533 
    534 		pu->pu_kargp->pa_fd = pu->pu_fd = fd;
    535 		if ((rv = mount(MOUNT_PUFFS, rp, mntflags,
    536 		    pu->pu_kargp, sizeof(struct puffs_kargs))) == -1)
    537 			goto out;
    538 	}
    539 
    540 	PU_SETSTATE(pu, PUFFS_STATE_RUNNING);
    541 
    542  out:
    543 	if (rv != 0)
    544 		sverrno = errno;
    545 	else
    546 		sverrno = 0;
    547 	free(pu->pu_kargp);
    548 	pu->pu_kargp = NULL;
    549 
    550 	if (pu->pu_state & PU_PUFFSDAEMON) {
    551 		n = write(pu->pu_dpipe[1], &sverrno, sizeof(int));
    552 		assert(n == 4);
    553 		close(pu->pu_dpipe[0]);
    554 		close(pu->pu_dpipe[1]);
    555 	}
    556 
    557 	errno = sverrno;
    558 	return rv;
    559 }
    560 
    561 struct puffs_usermount *
    562 _puffs_init(int develv, struct puffs_ops *pops, const char *mntfromname,
    563 	const char *puffsname, void *priv, uint32_t pflags)
    564 {
    565 	struct puffs_usermount *pu;
    566 	struct puffs_kargs *pargs;
    567 	int sverrno;
    568 
    569 	if (develv != PUFFS_DEVEL_LIBVERSION) {
    570 		warnx("puffs_init: mounting with lib version %d, need %d",
    571 		    develv, PUFFS_DEVEL_LIBVERSION);
    572 		errno = EINVAL;
    573 		return NULL;
    574 	}
    575 
    576 	pu = malloc(sizeof(struct puffs_usermount));
    577 	if (pu == NULL)
    578 		goto failfree;
    579 	memset(pu, 0, sizeof(struct puffs_usermount));
    580 
    581 	pargs = pu->pu_kargp = malloc(sizeof(struct puffs_kargs));
    582 	if (pargs == NULL)
    583 		goto failfree;
    584 	memset(pargs, 0, sizeof(struct puffs_kargs));
    585 
    586 	pargs->pa_vers = PUFFSDEVELVERS | PUFFSVERSION;
    587 	pargs->pa_flags = PUFFS_FLAG_KERN(pflags);
    588 	fillvnopmask(pops, pargs->pa_vnopmask);
    589 	(void)strlcpy(pargs->pa_typename, puffsname,
    590 	    sizeof(pargs->pa_typename));
    591 	(void)strlcpy(pargs->pa_mntfromname, mntfromname,
    592 	    sizeof(pargs->pa_mntfromname));
    593 
    594 	puffs_zerostatvfs(&pargs->pa_svfsb);
    595 	pargs->pa_root_cookie = NULL;
    596 	pargs->pa_root_vtype = VDIR;
    597 	pargs->pa_root_vsize = 0;
    598 	pargs->pa_root_rdev = 0;
    599 	pargs->pa_maxmsglen = 0;
    600 
    601 	pu->pu_flags = pflags;
    602 	pu->pu_ops = *pops;
    603 	free(pops); /* XXX */
    604 
    605 	pu->pu_privdata = priv;
    606 	pu->pu_cc_stackshift = PUFFS_CC_STACKSHIFT_DEFAULT;
    607 	LIST_INIT(&pu->pu_pnodelst);
    608 	LIST_INIT(&pu->pu_ios);
    609 	LIST_INIT(&pu->pu_ios_rmlist);
    610 	LIST_INIT(&pu->pu_ccnukelst);
    611 	TAILQ_INIT(&pu->pu_sched);
    612 	TAILQ_INIT(&pu->pu_exq);
    613 
    614 	pu->pu_framectrl[PU_FRAMECTRL_FS].rfb = puffs_fsframe_read;
    615 	pu->pu_framectrl[PU_FRAMECTRL_FS].wfb = puffs_fsframe_write;
    616 	pu->pu_framectrl[PU_FRAMECTRL_FS].cmpfb = puffs_fsframe_cmp;
    617 	pu->pu_framectrl[PU_FRAMECTRL_FS].gotfb = puffs_fsframe_gotframe;
    618 	pu->pu_framectrl[PU_FRAMECTRL_FS].fdnotfn = puffs_framev_unmountonclose;
    619 
    620 	/* defaults for some user-settable translation functions */
    621 	pu->pu_cmap = NULL; /* identity translation */
    622 
    623 	pu->pu_pathbuild = puffs_stdpath_buildpath;
    624 	pu->pu_pathfree = puffs_stdpath_freepath;
    625 	pu->pu_pathcmp = puffs_stdpath_cmppath;
    626 	pu->pu_pathtransform = NULL;
    627 	pu->pu_namemod = NULL;
    628 
    629 	pu->pu_errnotify = puffs_defaulterror;
    630 
    631 	PU_SETSTATE(pu, PUFFS_STATE_BEFOREMOUNT);
    632 
    633 	return pu;
    634 
    635  failfree:
    636 	/* can't unmount() from here for obvious reasons */
    637 	sverrno = errno;
    638 	free(pu);
    639 	errno = sverrno;
    640 	return NULL;
    641 }
    642 
    643 /*
    644  * XXX: there's currently no clean way to request unmount from
    645  * within the user server, so be very brutal about it.
    646  */
    647 /*ARGSUSED1*/
    648 int
    649 puffs_exit(struct puffs_usermount *pu, int force)
    650 {
    651 	struct puffs_node *pn;
    652 
    653 	force = 1; /* currently */
    654 
    655 	if (pu->pu_fd)
    656 		close(pu->pu_fd);
    657 
    658 	while ((pn = LIST_FIRST(&pu->pu_pnodelst)) != NULL)
    659 		puffs_pn_put(pn);
    660 
    661 	finalpush(pu);
    662 	puffs_framev_exit(pu);
    663 	if (pu->pu_state & PU_HASKQ)
    664 		close(pu->pu_kq);
    665 	free(pu);
    666 
    667 	return 0; /* always succesful for now, WILL CHANGE */
    668 }
    669 
    670 int
    671 puffs_mainloop(struct puffs_usermount *pu)
    672 {
    673 	struct puffs_framectrl *pfctrl;
    674 	struct puffs_fctrl_io *fio;
    675 	struct puffs_cc *pcc;
    676 	struct kevent *curev;
    677 	size_t nchanges;
    678 	int ndone, sverrno;
    679 
    680 	assert(puffs_getstate(pu) >= PUFFS_STATE_RUNNING);
    681 
    682 	pu->pu_kq = kqueue();
    683 	if (pu->pu_kq == -1)
    684 		goto out;
    685 	pu->pu_state |= PU_HASKQ;
    686 
    687 	puffs_setblockingmode(pu, PUFFSDEV_NONBLOCK);
    688 	if (puffs__framev_addfd_ctrl(pu, puffs_getselectable(pu),
    689 	    PUFFS_FBIO_READ | PUFFS_FBIO_WRITE,
    690 	    &pu->pu_framectrl[PU_FRAMECTRL_FS]) == -1)
    691 		goto out;
    692 
    693 	curev = realloc(pu->pu_evs, (2*pu->pu_nfds)*sizeof(struct kevent));
    694 	if (curev == NULL)
    695 		goto out;
    696 	pu->pu_evs = curev;
    697 
    698 	LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    699 		EV_SET(curev, fio->io_fd, EVFILT_READ, EV_ADD,
    700 		    0, 0, (uintptr_t)fio);
    701 		curev++;
    702 		EV_SET(curev, fio->io_fd, EVFILT_WRITE, EV_ADD | EV_DISABLE,
    703 		    0, 0, (uintptr_t)fio);
    704 		curev++;
    705 	}
    706 	if (kevent(pu->pu_kq, pu->pu_evs, 2*pu->pu_nfds, NULL, 0, NULL) == -1)
    707 		goto out;
    708 
    709 	pu->pu_state |= PU_INLOOP;
    710 
    711 	while (puffs_getstate(pu) != PUFFS_STATE_UNMOUNTED) {
    712 		if (pu->pu_ml_lfn)
    713 			pu->pu_ml_lfn(pu);
    714 
    715 		/* XXX: can we still do these optimizations? */
    716 #if 0
    717 		/*
    718 		 * Do this here, because:
    719 		 *  a) loopfunc might generate some results
    720 		 *  b) it's still "after" event handling (except for round 1)
    721 		 */
    722 		if (puffs_req_putput(ppr) == -1)
    723 			goto out;
    724 		puffs_req_resetput(ppr);
    725 
    726 		/* micro optimization: skip kevent syscall if possible */
    727 		if (pu->pu_nfds == 1 && pu->pu_ml_timep == NULL
    728 		    && (pu->pu_state & PU_ASYNCFD) == 0) {
    729 			pfctrl = XXX->fctrl;
    730 			puffs_framev_input(pu, pfctrl, XXX);
    731 			continue;
    732 		}
    733 #endif
    734 
    735 		/* else: do full processing */
    736 		/* Don't bother worrying about O(n) for now */
    737 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    738 			if (fio->stat & FIO_WRGONE)
    739 				continue;
    740 
    741 			pfctrl = fio->fctrl;
    742 
    743 			/*
    744 			 * Try to write out everything to avoid the
    745 			 * need for enabling EVFILT_WRITE.  The likely
    746 			 * case is that we can fit everything into the
    747 			 * socket buffer.
    748 			 */
    749 			puffs_framev_output(pu, pfctrl, fio);
    750 		}
    751 
    752 		/*
    753 		 * Build list of which to enable/disable in writecheck.
    754 		 */
    755 		nchanges = 0;
    756 		LIST_FOREACH(fio, &pu->pu_ios, fio_entries) {
    757 			if (fio->stat & FIO_WRGONE)
    758 				continue;
    759 
    760 			/* en/disable write checks for kqueue as needed */
    761 			assert((FIO_EN_WRITE(fio) && FIO_RM_WRITE(fio)) == 0);
    762 			if (FIO_EN_WRITE(fio)) {
    763 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    764 				    EVFILT_WRITE, EV_ENABLE, 0, 0,
    765 				    (uintptr_t)fio);
    766 				fio->stat |= FIO_WR;
    767 				nchanges++;
    768 			}
    769 			if (FIO_RM_WRITE(fio)) {
    770 				EV_SET(&pu->pu_evs[nchanges], fio->io_fd,
    771 				    EVFILT_WRITE, EV_DISABLE, 0, 0,
    772 				    (uintptr_t)fio);
    773 				fio->stat &= ~FIO_WR;
    774 				nchanges++;
    775 			}
    776 			assert(nchanges <= pu->pu_nfds);
    777 		}
    778 
    779 		ndone = kevent(pu->pu_kq, pu->pu_evs, nchanges,
    780 		    pu->pu_evs, 2*pu->pu_nfds, pu->pu_ml_timep);
    781 
    782 		if (ndone == -1) {
    783 			if (errno != EINTR)
    784 				goto out;
    785 			else
    786 				continue;
    787 		}
    788 
    789 		/* uoptimize */
    790 		if (ndone == 0)
    791 			continue;
    792 
    793 		/* iterate over the results */
    794 		for (curev = pu->pu_evs; ndone--; curev++) {
    795 			int what;
    796 
    797 #if 0
    798 			/* get & possibly dispatch events from kernel */
    799 			if (curev->ident == puffsfd) {
    800 				if (puffs_req_handle(pgr, ppr, 0) == -1)
    801 					goto out;
    802 				continue;
    803 			}
    804 #endif
    805 
    806 			fio = (void *)curev->udata;
    807 			pfctrl = fio->fctrl;
    808 			if (curev->flags & EV_ERROR) {
    809 				assert(curev->filter == EVFILT_WRITE);
    810 				fio->stat &= ~FIO_WR;
    811 
    812 				/* XXX: how to know if it's a transient error */
    813 				puffs_framev_writeclose(pu, fio,
    814 				    (int)curev->data);
    815 				puffs_framev_notify(fio, PUFFS_FBIO_ERROR);
    816 				continue;
    817 			}
    818 
    819 			what = 0;
    820 			if (curev->filter == EVFILT_READ) {
    821 				puffs_framev_input(pu, pfctrl, fio);
    822 				what |= PUFFS_FBIO_READ;
    823 			}
    824 
    825 			else if (curev->filter == EVFILT_WRITE) {
    826 				puffs_framev_output(pu, pfctrl, fio);
    827 				what |= PUFFS_FBIO_WRITE;
    828 			}
    829 			if (what)
    830 				puffs_framev_notify(fio, what);
    831 		}
    832 
    833 		/*
    834 		 * Schedule continuations.
    835 		 */
    836 		while ((pcc = TAILQ_FIRST(&pu->pu_sched)) != NULL) {
    837 			TAILQ_REMOVE(&pu->pu_sched, pcc, entries);
    838 			puffs_goto(pcc);
    839 		}
    840 
    841 		/*
    842 		 * Really free fd's now that we don't have references
    843 		 * to them.
    844 		 */
    845 		while ((fio = LIST_FIRST(&pu->pu_ios_rmlist)) != NULL) {
    846 			LIST_REMOVE(fio, fio_entries);
    847 			free(fio);
    848 		}
    849 	}
    850 	finalpush(pu);
    851 	errno = 0;
    852 
    853  out:
    854 	/* store the real error for a while */
    855 	sverrno = errno;
    856 
    857 	errno = sverrno;
    858 	if (errno)
    859 		return -1;
    860 	else
    861 		return 0;
    862 }
    863