Home | History | Annotate | Line # | Download | only in puffs
puffs_msgif.c revision 1.80
      1 /*	$NetBSD: puffs_msgif.c,v 1.80 2010/01/14 19:50:07 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 __KERNEL_RCSID(0, "$NetBSD: puffs_msgif.c,v 1.80 2010/01/14 19:50:07 pooka Exp $");
     34 
     35 #include <sys/param.h>
     36 #include <sys/atomic.h>
     37 #include <sys/kmem.h>
     38 #include <sys/kthread.h>
     39 #include <sys/lock.h>
     40 #include <sys/malloc.h>
     41 #include <sys/mount.h>
     42 #include <sys/namei.h>
     43 #include <sys/proc.h>
     44 #include <sys/vnode.h>
     45 #include <sys/atomic.h>
     46 
     47 #include <dev/putter/putter_sys.h>
     48 
     49 #include <fs/puffs/puffs_msgif.h>
     50 #include <fs/puffs/puffs_sys.h>
     51 
     52 #include <miscfs/syncfs/syncfs.h> /* XXX: for syncer_mutex reference */
     53 
     54 /*
     55  * waitq data structures
     56  */
     57 
     58 /*
     59  * While a request is going to userspace, park the caller within the
     60  * kernel.  This is the kernel counterpart of "struct puffs_req".
     61  */
     62 struct puffs_msgpark {
     63 	struct puffs_req	*park_preq;	/* req followed by buf	*/
     64 
     65 	size_t			park_copylen;	/* userspace copylength	*/
     66 	size_t			park_maxlen;	/* max size in comeback */
     67 
     68 	parkdone_fn		park_done;	/* "biodone" a'la puffs	*/
     69 	void			*park_donearg;
     70 
     71 	int			park_flags;
     72 	int			park_refcount;
     73 
     74 	kcondvar_t		park_cv;
     75 	kmutex_t		park_mtx;
     76 
     77 	TAILQ_ENTRY(puffs_msgpark) park_entries;
     78 };
     79 #define PARKFLAG_WAITERGONE	0x01
     80 #define PARKFLAG_DONE		0x02
     81 #define PARKFLAG_ONQUEUE1	0x04
     82 #define PARKFLAG_ONQUEUE2	0x08
     83 #define PARKFLAG_CALL		0x10
     84 #define PARKFLAG_WANTREPLY	0x20
     85 #define	PARKFLAG_HASERROR	0x40
     86 
     87 static pool_cache_t parkpc;
     88 #ifdef PUFFSDEBUG
     89 static int totalpark;
     90 #endif
     91 
     92 static int
     93 makepark(void *arg, void *obj, int flags)
     94 {
     95 	struct puffs_msgpark *park = obj;
     96 
     97 	mutex_init(&park->park_mtx, MUTEX_DEFAULT, IPL_NONE);
     98 	cv_init(&park->park_cv, "puffsrpl");
     99 
    100 	return 0;
    101 }
    102 
    103 static void
    104 nukepark(void *arg, void *obj)
    105 {
    106 	struct puffs_msgpark *park = obj;
    107 
    108 	cv_destroy(&park->park_cv);
    109 	mutex_destroy(&park->park_mtx);
    110 }
    111 
    112 void
    113 puffs_msgif_init(void)
    114 {
    115 
    116 	parkpc = pool_cache_init(sizeof(struct puffs_msgpark), 0, 0, 0,
    117 	    "puffprkl", NULL, IPL_NONE, makepark, nukepark, NULL);
    118 }
    119 
    120 void
    121 puffs_msgif_destroy(void)
    122 {
    123 
    124 	pool_cache_destroy(parkpc);
    125 }
    126 
    127 static int alloced;
    128 
    129 static struct puffs_msgpark *
    130 puffs_msgpark_alloc(int waitok)
    131 {
    132 	struct puffs_msgpark *park;
    133 
    134 	park = pool_cache_get(parkpc, waitok ? PR_WAITOK : PR_NOWAIT);
    135 	if (park == NULL)
    136 		return park;
    137 
    138 	park->park_refcount = 1;
    139 	park->park_preq = NULL;
    140 	park->park_flags = PARKFLAG_WANTREPLY;
    141 
    142 #ifdef PUFFSDEBUG
    143 	totalpark++;
    144 #endif
    145 
    146 	return park;
    147 }
    148 
    149 static void
    150 puffs_msgpark_reference(struct puffs_msgpark *park)
    151 {
    152 
    153 	KASSERT(mutex_owned(&park->park_mtx));
    154 	park->park_refcount++;
    155 }
    156 
    157 /*
    158  * Release reference to park structure.
    159  */
    160 static void
    161 puffs_msgpark_release1(struct puffs_msgpark *park, int howmany)
    162 {
    163 	struct puffs_req *preq = park->park_preq;
    164 	int refcnt;
    165 
    166 	KASSERT(mutex_owned(&park->park_mtx));
    167 	refcnt = park->park_refcount -= howmany;
    168 	mutex_exit(&park->park_mtx);
    169 
    170 	KASSERT(refcnt >= 0);
    171 
    172 	if (refcnt == 0) {
    173 		alloced--;
    174 		if (preq)
    175 			kmem_free(preq, park->park_maxlen);
    176 		pool_cache_put(parkpc, park);
    177 
    178 #ifdef PUFFSDEBUG
    179 		totalpark--;
    180 #endif
    181 	}
    182 }
    183 #define puffs_msgpark_release(a) puffs_msgpark_release1(a, 1)
    184 
    185 #ifdef PUFFSDEBUG
    186 static void
    187 parkdump(struct puffs_msgpark *park)
    188 {
    189 
    190 	DPRINTF(("park %p, preq %p, id %" PRIu64 "\n"
    191 	    "\tcopy %zu, max %zu - done: %p/%p\n"
    192 	    "\tflags 0x%08x, refcount %d, cv/mtx: %p/%p\n",
    193 	    park, park->park_preq, park->park_preq->preq_id,
    194 	    park->park_copylen, park->park_maxlen,
    195 	    park->park_done, park->park_donearg,
    196 	    park->park_flags, park->park_refcount,
    197 	    &park->park_cv, &park->park_mtx));
    198 }
    199 
    200 static void
    201 parkqdump(struct puffs_wq *q, int dumpall)
    202 {
    203 	struct puffs_msgpark *park;
    204 	int total = 0;
    205 
    206 	TAILQ_FOREACH(park, q, park_entries) {
    207 		if (dumpall)
    208 			parkdump(park);
    209 		total++;
    210 	}
    211 	DPRINTF(("puffs waitqueue at %p dumped, %d total\n", q, total));
    212 
    213 }
    214 #endif /* PUFFSDEBUG */
    215 
    216 /*
    217  * A word about locking in the park structures: the lock protects the
    218  * fields of the *park* structure (not preq) and acts as an interlock
    219  * in cv operations.  The lock is always internal to this module and
    220  * callers do not need to worry about it.
    221  */
    222 
    223 int
    224 puffs_msgmem_alloc(size_t len, struct puffs_msgpark **ppark, void **mem,
    225 	int cansleep)
    226 {
    227 	struct puffs_msgpark *park;
    228 	void *m;
    229 
    230 	m = kmem_zalloc(len, cansleep ? KM_SLEEP : KM_NOSLEEP);
    231 	if (m == NULL) {
    232 		KASSERT(cansleep == 0);
    233 		return ENOMEM;
    234 	}
    235 
    236 	park = puffs_msgpark_alloc(cansleep);
    237 	if (park == NULL) {
    238 		KASSERT(cansleep == 0);
    239 		kmem_free(m, len);
    240 		return ENOMEM;
    241 	}
    242 
    243 	park->park_preq = m;
    244 	park->park_maxlen = park->park_copylen = len;
    245 
    246 	*ppark = park;
    247 	*mem = m;
    248 
    249 	return 0;
    250 }
    251 
    252 void
    253 puffs_msgmem_release(struct puffs_msgpark *park)
    254 {
    255 
    256 	if (park == NULL)
    257 		return;
    258 
    259 	mutex_enter(&park->park_mtx);
    260 	puffs_msgpark_release(park);
    261 }
    262 
    263 void
    264 puffs_msg_setfaf(struct puffs_msgpark *park)
    265 {
    266 
    267 	KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
    268 	park->park_flags &= ~PARKFLAG_WANTREPLY;
    269 }
    270 
    271 void
    272 puffs_msg_setdelta(struct puffs_msgpark *park, size_t delta)
    273 {
    274 
    275 	KASSERT(delta < park->park_maxlen); /* "<=" wouldn't make sense */
    276 	park->park_copylen = park->park_maxlen - delta;
    277 }
    278 
    279 void
    280 puffs_msg_setinfo(struct puffs_msgpark *park, int class, int type,
    281 	puffs_cookie_t ck)
    282 {
    283 
    284 	park->park_preq->preq_opclass = PUFFSOP_OPCLASS(class);
    285 	park->park_preq->preq_optype = type;
    286 	park->park_preq->preq_cookie = ck;
    287 }
    288 
    289 void
    290 puffs_msg_setcall(struct puffs_msgpark *park, parkdone_fn donefn, void *donearg)
    291 {
    292 
    293 	KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
    294 	park->park_done = donefn;
    295 	park->park_donearg = donearg;
    296 	park->park_flags |= PARKFLAG_CALL;
    297 }
    298 
    299 /*
    300  * kernel-user-kernel waitqueues
    301  */
    302 
    303 static uint64_t
    304 puffs_getmsgid(struct puffs_mount *pmp)
    305 {
    306 	uint64_t rv;
    307 
    308 	mutex_enter(&pmp->pmp_lock);
    309 	rv = pmp->pmp_nextmsgid++;
    310 	mutex_exit(&pmp->pmp_lock);
    311 
    312 	return rv;
    313 }
    314 
    315 /*
    316  * A word about reference counting of parks.  A reference must be taken
    317  * when accessing a park and additionally when it is on a queue.  So
    318  * when taking it off a queue and releasing the access reference, the
    319  * reference count is generally decremented by 2.
    320  */
    321 
    322 void
    323 puffs_msg_enqueue(struct puffs_mount *pmp, struct puffs_msgpark *park)
    324 {
    325 	struct lwp *l = curlwp;
    326 	struct mount *mp;
    327 	struct puffs_req *preq;
    328 
    329 	mp = PMPTOMP(pmp);
    330 	preq = park->park_preq;
    331 	preq->preq_buflen = park->park_maxlen;
    332 	KASSERT(preq->preq_id == 0
    333 	    || (preq->preq_opclass & PUFFSOPFLAG_ISRESPONSE));
    334 
    335 	if ((park->park_flags & PARKFLAG_WANTREPLY) == 0)
    336 		preq->preq_opclass |= PUFFSOPFLAG_FAF;
    337 	else
    338 		preq->preq_id = puffs_getmsgid(pmp);
    339 
    340 	/* fill in caller information */
    341 	preq->preq_pid = l->l_proc->p_pid;
    342 	preq->preq_lid = l->l_lid;
    343 
    344 	/*
    345 	 * To support cv_sig, yet another movie: check if there are signals
    346 	 * pending and we are issueing a non-FAF.  If so, return an error
    347 	 * directly UNLESS we are issueing INACTIVE/RECLAIM.  In that case,
    348 	 * convert it to a FAF, fire off to the file server and return
    349 	 * an error.  Yes, this is bordering disgusting.  Barfbags are on me.
    350 	 */
    351 	if (__predict_false((park->park_flags & PARKFLAG_WANTREPLY)
    352 	   && (park->park_flags & PARKFLAG_CALL) == 0
    353 	   && (l->l_flag & LW_PENDSIG) != 0 && sigispending(l, 0))) {
    354 		park->park_flags |= PARKFLAG_HASERROR;
    355 		preq->preq_rv = EINTR;
    356 		if (PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN
    357 		    && (preq->preq_optype == PUFFS_VN_INACTIVE
    358 		     || preq->preq_optype == PUFFS_VN_RECLAIM)) {
    359 			park->park_preq->preq_opclass |= PUFFSOPFLAG_FAF;
    360 			park->park_flags &= ~PARKFLAG_WANTREPLY;
    361 			DPRINTF(("puffs_msg_enqueue: converted to FAF %p\n",
    362 			    park));
    363 		} else {
    364 			return;
    365 		}
    366 	}
    367 
    368 	mutex_enter(&pmp->pmp_lock);
    369 	if (pmp->pmp_status != PUFFSTAT_RUNNING) {
    370 		mutex_exit(&pmp->pmp_lock);
    371 		park->park_flags |= PARKFLAG_HASERROR;
    372 		preq->preq_rv = ENXIO;
    373 		return;
    374 	}
    375 
    376 #ifdef PUFFSDEBUG
    377 	parkqdump(&pmp->pmp_msg_touser, puffsdebug > 1);
    378 	parkqdump(&pmp->pmp_msg_replywait, puffsdebug > 1);
    379 #endif
    380 
    381 	/*
    382 	 * Note: we don't need to lock park since we have the only
    383 	 * reference to it at this point.
    384 	 */
    385 	TAILQ_INSERT_TAIL(&pmp->pmp_msg_touser, park, park_entries);
    386 	park->park_flags |= PARKFLAG_ONQUEUE1;
    387 	pmp->pmp_msg_touser_count++;
    388 	park->park_refcount++;
    389 	mutex_exit(&pmp->pmp_lock);
    390 
    391 	cv_broadcast(&pmp->pmp_msg_waiter_cv);
    392 	putter_notify(pmp->pmp_pi);
    393 
    394 	DPRINTF(("touser: req %" PRIu64 ", preq: %p, park: %p, "
    395 	    "c/t: 0x%x/0x%x, f: 0x%x\n", preq->preq_id, preq, park,
    396 	    preq->preq_opclass, preq->preq_optype, park->park_flags));
    397 }
    398 
    399 int
    400 puffs_msg_wait(struct puffs_mount *pmp, struct puffs_msgpark *park)
    401 {
    402 	struct puffs_req *preq = park->park_preq; /* XXX: hmmm */
    403 	int error = 0;
    404 	int rv;
    405 
    406 	mutex_enter(&pmp->pmp_lock);
    407 	puffs_mp_reference(pmp);
    408 	mutex_exit(&pmp->pmp_lock);
    409 
    410 	mutex_enter(&park->park_mtx);
    411 	if ((park->park_flags & PARKFLAG_WANTREPLY) == 0
    412 	    || (park->park_flags & PARKFLAG_CALL)) {
    413 		mutex_exit(&park->park_mtx);
    414 		rv = 0;
    415 		goto skipwait;
    416 	}
    417 
    418 	/* did the response beat us to the wait? */
    419 	if (__predict_false((park->park_flags & PARKFLAG_DONE)
    420 	    || (park->park_flags & PARKFLAG_HASERROR))) {
    421 		rv = park->park_preq->preq_rv;
    422 		mutex_exit(&park->park_mtx);
    423 		goto skipwait;
    424 	}
    425 
    426 	error = cv_wait_sig(&park->park_cv, &park->park_mtx);
    427 	DPRINTF(("puffs_touser: waiter for %p woke up with %d\n",
    428 	    park, error));
    429 	if (error) {
    430 		park->park_flags |= PARKFLAG_WAITERGONE;
    431 		if (park->park_flags & PARKFLAG_DONE) {
    432 			rv = preq->preq_rv;
    433 			mutex_exit(&park->park_mtx);
    434 		} else {
    435 			/*
    436 			 * ok, we marked it as going away, but
    437 			 * still need to do queue ops.  take locks
    438 			 * in correct order.
    439 			 *
    440 			 * We don't want to release our reference
    441 			 * if it's on replywait queue to avoid error
    442 			 * to file server.  putop() code will DTRT.
    443 			 */
    444 			mutex_exit(&park->park_mtx);
    445 			mutex_enter(&pmp->pmp_lock);
    446 			mutex_enter(&park->park_mtx);
    447 
    448 			/*
    449 			 * Still on queue1?  We can safely remove it
    450 			 * without any consequences since the file
    451 			 * server hasn't seen it.  "else" we need to
    452 			 * wait for the response and just ignore it
    453 			 * to avoid signalling an incorrect error to
    454 			 * the file server.
    455 			 */
    456 			if (park->park_flags & PARKFLAG_ONQUEUE1) {
    457 				TAILQ_REMOVE(&pmp->pmp_msg_touser,
    458 				    park, park_entries);
    459 				puffs_msgpark_release(park);
    460 				pmp->pmp_msg_touser_count--;
    461 				park->park_flags &= ~PARKFLAG_ONQUEUE1;
    462 			} else {
    463 				mutex_exit(&park->park_mtx);
    464 			}
    465 			mutex_exit(&pmp->pmp_lock);
    466 
    467 			rv = EINTR;
    468 		}
    469 	} else {
    470 		rv = preq->preq_rv;
    471 		mutex_exit(&park->park_mtx);
    472 	}
    473 
    474  skipwait:
    475 	mutex_enter(&pmp->pmp_lock);
    476 	puffs_mp_release(pmp);
    477 	mutex_exit(&pmp->pmp_lock);
    478 
    479 	return rv;
    480 }
    481 
    482 /*
    483  * XXX: this suuuucks.  Hopefully I'll get rid of this lossage once
    484  * the whole setback-nonsense gets fixed.
    485  */
    486 int
    487 puffs_msg_wait2(struct puffs_mount *pmp, struct puffs_msgpark *park,
    488 	struct puffs_node *pn1, struct puffs_node *pn2)
    489 {
    490 	struct puffs_req *preq;
    491 	int rv;
    492 
    493 	rv = puffs_msg_wait(pmp, park);
    494 
    495 	preq = park->park_preq;
    496 	if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N1)
    497 		pn1->pn_stat |= PNODE_DOINACT;
    498 	if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N2)
    499 		pn2->pn_stat |= PNODE_DOINACT;
    500 
    501 	if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N1)
    502 		pn1->pn_stat |= PNODE_NOREFS;
    503 	if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N2)
    504 		pn2->pn_stat |= PNODE_NOREFS;
    505 
    506 	return rv;
    507 
    508 }
    509 
    510 /*
    511  * XXX: lazy bum.  please, for the love of foie gras, fix me.
    512  * This should *NOT* depend on setfaf.  Also "memcpy" could
    513  * be done more nicely.
    514  */
    515 void
    516 puffs_msg_sendresp(struct puffs_mount *pmp, struct puffs_req *origpreq, int rv)
    517 {
    518 	struct puffs_msgpark *park;
    519 	struct puffs_req *preq;
    520 
    521 	puffs_msgmem_alloc(sizeof(struct puffs_req), &park, (void *)&preq, 1);
    522 	puffs_msg_setfaf(park); /* XXXXXX: avoids reqid override */
    523 
    524 	memcpy(preq, origpreq, sizeof(struct puffs_req));
    525 	preq->preq_rv = rv;
    526 	preq->preq_opclass |= PUFFSOPFLAG_ISRESPONSE;
    527 
    528 	puffs_msg_enqueue(pmp, park);
    529 	puffs_msgmem_release(park);
    530 }
    531 
    532 /*
    533  * Get next request in the outgoing queue.  "maxsize" controls the
    534  * size the caller can accommodate and "nonblock" signals if this
    535  * should block while waiting for input.  Handles all locking internally.
    536  */
    537 int
    538 puffs_msgif_getout(void *this, size_t maxsize, int nonblock,
    539 	uint8_t **data, size_t *dlen, void **parkptr)
    540 {
    541 	struct puffs_mount *pmp = this;
    542 	struct puffs_msgpark *park;
    543 	struct puffs_req *preq;
    544 	int error;
    545 
    546 	error = 0;
    547 	mutex_enter(&pmp->pmp_lock);
    548 	puffs_mp_reference(pmp);
    549 	for (;;) {
    550 		/* RIP? */
    551 		if (pmp->pmp_status != PUFFSTAT_RUNNING) {
    552 			error = ENXIO;
    553 			break;
    554 		}
    555 
    556 		/* need platinum yendorian express card? */
    557 		if (TAILQ_EMPTY(&pmp->pmp_msg_touser)) {
    558 			DPRINTF(("puffs_getout: no outgoing op, "));
    559 			if (nonblock) {
    560 				DPRINTF(("returning EWOULDBLOCK\n"));
    561 				error = EWOULDBLOCK;
    562 				break;
    563 			}
    564 			DPRINTF(("waiting ...\n"));
    565 
    566 			error = cv_wait_sig(&pmp->pmp_msg_waiter_cv,
    567 			    &pmp->pmp_lock);
    568 			if (error)
    569 				break;
    570 			else
    571 				continue;
    572 		}
    573 
    574 		park = TAILQ_FIRST(&pmp->pmp_msg_touser);
    575 		if (park == NULL)
    576 			continue;
    577 
    578 		mutex_enter(&park->park_mtx);
    579 		puffs_msgpark_reference(park);
    580 
    581 		DPRINTF(("puffs_getout: found park at %p, ", park));
    582 
    583 		/* If it's a goner, don't process any furher */
    584 		if (park->park_flags & PARKFLAG_WAITERGONE) {
    585 			DPRINTF(("waitergone!\n"));
    586 			puffs_msgpark_release(park);
    587 			continue;
    588 		}
    589 		preq = park->park_preq;
    590 
    591 #if 0
    592 		/* check size */
    593 		/*
    594 		 * XXX: this check is not valid for now, we don't know
    595 		 * the size of the caller's input buffer.  i.e. this
    596 		 * will most likely go away
    597 		 */
    598 		if (maxsize < preq->preq_frhdr.pfr_len) {
    599 			DPRINTF(("buffer too small\n"));
    600 			puffs_msgpark_release(park);
    601 			error = E2BIG;
    602 			break;
    603 		}
    604 #endif
    605 
    606 		DPRINTF(("returning\n"));
    607 
    608 		/*
    609 		 * Ok, we found what we came for.  Release it from the
    610 		 * outgoing queue but do not unlock.  We will unlock
    611 		 * only after we "releaseout" it to avoid complications:
    612 		 * otherwise it is (theoretically) possible for userland
    613 		 * to race us into "put" before we have a change to put
    614 		 * this baby on the receiving queue.
    615 		 */
    616 		TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries);
    617 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
    618 		park->park_flags &= ~PARKFLAG_ONQUEUE1;
    619 		mutex_exit(&park->park_mtx);
    620 
    621 		pmp->pmp_msg_touser_count--;
    622 		KASSERT(pmp->pmp_msg_touser_count >= 0);
    623 
    624 		break;
    625 	}
    626 	puffs_mp_release(pmp);
    627 	mutex_exit(&pmp->pmp_lock);
    628 
    629 	if (error == 0) {
    630 		*data = (uint8_t *)preq;
    631 		preq->preq_pth.pth_framelen = park->park_copylen;
    632 		*dlen = preq->preq_pth.pth_framelen;
    633 		*parkptr = park;
    634 	}
    635 
    636 	return error;
    637 }
    638 
    639 /*
    640  * Release outgoing structure.  Now, depending on the success of the
    641  * outgoing send, it is either going onto the result waiting queue
    642  * or the death chamber.
    643  */
    644 void
    645 puffs_msgif_releaseout(void *this, void *parkptr, int status)
    646 {
    647 	struct puffs_mount *pmp = this;
    648 	struct puffs_msgpark *park = parkptr;
    649 
    650 	DPRINTF(("puffs_releaseout: returning park %p, errno %d: " ,
    651 	    park, status));
    652 	mutex_enter(&pmp->pmp_lock);
    653 	mutex_enter(&park->park_mtx);
    654 	if (park->park_flags & PARKFLAG_WANTREPLY) {
    655 		if (status == 0) {
    656 			DPRINTF(("enqueue replywait\n"));
    657 			TAILQ_INSERT_TAIL(&pmp->pmp_msg_replywait, park,
    658 			    park_entries);
    659 			park->park_flags |= PARKFLAG_ONQUEUE2;
    660 		} else {
    661 			DPRINTF(("error path!\n"));
    662 			park->park_preq->preq_rv = status;
    663 			park->park_flags |= PARKFLAG_DONE;
    664 			cv_signal(&park->park_cv);
    665 		}
    666 		puffs_msgpark_release(park);
    667 	} else {
    668 		DPRINTF(("release\n"));
    669 		puffs_msgpark_release1(park, 2);
    670 	}
    671 	mutex_exit(&pmp->pmp_lock);
    672 }
    673 
    674 size_t
    675 puffs_msgif_waitcount(void *this)
    676 {
    677 	struct puffs_mount *pmp = this;
    678 	size_t rv;
    679 
    680 	mutex_enter(&pmp->pmp_lock);
    681 	rv = pmp->pmp_msg_touser_count;
    682 	mutex_exit(&pmp->pmp_lock);
    683 
    684 	return rv;
    685 }
    686 
    687 /*
    688  * XXX: locking with this one?
    689  */
    690 static void
    691 puffsop_msg(void *this, struct puffs_req *preq)
    692 {
    693 	struct puffs_mount *pmp = this;
    694 	struct putter_hdr *pth = &preq->preq_pth;
    695 	struct puffs_msgpark *park;
    696 	int wgone;
    697 
    698 	mutex_enter(&pmp->pmp_lock);
    699 
    700 	/* Locate waiter */
    701 	TAILQ_FOREACH(park, &pmp->pmp_msg_replywait, park_entries) {
    702 		if (park->park_preq->preq_id == preq->preq_id)
    703 			break;
    704 	}
    705 	if (park == NULL) {
    706 		DPRINTF(("puffsop_msg: no request: %" PRIu64 "\n",
    707 		    preq->preq_id));
    708 		mutex_exit(&pmp->pmp_lock);
    709 		return; /* XXX send error */
    710 	}
    711 
    712 	mutex_enter(&park->park_mtx);
    713 	puffs_msgpark_reference(park);
    714 	if (pth->pth_framelen > park->park_maxlen) {
    715 		DPRINTF(("puffsop_msg: invalid buffer length: "
    716 		    "%" PRIu64 " (req %" PRIu64 ", \n", pth->pth_framelen,
    717 		    preq->preq_id));
    718 		park->park_preq->preq_rv = EPROTO;
    719 		cv_signal(&park->park_cv);
    720 		puffs_msgpark_release1(park, 2);
    721 		mutex_exit(&pmp->pmp_lock);
    722 		return; /* XXX: error */
    723 	}
    724 	wgone = park->park_flags & PARKFLAG_WAITERGONE;
    725 
    726 	KASSERT(park->park_flags & PARKFLAG_ONQUEUE2);
    727 	TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries);
    728 	park->park_flags &= ~PARKFLAG_ONQUEUE2;
    729 	mutex_exit(&pmp->pmp_lock);
    730 
    731 	if (wgone) {
    732 		DPRINTF(("puffsop_msg: bad service - waiter gone for "
    733 		    "park %p\n", park));
    734 	} else {
    735 		if (park->park_flags & PARKFLAG_CALL) {
    736 			DPRINTF(("puffsop_msg: call for %p, arg %p\n",
    737 			    park->park_preq, park->park_donearg));
    738 			park->park_done(pmp, preq, park->park_donearg);
    739 		} else {
    740 			/* XXX: yes, I know */
    741 			memcpy(park->park_preq, preq, pth->pth_framelen);
    742 		}
    743 	}
    744 
    745 	if (!wgone) {
    746 		DPRINTF(("puffs_putop: flagging done for "
    747 		    "park %p\n", park));
    748 		cv_signal(&park->park_cv);
    749 	}
    750 
    751 	park->park_flags |= PARKFLAG_DONE;
    752 	puffs_msgpark_release1(park, 2);
    753 }
    754 
    755 static void
    756 puffsop_flush(struct puffs_mount *pmp, struct puffs_flush *pf)
    757 {
    758 	struct vnode *vp;
    759 	voff_t offlo, offhi;
    760 	int rv, flags = 0;
    761 
    762 	KASSERT(pf->pf_req.preq_pth.pth_framelen == sizeof(struct puffs_flush));
    763 
    764 	/* XXX: slurry */
    765 	if (pf->pf_op == PUFFS_INVAL_NAMECACHE_ALL) {
    766 		cache_purgevfs(PMPTOMP(pmp));
    767 		rv = 0;
    768 		goto out;
    769 	}
    770 
    771 	/*
    772 	 * Get vnode, don't lock it.  Namecache is protected by its own lock
    773 	 * and we have a reference to protect against premature harvesting.
    774 	 *
    775 	 * The node we want here might be locked and the op is in
    776 	 * userspace waiting for us to complete ==> deadlock.  Another
    777 	 * reason we need to eventually bump locking to userspace, as we
    778 	 * will need to lock the node if we wish to do flushes.
    779 	 */
    780 	rv = puffs_cookie2vnode(pmp, pf->pf_cookie, 0, 0, &vp);
    781 	if (rv) {
    782 		if (rv == PUFFS_NOSUCHCOOKIE)
    783 			rv = ENOENT;
    784 		goto out;
    785 	}
    786 
    787 	switch (pf->pf_op) {
    788 #if 0
    789 	/* not quite ready, yet */
    790 	case PUFFS_INVAL_NAMECACHE_NODE:
    791 	struct componentname *pf_cn;
    792 	char *name;
    793 		/* get comfortab^Wcomponentname */
    794 		pf_cn = kmem_alloc(componentname);
    795 		memset(pf_cn, 0, sizeof(struct componentname));
    796 		break;
    797 
    798 #endif
    799 	case PUFFS_INVAL_NAMECACHE_DIR:
    800 		if (vp->v_type != VDIR) {
    801 			rv = EINVAL;
    802 			break;
    803 		}
    804 		cache_purge1(vp, NULL, PURGE_CHILDREN);
    805 		break;
    806 
    807 	case PUFFS_INVAL_PAGECACHE_NODE_RANGE:
    808 		flags = PGO_FREE;
    809 		/*FALLTHROUGH*/
    810 	case PUFFS_FLUSH_PAGECACHE_NODE_RANGE:
    811 		if (flags == 0)
    812 			flags = PGO_CLEANIT;
    813 
    814 		if (pf->pf_end > vp->v_size || vp->v_type != VREG) {
    815 			rv = EINVAL;
    816 			break;
    817 		}
    818 
    819 		offlo = trunc_page(pf->pf_start);
    820 		offhi = round_page(pf->pf_end);
    821 		if (offhi != 0 && offlo >= offhi) {
    822 			rv = EINVAL;
    823 			break;
    824 		}
    825 
    826 		mutex_enter(&vp->v_uobj.vmobjlock);
    827 		rv = VOP_PUTPAGES(vp, offlo, offhi, flags);
    828 		break;
    829 
    830 	default:
    831 		rv = EINVAL;
    832 	}
    833 
    834 	vrele(vp);
    835 
    836  out:
    837 	puffs_msg_sendresp(pmp, &pf->pf_req, rv);
    838 }
    839 
    840 int
    841 puffs_msgif_dispatch(void *this, struct putter_hdr *pth)
    842 {
    843 	struct puffs_mount *pmp = this;
    844 	struct puffs_req *preq = (struct puffs_req *)pth;
    845 	struct puffs_sopreq *psopr;
    846 
    847 	if (pth->pth_framelen < sizeof(struct puffs_req)) {
    848 		puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */
    849 		return 0;
    850 	}
    851 
    852 	switch (PUFFSOP_OPCLASS(preq->preq_opclass)) {
    853 	case PUFFSOP_VN:
    854 	case PUFFSOP_VFS:
    855 		DPRINTF(("dispatch: vn/vfs message 0x%x\n", preq->preq_optype));
    856 		puffsop_msg(pmp, preq);
    857 		break;
    858 
    859 	case PUFFSOP_FLUSH: /* process in sop thread */
    860 	{
    861 		struct puffs_flush *pf;
    862 
    863 		DPRINTF(("dispatch: flush 0x%x\n", preq->preq_optype));
    864 
    865 		if (preq->preq_pth.pth_framelen != sizeof(struct puffs_flush)) {
    866 			puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */
    867 			break;
    868 		}
    869 		pf = (struct puffs_flush *)preq;
    870 
    871 		psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP);
    872 		memcpy(&psopr->psopr_pf, pf, sizeof(*pf));
    873 		psopr->psopr_sopreq = PUFFS_SOPREQ_FLUSH;
    874 
    875 		mutex_enter(&pmp->pmp_sopmtx);
    876 		if (pmp->pmp_sopthrcount == 0) {
    877 			mutex_exit(&pmp->pmp_sopmtx);
    878 			kmem_free(psopr, sizeof(*psopr));
    879 			puffs_msg_sendresp(pmp, preq, ENXIO);
    880 		} else {
    881 			TAILQ_INSERT_TAIL(&pmp->pmp_sopreqs,
    882 			    psopr, psopr_entries);
    883 			cv_signal(&pmp->pmp_sopcv);
    884 			mutex_exit(&pmp->pmp_sopmtx);
    885 		}
    886 		break;
    887 	}
    888 
    889 	case PUFFSOP_UNMOUNT: /* process in sop thread */
    890 	{
    891 
    892 		DPRINTF(("dispatch: unmount 0x%x\n", preq->preq_optype));
    893 
    894 		psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP);
    895 		psopr->psopr_preq = *preq;
    896 		psopr->psopr_sopreq = PUFFS_SOPREQ_UNMOUNT;
    897 
    898 		mutex_enter(&pmp->pmp_sopmtx);
    899 		if (pmp->pmp_sopthrcount == 0) {
    900 			mutex_exit(&pmp->pmp_sopmtx);
    901 			kmem_free(psopr, sizeof(*psopr));
    902 			puffs_msg_sendresp(pmp, preq, ENXIO);
    903 		} else {
    904 			TAILQ_INSERT_TAIL(&pmp->pmp_sopreqs,
    905 			    psopr, psopr_entries);
    906 			cv_signal(&pmp->pmp_sopcv);
    907 			mutex_exit(&pmp->pmp_sopmtx);
    908 		}
    909 		break;
    910 	}
    911 
    912 	default:
    913 		DPRINTF(("dispatch: invalid class 0x%x\n", preq->preq_opclass));
    914 		puffs_msg_sendresp(pmp, preq, EOPNOTSUPP);
    915 		break;
    916 	}
    917 
    918 	return 0;
    919 }
    920 
    921 /*
    922  * Work loop for thread processing all ops from server which
    923  * cannot safely be handled in caller context.  This includes
    924  * everything which might need a lock currently "held" by the file
    925  * server, i.e. a long-term kernel lock which will be released only
    926  * once the file server acknowledges a request
    927  */
    928 void
    929 puffs_sop_thread(void *arg)
    930 {
    931 	struct puffs_mount *pmp = arg;
    932 	struct mount *mp = PMPTOMP(pmp);
    933 	struct puffs_sopreq *psopr;
    934 	bool keeprunning;
    935 	bool unmountme = false;
    936 
    937 	mutex_enter(&pmp->pmp_sopmtx);
    938 	for (keeprunning = true; keeprunning; ) {
    939 		while ((psopr = TAILQ_FIRST(&pmp->pmp_sopreqs)) == NULL)
    940 			cv_wait(&pmp->pmp_sopcv, &pmp->pmp_sopmtx);
    941 		TAILQ_REMOVE(&pmp->pmp_sopreqs, psopr, psopr_entries);
    942 		mutex_exit(&pmp->pmp_sopmtx);
    943 
    944 		switch (psopr->psopr_sopreq) {
    945 		case PUFFS_SOPREQSYS_EXIT:
    946 			keeprunning = false;
    947 			break;
    948 		case PUFFS_SOPREQ_FLUSH:
    949 			puffsop_flush(pmp, &psopr->psopr_pf);
    950 			break;
    951 		case PUFFS_SOPREQ_UNMOUNT:
    952 			puffs_msg_sendresp(pmp, &psopr->psopr_preq, 0);
    953 
    954 			unmountme = true;
    955 			keeprunning = false;
    956 
    957 			/*
    958 			 * We know the mountpoint is still alive because
    959 			 * the thread that is us (poetic?) is still alive.
    960 			 */
    961 			atomic_inc_uint((unsigned int*)&mp->mnt_refcnt);
    962 			break;
    963 		}
    964 
    965 		kmem_free(psopr, sizeof(*psopr));
    966 		mutex_enter(&pmp->pmp_sopmtx);
    967 	}
    968 
    969 	/*
    970 	 * Purge remaining ops.
    971 	 */
    972 	while ((psopr = TAILQ_FIRST(&pmp->pmp_sopreqs)) != NULL) {
    973 		TAILQ_REMOVE(&pmp->pmp_sopreqs, psopr, psopr_entries);
    974 		mutex_exit(&pmp->pmp_sopmtx);
    975 		puffs_msg_sendresp(pmp, &psopr->psopr_preq, ENXIO);
    976 		kmem_free(psopr, sizeof(*psopr));
    977 		mutex_enter(&pmp->pmp_sopmtx);
    978 	}
    979 
    980 	pmp->pmp_sopthrcount--;
    981 	cv_broadcast(&pmp->pmp_sopcv);
    982 	mutex_exit(&pmp->pmp_sopmtx); /* not allowed to access fs after this */
    983 
    984 	/*
    985 	 * If unmount was requested, we can now safely do it here, since
    986 	 * our context is dead from the point-of-view of puffs_unmount()
    987 	 * and we are just another thread.  dounmount() makes internally
    988 	 * sure that VFS_UNMOUNT() isn't called reentrantly and that it
    989 	 * is eventually completed.
    990 	 */
    991 	if (unmountme) {
    992 		(void)dounmount(mp, MNT_FORCE, curlwp);
    993 		vfs_destroy(mp);
    994 	}
    995 
    996 	kthread_exit(0);
    997 }
    998 
    999 int
   1000 puffs_msgif_close(void *this)
   1001 {
   1002 	struct puffs_mount *pmp = this;
   1003 	struct mount *mp = PMPTOMP(pmp);
   1004 
   1005 	mutex_enter(&pmp->pmp_lock);
   1006 	puffs_mp_reference(pmp);
   1007 
   1008 	/*
   1009 	 * Free the waiting callers before proceeding any further.
   1010 	 * The syncer might be jogging around in this file system
   1011 	 * currently.  If we allow it to go to the userspace of no
   1012 	 * return while trying to get the syncer lock, well ...
   1013 	 */
   1014 	puffs_userdead(pmp);
   1015 
   1016 	/*
   1017 	 * Make sure someone from puffs_unmount() isn't currently in
   1018 	 * userspace.  If we don't take this precautionary step,
   1019 	 * they might notice that the mountpoint has disappeared
   1020 	 * from under them once they return.  Especially note that we
   1021 	 * cannot simply test for an unmounter before calling
   1022 	 * dounmount(), since it might be possible that that particular
   1023 	 * invocation of unmount was called without MNT_FORCE.  Here we
   1024 	 * *must* make sure unmount succeeds.  Also, restart is necessary
   1025 	 * since pmp isn't locked.  We might end up with PUTTER_DEAD after
   1026 	 * restart and exit from there.
   1027 	 */
   1028 	if (pmp->pmp_unmounting) {
   1029 		cv_wait(&pmp->pmp_unmounting_cv, &pmp->pmp_lock);
   1030 		puffs_mp_release(pmp);
   1031 		mutex_exit(&pmp->pmp_lock);
   1032 		DPRINTF(("puffs_fop_close: unmount was in progress for pmp %p, "
   1033 		    "restart\n", pmp));
   1034 		return ERESTART;
   1035 	}
   1036 
   1037 	/* Won't access pmp from here anymore */
   1038 	atomic_inc_uint((unsigned int*)&mp->mnt_refcnt);
   1039 	puffs_mp_release(pmp);
   1040 	mutex_exit(&pmp->pmp_lock);
   1041 
   1042 	/* Detach from VFS. */
   1043 	(void)dounmount(mp, MNT_FORCE, curlwp);
   1044 	vfs_destroy(mp);
   1045 
   1046 	return 0;
   1047 }
   1048 
   1049 /*
   1050  * We're dead, kaput, RIP, slightly more than merely pining for the
   1051  * fjords, belly-up, fallen, lifeless, finished, expired, gone to meet
   1052  * our maker, ceased to be, etcetc.  YASD.  It's a dead FS!
   1053  *
   1054  * Caller must hold puffs mutex.
   1055  */
   1056 void
   1057 puffs_userdead(struct puffs_mount *pmp)
   1058 {
   1059 	struct puffs_msgpark *park, *park_next;
   1060 
   1061 	/*
   1062 	 * Mark filesystem status as dying so that operations don't
   1063 	 * attempt to march to userspace any longer.
   1064 	 */
   1065 	pmp->pmp_status = PUFFSTAT_DYING;
   1066 
   1067 	/* signal waiters on REQUEST TO file server queue */
   1068 	for (park = TAILQ_FIRST(&pmp->pmp_msg_touser); park; park = park_next) {
   1069 		uint8_t opclass;
   1070 
   1071 		mutex_enter(&park->park_mtx);
   1072 		puffs_msgpark_reference(park);
   1073 		park_next = TAILQ_NEXT(park, park_entries);
   1074 
   1075 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
   1076 		TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries);
   1077 		park->park_flags &= ~PARKFLAG_ONQUEUE1;
   1078 		pmp->pmp_msg_touser_count--;
   1079 
   1080 		/*
   1081 		 * Even though waiters on QUEUE1 are removed in touser()
   1082 		 * in case of WAITERGONE, it is still possible for us to
   1083 		 * get raced here due to having to retake locks in said
   1084 		 * touser().  In the race case simply "ignore" the item
   1085 		 * on the queue and move on to the next one.
   1086 		 */
   1087 		if (park->park_flags & PARKFLAG_WAITERGONE) {
   1088 			KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
   1089 			KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
   1090 			puffs_msgpark_release(park);
   1091 
   1092 		} else {
   1093 			opclass = park->park_preq->preq_opclass;
   1094 			park->park_preq->preq_rv = ENXIO;
   1095 
   1096 			if (park->park_flags & PARKFLAG_CALL) {
   1097 				park->park_done(pmp, park->park_preq,
   1098 				    park->park_donearg);
   1099 				puffs_msgpark_release1(park, 2);
   1100 			} else if ((park->park_flags & PARKFLAG_WANTREPLY)==0) {
   1101 				puffs_msgpark_release1(park, 2);
   1102 			} else {
   1103 				park->park_preq->preq_rv = ENXIO;
   1104 				cv_signal(&park->park_cv);
   1105 				puffs_msgpark_release(park);
   1106 			}
   1107 		}
   1108 	}
   1109 
   1110 	/* signal waiters on RESPONSE FROM file server queue */
   1111 	for (park=TAILQ_FIRST(&pmp->pmp_msg_replywait); park; park=park_next) {
   1112 		mutex_enter(&park->park_mtx);
   1113 		puffs_msgpark_reference(park);
   1114 		park_next = TAILQ_NEXT(park, park_entries);
   1115 
   1116 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE2);
   1117 		KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
   1118 
   1119 		TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries);
   1120 		park->park_flags &= ~PARKFLAG_ONQUEUE2;
   1121 
   1122 		if (park->park_flags & PARKFLAG_WAITERGONE) {
   1123 			KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
   1124 			puffs_msgpark_release(park);
   1125 		} else {
   1126 			park->park_preq->preq_rv = ENXIO;
   1127 			if (park->park_flags & PARKFLAG_CALL) {
   1128 				park->park_done(pmp, park->park_preq,
   1129 				    park->park_donearg);
   1130 				puffs_msgpark_release1(park, 2);
   1131 			} else {
   1132 				cv_signal(&park->park_cv);
   1133 				puffs_msgpark_release(park);
   1134 			}
   1135 		}
   1136 	}
   1137 
   1138 	cv_broadcast(&pmp->pmp_msg_waiter_cv);
   1139 }
   1140