xref: /netbsd-src/sys/fs/puffs/puffs_msgif.c (revision a5847cc334d9a7029f6352b847e9e8d71a0f9e0c)
1 /*	$NetBSD: puffs_msgif.c,v 1.89 2011/10/19 01:39:29 manu 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 1.89 2011/10/19 01:39:29 manu 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 <uvm/uvm.h>
48 
49 #include <dev/putter/putter_sys.h>
50 
51 #include <fs/puffs/puffs_msgif.h>
52 #include <fs/puffs/puffs_sys.h>
53 
54 #include <miscfs/syncfs/syncfs.h> /* XXX: for syncer_mutex reference */
55 
56 /*
57  * waitq data structures
58  */
59 
60 /*
61  * While a request is going to userspace, park the caller within the
62  * kernel.  This is the kernel counterpart of "struct puffs_req".
63  */
64 struct puffs_msgpark {
65 	struct puffs_req	*park_preq;	/* req followed by buf	*/
66 
67 	size_t			park_copylen;	/* userspace copylength	*/
68 	size_t			park_maxlen;	/* max size in comeback */
69 
70 	struct puffs_req	*park_creq;	/* non-compat preq	*/
71 	size_t			park_creqlen;	/* non-compat preq len	*/
72 
73 	parkdone_fn		park_done;	/* "biodone" a'la puffs	*/
74 	void			*park_donearg;
75 
76 	int			park_flags;
77 	int			park_refcount;
78 
79 	kcondvar_t		park_cv;
80 	kmutex_t		park_mtx;
81 
82 	TAILQ_ENTRY(puffs_msgpark) park_entries;
83 };
84 #define PARKFLAG_WAITERGONE	0x01
85 #define PARKFLAG_DONE		0x02
86 #define PARKFLAG_ONQUEUE1	0x04
87 #define PARKFLAG_ONQUEUE2	0x08
88 #define PARKFLAG_CALL		0x10
89 #define PARKFLAG_WANTREPLY	0x20
90 #define	PARKFLAG_HASERROR	0x40
91 
92 static pool_cache_t parkpc;
93 #ifdef PUFFSDEBUG
94 static int totalpark;
95 #endif
96 
97 static int
98 makepark(void *arg, void *obj, int flags)
99 {
100 	struct puffs_msgpark *park = obj;
101 
102 	mutex_init(&park->park_mtx, MUTEX_DEFAULT, IPL_NONE);
103 	cv_init(&park->park_cv, "puffsrpl");
104 
105 	return 0;
106 }
107 
108 static void
109 nukepark(void *arg, void *obj)
110 {
111 	struct puffs_msgpark *park = obj;
112 
113 	cv_destroy(&park->park_cv);
114 	mutex_destroy(&park->park_mtx);
115 }
116 
117 void
118 puffs_msgif_init(void)
119 {
120 
121 	parkpc = pool_cache_init(sizeof(struct puffs_msgpark), 0, 0, 0,
122 	    "puffprkl", NULL, IPL_NONE, makepark, nukepark, NULL);
123 }
124 
125 void
126 puffs_msgif_destroy(void)
127 {
128 
129 	pool_cache_destroy(parkpc);
130 }
131 
132 static struct puffs_msgpark *
133 puffs_msgpark_alloc(int waitok)
134 {
135 	struct puffs_msgpark *park;
136 
137 	KASSERT(curlwp != uvm.pagedaemon_lwp || !waitok);
138 
139 	park = pool_cache_get(parkpc, waitok ? PR_WAITOK : PR_NOWAIT);
140 	if (park == NULL)
141 		return park;
142 
143 	park->park_refcount = 1;
144 	park->park_preq = park->park_creq = NULL;
145 	park->park_flags = PARKFLAG_WANTREPLY;
146 
147 #ifdef PUFFSDEBUG
148 	totalpark++;
149 #endif
150 
151 	return park;
152 }
153 
154 static void
155 puffs_msgpark_reference(struct puffs_msgpark *park)
156 {
157 
158 	KASSERT(mutex_owned(&park->park_mtx));
159 	park->park_refcount++;
160 }
161 
162 /*
163  * Release reference to park structure.
164  */
165 static void
166 puffs_msgpark_release1(struct puffs_msgpark *park, int howmany)
167 {
168 	struct puffs_req *preq = park->park_preq;
169 	struct puffs_req *creq = park->park_creq;
170 	int refcnt;
171 
172 	KASSERT(mutex_owned(&park->park_mtx));
173 	refcnt = park->park_refcount -= howmany;
174 	mutex_exit(&park->park_mtx);
175 
176 	KASSERT(refcnt >= 0);
177 
178 	if (refcnt == 0) {
179 		if (preq)
180 			kmem_free(preq, park->park_maxlen);
181 #if 1
182 		if (creq)
183 			kmem_free(creq, park->park_creqlen);
184 #endif
185 		pool_cache_put(parkpc, park);
186 
187 #ifdef PUFFSDEBUG
188 		totalpark--;
189 #endif
190 	}
191 }
192 #define puffs_msgpark_release(a) puffs_msgpark_release1(a, 1)
193 
194 #ifdef PUFFSDEBUG
195 static void
196 parkdump(struct puffs_msgpark *park)
197 {
198 
199 	DPRINTF(("park %p, preq %p, id %" PRIu64 "\n"
200 	    "\tcopy %zu, max %zu - done: %p/%p\n"
201 	    "\tflags 0x%08x, refcount %d, cv/mtx: %p/%p\n",
202 	    park, park->park_preq, park->park_preq->preq_id,
203 	    park->park_copylen, park->park_maxlen,
204 	    park->park_done, park->park_donearg,
205 	    park->park_flags, park->park_refcount,
206 	    &park->park_cv, &park->park_mtx));
207 }
208 
209 static void
210 parkqdump(struct puffs_wq *q, int dumpall)
211 {
212 	struct puffs_msgpark *park;
213 	int total = 0;
214 
215 	TAILQ_FOREACH(park, q, park_entries) {
216 		if (dumpall)
217 			parkdump(park);
218 		total++;
219 	}
220 	DPRINTF(("puffs waitqueue at %p dumped, %d total\n", q, total));
221 
222 }
223 #endif /* PUFFSDEBUG */
224 
225 /*
226  * A word about locking in the park structures: the lock protects the
227  * fields of the *park* structure (not preq) and acts as an interlock
228  * in cv operations.  The lock is always internal to this module and
229  * callers do not need to worry about it.
230  */
231 
232 int
233 puffs_msgmem_alloc(size_t len, struct puffs_msgpark **ppark, void **mem,
234 	int cansleep)
235 {
236 	struct puffs_msgpark *park;
237 	void *m;
238 
239 	KASSERT(curlwp != uvm.pagedaemon_lwp || !cansleep);
240 	m = kmem_zalloc(len, cansleep ? KM_SLEEP : KM_NOSLEEP);
241 	if (m == NULL) {
242 		KASSERT(cansleep == 0);
243 		return ENOMEM;
244 	}
245 
246 	park = puffs_msgpark_alloc(cansleep);
247 	if (park == NULL) {
248 		KASSERT(cansleep == 0);
249 		kmem_free(m, len);
250 		return ENOMEM;
251 	}
252 
253 	park->park_preq = m;
254 	park->park_maxlen = park->park_copylen = len;
255 
256 	*ppark = park;
257 	*mem = m;
258 
259 	return 0;
260 }
261 
262 void
263 puffs_msgmem_release(struct puffs_msgpark *park)
264 {
265 
266 	if (park == NULL)
267 		return;
268 
269 	mutex_enter(&park->park_mtx);
270 	puffs_msgpark_release(park);
271 }
272 
273 void
274 puffs_msg_setfaf(struct puffs_msgpark *park)
275 {
276 
277 	KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
278 	park->park_flags &= ~PARKFLAG_WANTREPLY;
279 }
280 
281 void
282 puffs_msg_setdelta(struct puffs_msgpark *park, size_t delta)
283 {
284 
285 	KASSERT(delta < park->park_maxlen); /* "<=" wouldn't make sense */
286 	park->park_copylen = park->park_maxlen - delta;
287 }
288 
289 void
290 puffs_msg_setinfo(struct puffs_msgpark *park, int class, int type,
291 	puffs_cookie_t ck)
292 {
293 
294 	park->park_preq->preq_opclass = PUFFSOP_OPCLASS(class);
295 	park->park_preq->preq_optype = type;
296 	park->park_preq->preq_cookie = ck;
297 }
298 
299 void
300 puffs_msg_setcall(struct puffs_msgpark *park, parkdone_fn donefn, void *donearg)
301 {
302 
303 	KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
304 	park->park_done = donefn;
305 	park->park_donearg = donearg;
306 	park->park_flags |= PARKFLAG_CALL;
307 }
308 
309 /*
310  * kernel-user-kernel waitqueues
311  */
312 
313 static uint64_t
314 puffs_getmsgid(struct puffs_mount *pmp)
315 {
316 	uint64_t rv;
317 
318 	mutex_enter(&pmp->pmp_lock);
319 	rv = pmp->pmp_nextmsgid++;
320 	mutex_exit(&pmp->pmp_lock);
321 
322 	return rv;
323 }
324 
325 /*
326  * A word about reference counting of parks.  A reference must be taken
327  * when accessing a park and additionally when it is on a queue.  So
328  * when taking it off a queue and releasing the access reference, the
329  * reference count is generally decremented by 2.
330  */
331 
332 void
333 puffs_msg_enqueue(struct puffs_mount *pmp, struct puffs_msgpark *park)
334 {
335 	struct lwp *l = curlwp;
336 	struct mount *mp;
337 	struct puffs_req *preq, *creq;
338 	ssize_t delta;
339 
340 	/*
341 	 * Some clients reuse a park, so reset some flags.  We might
342 	 * want to provide a caller-side interface for this and add
343 	 * a few more invariant checks here, but this will do for now.
344 	 */
345 	park->park_flags &= ~(PARKFLAG_DONE | PARKFLAG_HASERROR);
346 	KASSERT((park->park_flags & PARKFLAG_WAITERGONE) == 0);
347 
348 	mp = PMPTOMP(pmp);
349 	preq = park->park_preq;
350 
351 #if 1
352 	/* check if we do compat adjustments */
353 	if (pmp->pmp_docompat && puffs_compat_outgoing(preq, &creq, &delta)) {
354 		park->park_creq = park->park_preq;
355 		park->park_creqlen = park->park_maxlen;
356 
357 		park->park_maxlen += delta;
358 		park->park_copylen += delta;
359 		park->park_preq = preq = creq;
360 	}
361 #endif
362 
363 	preq->preq_buflen = park->park_maxlen;
364 	KASSERT(preq->preq_id == 0
365 	    || (preq->preq_opclass & PUFFSOPFLAG_ISRESPONSE));
366 
367 	if ((park->park_flags & PARKFLAG_WANTREPLY) == 0)
368 		preq->preq_opclass |= PUFFSOPFLAG_FAF;
369 	else
370 		preq->preq_id = puffs_getmsgid(pmp);
371 
372 	/* fill in caller information */
373 	preq->preq_pid = l->l_proc->p_pid;
374 	preq->preq_lid = l->l_lid;
375 
376 	/*
377 	 * To support cv_sig, yet another movie: check if there are signals
378 	 * pending and we are issueing a non-FAF.  If so, return an error
379 	 * directly UNLESS we are issueing INACTIVE/RECLAIM.  In that case,
380 	 * convert it to a FAF, fire off to the file server and return
381 	 * an error.  Yes, this is bordering disgusting.  Barfbags are on me.
382 	 */
383 	if (__predict_false((park->park_flags & PARKFLAG_WANTREPLY)
384 	   && (park->park_flags & PARKFLAG_CALL) == 0
385 	   && (l->l_flag & LW_PENDSIG) != 0 && sigispending(l, 0))) {
386 		sigset_t ss;
387 
388 		/*
389 		 * see the comment about signals in puffs_msg_wait.
390 		 */
391 		sigpending1(l, &ss);
392 		if (sigismember(&ss, SIGINT) ||
393 		    sigismember(&ss, SIGTERM) ||
394 		    sigismember(&ss, SIGKILL) ||
395 		    sigismember(&ss, SIGHUP) ||
396 		    sigismember(&ss, SIGQUIT)) {
397 			park->park_flags |= PARKFLAG_HASERROR;
398 			preq->preq_rv = EINTR;
399 			if (PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN
400 			    && (preq->preq_optype == PUFFS_VN_INACTIVE
401 			     || preq->preq_optype == PUFFS_VN_RECLAIM)) {
402 				park->park_preq->preq_opclass |=
403 				    PUFFSOPFLAG_FAF;
404 				park->park_flags &= ~PARKFLAG_WANTREPLY;
405 				DPRINTF(("puffs_msg_enqueue: "
406 				    "converted to FAF %p\n", park));
407 			} else {
408 				return;
409 			}
410 		}
411 	}
412 
413 	mutex_enter(&pmp->pmp_lock);
414 	if (pmp->pmp_status != PUFFSTAT_RUNNING) {
415 		mutex_exit(&pmp->pmp_lock);
416 		park->park_flags |= PARKFLAG_HASERROR;
417 		preq->preq_rv = ENXIO;
418 		return;
419 	}
420 
421 #ifdef PUFFSDEBUG
422 	parkqdump(&pmp->pmp_msg_touser, puffsdebug > 1);
423 	parkqdump(&pmp->pmp_msg_replywait, puffsdebug > 1);
424 #endif
425 
426 	/*
427 	 * Note: we don't need to lock park since we have the only
428 	 * reference to it at this point.
429 	 */
430 	TAILQ_INSERT_TAIL(&pmp->pmp_msg_touser, park, park_entries);
431 	park->park_flags |= PARKFLAG_ONQUEUE1;
432 	pmp->pmp_msg_touser_count++;
433 	park->park_refcount++;
434 	mutex_exit(&pmp->pmp_lock);
435 
436 	cv_broadcast(&pmp->pmp_msg_waiter_cv);
437 	putter_notify(pmp->pmp_pi);
438 
439 	DPRINTF(("touser: req %" PRIu64 ", preq: %p, park: %p, "
440 	    "c/t: 0x%x/0x%x, f: 0x%x\n", preq->preq_id, preq, park,
441 	    preq->preq_opclass, preq->preq_optype, park->park_flags));
442 }
443 
444 int
445 puffs_msg_wait(struct puffs_mount *pmp, struct puffs_msgpark *park)
446 {
447 	lwp_t *l = curlwp;
448 	proc_t *p = l->l_proc;
449 	struct puffs_req *preq = park->park_preq; /* XXX: hmmm */
450 	sigset_t ss;
451 	sigset_t oss;
452 	int error = 0;
453 	int rv;
454 
455 	/*
456 	 * block unimportant signals.
457 	 *
458 	 * The set of "important" signals here was chosen to be same as
459 	 * nfs interruptible mount.
460 	 */
461 	sigfillset(&ss);
462 	sigdelset(&ss, SIGINT);
463 	sigdelset(&ss, SIGTERM);
464 	sigdelset(&ss, SIGKILL);
465 	sigdelset(&ss, SIGHUP);
466 	sigdelset(&ss, SIGQUIT);
467 	mutex_enter(p->p_lock);
468 	sigprocmask1(l, SIG_BLOCK, &ss, &oss);
469 	mutex_exit(p->p_lock);
470 
471 	mutex_enter(&pmp->pmp_lock);
472 	puffs_mp_reference(pmp);
473 	mutex_exit(&pmp->pmp_lock);
474 
475 	mutex_enter(&park->park_mtx);
476 	/* did the response beat us to the wait? */
477 	if (__predict_false((park->park_flags & PARKFLAG_DONE)
478 	    || (park->park_flags & PARKFLAG_HASERROR))) {
479 		rv = park->park_preq->preq_rv;
480 		mutex_exit(&park->park_mtx);
481 		goto skipwait;
482 	}
483 
484 	if ((park->park_flags & PARKFLAG_WANTREPLY) == 0
485 	    || (park->park_flags & PARKFLAG_CALL)) {
486 		mutex_exit(&park->park_mtx);
487 		rv = 0;
488 		goto skipwait;
489 	}
490 
491 	error = cv_wait_sig(&park->park_cv, &park->park_mtx);
492 	DPRINTF(("puffs_touser: waiter for %p woke up with %d\n",
493 	    park, error));
494 	if (error) {
495 		park->park_flags |= PARKFLAG_WAITERGONE;
496 		if (park->park_flags & PARKFLAG_DONE) {
497 			rv = preq->preq_rv;
498 			mutex_exit(&park->park_mtx);
499 		} else {
500 			/*
501 			 * ok, we marked it as going away, but
502 			 * still need to do queue ops.  take locks
503 			 * in correct order.
504 			 *
505 			 * We don't want to release our reference
506 			 * if it's on replywait queue to avoid error
507 			 * to file server.  putop() code will DTRT.
508 			 */
509 			mutex_exit(&park->park_mtx);
510 			mutex_enter(&pmp->pmp_lock);
511 			mutex_enter(&park->park_mtx);
512 
513 			/*
514 			 * Still on queue1?  We can safely remove it
515 			 * without any consequences since the file
516 			 * server hasn't seen it.  "else" we need to
517 			 * wait for the response and just ignore it
518 			 * to avoid signalling an incorrect error to
519 			 * the file server.
520 			 */
521 			if (park->park_flags & PARKFLAG_ONQUEUE1) {
522 				TAILQ_REMOVE(&pmp->pmp_msg_touser,
523 				    park, park_entries);
524 				puffs_msgpark_release(park);
525 				pmp->pmp_msg_touser_count--;
526 				park->park_flags &= ~PARKFLAG_ONQUEUE1;
527 			} else {
528 				mutex_exit(&park->park_mtx);
529 			}
530 			mutex_exit(&pmp->pmp_lock);
531 
532 			rv = EINTR;
533 		}
534 	} else {
535 		rv = preq->preq_rv;
536 		mutex_exit(&park->park_mtx);
537 	}
538 
539  skipwait:
540 	mutex_enter(&pmp->pmp_lock);
541 	puffs_mp_release(pmp);
542 	mutex_exit(&pmp->pmp_lock);
543 
544 	mutex_enter(p->p_lock);
545 	sigprocmask1(l, SIG_SETMASK, &oss, NULL);
546 	mutex_exit(p->p_lock);
547 
548 	return rv;
549 }
550 
551 /*
552  * XXX: this suuuucks.  Hopefully I'll get rid of this lossage once
553  * the whole setback-nonsense gets fixed.
554  */
555 int
556 puffs_msg_wait2(struct puffs_mount *pmp, struct puffs_msgpark *park,
557 	struct puffs_node *pn1, struct puffs_node *pn2)
558 {
559 	struct puffs_req *preq;
560 	int rv;
561 
562 	rv = puffs_msg_wait(pmp, park);
563 
564 	preq = park->park_preq;
565 	if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N1)
566 		pn1->pn_stat |= PNODE_DOINACT;
567 	if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N2)
568 		pn2->pn_stat |= PNODE_DOINACT;
569 
570 	if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N1)
571 		pn1->pn_stat |= PNODE_NOREFS;
572 	if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N2)
573 		pn2->pn_stat |= PNODE_NOREFS;
574 
575 	return rv;
576 
577 }
578 
579 /*
580  * XXX: lazy bum.  please, for the love of foie gras, fix me.
581  * This should *NOT* depend on setfaf.  Also "memcpy" could
582  * be done more nicely.
583  */
584 void
585 puffs_msg_sendresp(struct puffs_mount *pmp, struct puffs_req *origpreq, int rv)
586 {
587 	struct puffs_msgpark *park;
588 	struct puffs_req *preq;
589 
590 	puffs_msgmem_alloc(sizeof(struct puffs_req), &park, (void *)&preq, 1);
591 	puffs_msg_setfaf(park); /* XXXXXX: avoids reqid override */
592 
593 	memcpy(preq, origpreq, sizeof(struct puffs_req));
594 	preq->preq_rv = rv;
595 	preq->preq_opclass |= PUFFSOPFLAG_ISRESPONSE;
596 
597 	puffs_msg_enqueue(pmp, park);
598 	puffs_msgmem_release(park);
599 }
600 
601 /*
602  * Get next request in the outgoing queue.  "maxsize" controls the
603  * size the caller can accommodate and "nonblock" signals if this
604  * should block while waiting for input.  Handles all locking internally.
605  */
606 int
607 puffs_msgif_getout(void *this, size_t maxsize, int nonblock,
608 	uint8_t **data, size_t *dlen, void **parkptr)
609 {
610 	struct puffs_mount *pmp = this;
611 	struct puffs_msgpark *park = NULL;
612 	struct puffs_req *preq = NULL;
613 	int error;
614 
615 	error = 0;
616 	mutex_enter(&pmp->pmp_lock);
617 	puffs_mp_reference(pmp);
618 	for (;;) {
619 		/* RIP? */
620 		if (pmp->pmp_status != PUFFSTAT_RUNNING) {
621 			error = ENXIO;
622 			break;
623 		}
624 
625 		/* need platinum yendorian express card? */
626 		if (TAILQ_EMPTY(&pmp->pmp_msg_touser)) {
627 			DPRINTF(("puffs_getout: no outgoing op, "));
628 			if (nonblock) {
629 				DPRINTF(("returning EWOULDBLOCK\n"));
630 				error = EWOULDBLOCK;
631 				break;
632 			}
633 			DPRINTF(("waiting ...\n"));
634 
635 			error = cv_wait_sig(&pmp->pmp_msg_waiter_cv,
636 			    &pmp->pmp_lock);
637 			if (error)
638 				break;
639 			else
640 				continue;
641 		}
642 
643 		park = TAILQ_FIRST(&pmp->pmp_msg_touser);
644 		if (park == NULL)
645 			continue;
646 
647 		mutex_enter(&park->park_mtx);
648 		puffs_msgpark_reference(park);
649 
650 		DPRINTF(("puffs_getout: found park at %p, ", park));
651 
652 		/* If it's a goner, don't process any furher */
653 		if (park->park_flags & PARKFLAG_WAITERGONE) {
654 			DPRINTF(("waitergone!\n"));
655 			puffs_msgpark_release(park);
656 			continue;
657 		}
658 		preq = park->park_preq;
659 
660 #if 0
661 		/* check size */
662 		/*
663 		 * XXX: this check is not valid for now, we don't know
664 		 * the size of the caller's input buffer.  i.e. this
665 		 * will most likely go away
666 		 */
667 		if (maxsize < preq->preq_frhdr.pfr_len) {
668 			DPRINTF(("buffer too small\n"));
669 			puffs_msgpark_release(park);
670 			error = E2BIG;
671 			break;
672 		}
673 #endif
674 
675 		DPRINTF(("returning\n"));
676 
677 		/*
678 		 * Ok, we found what we came for.  Release it from the
679 		 * outgoing queue but do not unlock.  We will unlock
680 		 * only after we "releaseout" it to avoid complications:
681 		 * otherwise it is (theoretically) possible for userland
682 		 * to race us into "put" before we have a change to put
683 		 * this baby on the receiving queue.
684 		 */
685 		TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries);
686 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
687 		park->park_flags &= ~PARKFLAG_ONQUEUE1;
688 		mutex_exit(&park->park_mtx);
689 
690 		pmp->pmp_msg_touser_count--;
691 		KASSERT(pmp->pmp_msg_touser_count >= 0);
692 
693 		break;
694 	}
695 	puffs_mp_release(pmp);
696 	mutex_exit(&pmp->pmp_lock);
697 
698 	if (error == 0) {
699 		*data = (uint8_t *)preq;
700 		preq->preq_pth.pth_framelen = park->park_copylen;
701 		*dlen = preq->preq_pth.pth_framelen;
702 		*parkptr = park;
703 	}
704 
705 	return error;
706 }
707 
708 /*
709  * Release outgoing structure.  Now, depending on the success of the
710  * outgoing send, it is either going onto the result waiting queue
711  * or the death chamber.
712  */
713 void
714 puffs_msgif_releaseout(void *this, void *parkptr, int status)
715 {
716 	struct puffs_mount *pmp = this;
717 	struct puffs_msgpark *park = parkptr;
718 
719 	DPRINTF(("puffs_releaseout: returning park %p, errno %d: " ,
720 	    park, status));
721 	mutex_enter(&pmp->pmp_lock);
722 	mutex_enter(&park->park_mtx);
723 	if (park->park_flags & PARKFLAG_WANTREPLY) {
724 		if (status == 0) {
725 			DPRINTF(("enqueue replywait\n"));
726 			TAILQ_INSERT_TAIL(&pmp->pmp_msg_replywait, park,
727 			    park_entries);
728 			park->park_flags |= PARKFLAG_ONQUEUE2;
729 		} else {
730 			DPRINTF(("error path!\n"));
731 			park->park_preq->preq_rv = status;
732 			park->park_flags |= PARKFLAG_DONE;
733 			cv_signal(&park->park_cv);
734 		}
735 		puffs_msgpark_release(park);
736 	} else {
737 		DPRINTF(("release\n"));
738 		puffs_msgpark_release1(park, 2);
739 	}
740 	mutex_exit(&pmp->pmp_lock);
741 }
742 
743 size_t
744 puffs_msgif_waitcount(void *this)
745 {
746 	struct puffs_mount *pmp = this;
747 	size_t rv;
748 
749 	mutex_enter(&pmp->pmp_lock);
750 	rv = pmp->pmp_msg_touser_count;
751 	mutex_exit(&pmp->pmp_lock);
752 
753 	return rv;
754 }
755 
756 /*
757  * XXX: locking with this one?
758  */
759 static void
760 puffsop_msg(void *this, struct puffs_req *preq)
761 {
762 	struct puffs_mount *pmp = this;
763 	struct putter_hdr *pth = &preq->preq_pth;
764 	struct puffs_msgpark *park;
765 	int wgone;
766 
767 	mutex_enter(&pmp->pmp_lock);
768 
769 	/* Locate waiter */
770 	TAILQ_FOREACH(park, &pmp->pmp_msg_replywait, park_entries) {
771 		if (park->park_preq->preq_id == preq->preq_id)
772 			break;
773 	}
774 	if (park == NULL) {
775 		DPRINTF(("puffsop_msg: no request: %" PRIu64 "\n",
776 		    preq->preq_id));
777 		mutex_exit(&pmp->pmp_lock);
778 		return; /* XXX send error */
779 	}
780 
781 	mutex_enter(&park->park_mtx);
782 	puffs_msgpark_reference(park);
783 	if (pth->pth_framelen > park->park_maxlen) {
784 		DPRINTF(("puffsop_msg: invalid buffer length: "
785 		    "%" PRIu64 " (req %" PRIu64 ", \n", pth->pth_framelen,
786 		    preq->preq_id));
787 		park->park_preq->preq_rv = EPROTO;
788 		cv_signal(&park->park_cv);
789 		puffs_msgpark_release1(park, 2);
790 		mutex_exit(&pmp->pmp_lock);
791 		return; /* XXX: error */
792 	}
793 	wgone = park->park_flags & PARKFLAG_WAITERGONE;
794 
795 	KASSERT(park->park_flags & PARKFLAG_ONQUEUE2);
796 	TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries);
797 	park->park_flags &= ~PARKFLAG_ONQUEUE2;
798 	mutex_exit(&pmp->pmp_lock);
799 
800 	if (wgone) {
801 		DPRINTF(("puffsop_msg: bad service - waiter gone for "
802 		    "park %p\n", park));
803 	} else {
804 #if 1
805 		if (park->park_creq) {
806 			struct puffs_req *creq;
807 			size_t csize;
808 
809 			KASSERT(pmp->pmp_docompat);
810 			puffs_compat_incoming(preq, park->park_creq);
811 			creq = park->park_creq;
812 			csize = park->park_creqlen;
813 			park->park_creq = park->park_preq;
814 			park->park_creqlen = park->park_maxlen;
815 
816 			park->park_preq = creq;
817 			park->park_maxlen = csize;
818 
819 			memcpy(park->park_creq, preq, pth->pth_framelen);
820 		} else {
821 #endif
822 			memcpy(park->park_preq, preq, pth->pth_framelen);
823 		}
824 
825 		if (park->park_flags & PARKFLAG_CALL) {
826 			DPRINTF(("puffsop_msg: call for %p, arg %p\n",
827 			    park->park_preq, park->park_donearg));
828 			park->park_done(pmp, preq, park->park_donearg);
829 		}
830 	}
831 
832 	if (!wgone) {
833 		DPRINTF(("puffs_putop: flagging done for "
834 		    "park %p\n", park));
835 		cv_signal(&park->park_cv);
836 	}
837 
838 	park->park_flags |= PARKFLAG_DONE;
839 	puffs_msgpark_release1(park, 2);
840 }
841 
842 static void
843 puffsop_flush(struct puffs_mount *pmp, struct puffs_flush *pf)
844 {
845 	struct vnode *vp;
846 	voff_t offlo, offhi;
847 	int rv, flags = 0;
848 
849 	KASSERT(pf->pf_req.preq_pth.pth_framelen == sizeof(struct puffs_flush));
850 
851 	/* XXX: slurry */
852 	if (pf->pf_op == PUFFS_INVAL_NAMECACHE_ALL) {
853 		cache_purgevfs(PMPTOMP(pmp));
854 		rv = 0;
855 		goto out;
856 	}
857 
858 	/*
859 	 * Get vnode, don't lock it.  Namecache is protected by its own lock
860 	 * and we have a reference to protect against premature harvesting.
861 	 *
862 	 * The node we want here might be locked and the op is in
863 	 * userspace waiting for us to complete ==> deadlock.  Another
864 	 * reason we need to eventually bump locking to userspace, as we
865 	 * will need to lock the node if we wish to do flushes.
866 	 */
867 	rv = puffs_cookie2vnode(pmp, pf->pf_cookie, 0, 0, &vp);
868 	if (rv) {
869 		if (rv == PUFFS_NOSUCHCOOKIE)
870 			rv = ENOENT;
871 		goto out;
872 	}
873 
874 	switch (pf->pf_op) {
875 #if 0
876 	/* not quite ready, yet */
877 	case PUFFS_INVAL_NAMECACHE_NODE:
878 	struct componentname *pf_cn;
879 	char *name;
880 		/* get comfortab^Wcomponentname */
881 		pf_cn = kmem_alloc(componentname);
882 		memset(pf_cn, 0, sizeof(struct componentname));
883 		break;
884 
885 #endif
886 	case PUFFS_INVAL_NAMECACHE_DIR:
887 		if (vp->v_type != VDIR) {
888 			rv = EINVAL;
889 			break;
890 		}
891 		cache_purge1(vp, NULL, PURGE_CHILDREN);
892 		break;
893 
894 	case PUFFS_INVAL_PAGECACHE_NODE_RANGE:
895 		flags = PGO_FREE;
896 		/*FALLTHROUGH*/
897 	case PUFFS_FLUSH_PAGECACHE_NODE_RANGE:
898 		if (flags == 0)
899 			flags = PGO_CLEANIT;
900 
901 		if (pf->pf_end > vp->v_size || vp->v_type != VREG) {
902 			rv = EINVAL;
903 			break;
904 		}
905 
906 		offlo = trunc_page(pf->pf_start);
907 		offhi = round_page(pf->pf_end);
908 		if (offhi != 0 && offlo >= offhi) {
909 			rv = EINVAL;
910 			break;
911 		}
912 
913 		mutex_enter(vp->v_uobj.vmobjlock);
914 		rv = VOP_PUTPAGES(vp, offlo, offhi, flags);
915 		break;
916 
917 	default:
918 		rv = EINVAL;
919 	}
920 
921 	vrele(vp);
922 
923  out:
924 	puffs_msg_sendresp(pmp, &pf->pf_req, rv);
925 }
926 
927 int
928 puffs_msgif_dispatch(void *this, struct putter_hdr *pth)
929 {
930 	struct puffs_mount *pmp = this;
931 	struct puffs_req *preq = (struct puffs_req *)pth;
932 	struct puffs_sopreq *psopr;
933 
934 	if (pth->pth_framelen < sizeof(struct puffs_req)) {
935 		puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */
936 		return 0;
937 	}
938 
939 	switch (PUFFSOP_OPCLASS(preq->preq_opclass)) {
940 	case PUFFSOP_VN:
941 	case PUFFSOP_VFS:
942 		DPRINTF(("dispatch: vn/vfs message 0x%x\n", preq->preq_optype));
943 		puffsop_msg(pmp, preq);
944 		break;
945 
946 	case PUFFSOP_FLUSH: /* process in sop thread */
947 	{
948 		struct puffs_flush *pf;
949 
950 		DPRINTF(("dispatch: flush 0x%x\n", preq->preq_optype));
951 
952 		if (preq->preq_pth.pth_framelen != sizeof(struct puffs_flush)) {
953 			puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */
954 			break;
955 		}
956 		pf = (struct puffs_flush *)preq;
957 
958 		KASSERT(curlwp != uvm.pagedaemon_lwp);
959 		psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP);
960 		memcpy(&psopr->psopr_pf, pf, sizeof(*pf));
961 		psopr->psopr_sopreq = PUFFS_SOPREQ_FLUSH;
962 
963 		mutex_enter(&pmp->pmp_sopmtx);
964 		if (pmp->pmp_sopthrcount == 0) {
965 			mutex_exit(&pmp->pmp_sopmtx);
966 			kmem_free(psopr, sizeof(*psopr));
967 			puffs_msg_sendresp(pmp, preq, ENXIO);
968 		} else {
969 			TAILQ_INSERT_TAIL(&pmp->pmp_sopreqs,
970 			    psopr, psopr_entries);
971 			cv_signal(&pmp->pmp_sopcv);
972 			mutex_exit(&pmp->pmp_sopmtx);
973 		}
974 		break;
975 	}
976 
977 	case PUFFSOP_UNMOUNT: /* process in sop thread */
978 	{
979 
980 		DPRINTF(("dispatch: unmount 0x%x\n", preq->preq_optype));
981 
982 		KASSERT(curlwp != uvm.pagedaemon_lwp);
983 		psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP);
984 		psopr->psopr_preq = *preq;
985 		psopr->psopr_sopreq = PUFFS_SOPREQ_UNMOUNT;
986 
987 		mutex_enter(&pmp->pmp_sopmtx);
988 		if (pmp->pmp_sopthrcount == 0) {
989 			mutex_exit(&pmp->pmp_sopmtx);
990 			kmem_free(psopr, sizeof(*psopr));
991 			puffs_msg_sendresp(pmp, preq, ENXIO);
992 		} else {
993 			TAILQ_INSERT_TAIL(&pmp->pmp_sopreqs,
994 			    psopr, psopr_entries);
995 			cv_signal(&pmp->pmp_sopcv);
996 			mutex_exit(&pmp->pmp_sopmtx);
997 		}
998 		break;
999 	}
1000 
1001 	default:
1002 		DPRINTF(("dispatch: invalid class 0x%x\n", preq->preq_opclass));
1003 		puffs_msg_sendresp(pmp, preq, EOPNOTSUPP);
1004 		break;
1005 	}
1006 
1007 	return 0;
1008 }
1009 
1010 /*
1011  * Work loop for thread processing all ops from server which
1012  * cannot safely be handled in caller context.  This includes
1013  * everything which might need a lock currently "held" by the file
1014  * server, i.e. a long-term kernel lock which will be released only
1015  * once the file server acknowledges a request
1016  */
1017 void
1018 puffs_sop_thread(void *arg)
1019 {
1020 	struct puffs_mount *pmp = arg;
1021 	struct mount *mp = PMPTOMP(pmp);
1022 	struct puffs_sopreq *psopr;
1023 	bool keeprunning;
1024 	bool unmountme = false;
1025 
1026 	mutex_enter(&pmp->pmp_sopmtx);
1027 	for (keeprunning = true; keeprunning; ) {
1028 		while ((psopr = TAILQ_FIRST(&pmp->pmp_sopreqs)) == NULL)
1029 			cv_wait(&pmp->pmp_sopcv, &pmp->pmp_sopmtx);
1030 		TAILQ_REMOVE(&pmp->pmp_sopreqs, psopr, psopr_entries);
1031 		mutex_exit(&pmp->pmp_sopmtx);
1032 
1033 		switch (psopr->psopr_sopreq) {
1034 		case PUFFS_SOPREQSYS_EXIT:
1035 			keeprunning = false;
1036 			break;
1037 		case PUFFS_SOPREQ_FLUSH:
1038 			puffsop_flush(pmp, &psopr->psopr_pf);
1039 			break;
1040 		case PUFFS_SOPREQ_UNMOUNT:
1041 			puffs_msg_sendresp(pmp, &psopr->psopr_preq, 0);
1042 
1043 			unmountme = true;
1044 			keeprunning = false;
1045 
1046 			/*
1047 			 * We know the mountpoint is still alive because
1048 			 * the thread that is us (poetic?) is still alive.
1049 			 */
1050 			atomic_inc_uint((unsigned int*)&mp->mnt_refcnt);
1051 			break;
1052 		}
1053 
1054 		kmem_free(psopr, sizeof(*psopr));
1055 		mutex_enter(&pmp->pmp_sopmtx);
1056 	}
1057 
1058 	/*
1059 	 * Purge remaining ops.
1060 	 */
1061 	while ((psopr = TAILQ_FIRST(&pmp->pmp_sopreqs)) != NULL) {
1062 		TAILQ_REMOVE(&pmp->pmp_sopreqs, psopr, psopr_entries);
1063 		mutex_exit(&pmp->pmp_sopmtx);
1064 		puffs_msg_sendresp(pmp, &psopr->psopr_preq, ENXIO);
1065 		kmem_free(psopr, sizeof(*psopr));
1066 		mutex_enter(&pmp->pmp_sopmtx);
1067 	}
1068 
1069 	pmp->pmp_sopthrcount--;
1070 	cv_broadcast(&pmp->pmp_sopcv);
1071 	mutex_exit(&pmp->pmp_sopmtx); /* not allowed to access fs after this */
1072 
1073 	/*
1074 	 * If unmount was requested, we can now safely do it here, since
1075 	 * our context is dead from the point-of-view of puffs_unmount()
1076 	 * and we are just another thread.  dounmount() makes internally
1077 	 * sure that VFS_UNMOUNT() isn't called reentrantly and that it
1078 	 * is eventually completed.
1079 	 */
1080 	if (unmountme) {
1081 		(void)dounmount(mp, MNT_FORCE, curlwp);
1082 		vfs_destroy(mp);
1083 	}
1084 
1085 	kthread_exit(0);
1086 }
1087 
1088 int
1089 puffs_msgif_close(void *this)
1090 {
1091 	struct puffs_mount *pmp = this;
1092 	struct mount *mp = PMPTOMP(pmp);
1093 
1094 	mutex_enter(&pmp->pmp_lock);
1095 	puffs_mp_reference(pmp);
1096 
1097 	/*
1098 	 * Free the waiting callers before proceeding any further.
1099 	 * The syncer might be jogging around in this file system
1100 	 * currently.  If we allow it to go to the userspace of no
1101 	 * return while trying to get the syncer lock, well ...
1102 	 */
1103 	puffs_userdead(pmp);
1104 
1105 	/*
1106 	 * Make sure someone from puffs_unmount() isn't currently in
1107 	 * userspace.  If we don't take this precautionary step,
1108 	 * they might notice that the mountpoint has disappeared
1109 	 * from under them once they return.  Especially note that we
1110 	 * cannot simply test for an unmounter before calling
1111 	 * dounmount(), since it might be possible that that particular
1112 	 * invocation of unmount was called without MNT_FORCE.  Here we
1113 	 * *must* make sure unmount succeeds.  Also, restart is necessary
1114 	 * since pmp isn't locked.  We might end up with PUTTER_DEAD after
1115 	 * restart and exit from there.
1116 	 */
1117 	if (pmp->pmp_unmounting) {
1118 		cv_wait(&pmp->pmp_unmounting_cv, &pmp->pmp_lock);
1119 		puffs_mp_release(pmp);
1120 		mutex_exit(&pmp->pmp_lock);
1121 		DPRINTF(("puffs_fop_close: unmount was in progress for pmp %p, "
1122 		    "restart\n", pmp));
1123 		return ERESTART;
1124 	}
1125 
1126 	/* Won't access pmp from here anymore */
1127 	atomic_inc_uint((unsigned int*)&mp->mnt_refcnt);
1128 	puffs_mp_release(pmp);
1129 	mutex_exit(&pmp->pmp_lock);
1130 
1131 	/* Detach from VFS. */
1132 	(void)dounmount(mp, MNT_FORCE, curlwp);
1133 	vfs_destroy(mp);
1134 
1135 	return 0;
1136 }
1137 
1138 /*
1139  * We're dead, kaput, RIP, slightly more than merely pining for the
1140  * fjords, belly-up, fallen, lifeless, finished, expired, gone to meet
1141  * our maker, ceased to be, etcetc.  YASD.  It's a dead FS!
1142  *
1143  * Caller must hold puffs mutex.
1144  */
1145 void
1146 puffs_userdead(struct puffs_mount *pmp)
1147 {
1148 	struct puffs_msgpark *park, *park_next;
1149 
1150 	/*
1151 	 * Mark filesystem status as dying so that operations don't
1152 	 * attempt to march to userspace any longer.
1153 	 */
1154 	pmp->pmp_status = PUFFSTAT_DYING;
1155 
1156 	/* signal waiters on REQUEST TO file server queue */
1157 	for (park = TAILQ_FIRST(&pmp->pmp_msg_touser); park; park = park_next) {
1158 		uint8_t opclass;
1159 
1160 		mutex_enter(&park->park_mtx);
1161 		puffs_msgpark_reference(park);
1162 		park_next = TAILQ_NEXT(park, park_entries);
1163 
1164 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
1165 		TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries);
1166 		park->park_flags &= ~PARKFLAG_ONQUEUE1;
1167 		pmp->pmp_msg_touser_count--;
1168 
1169 		/*
1170 		 * Even though waiters on QUEUE1 are removed in touser()
1171 		 * in case of WAITERGONE, it is still possible for us to
1172 		 * get raced here due to having to retake locks in said
1173 		 * touser().  In the race case simply "ignore" the item
1174 		 * on the queue and move on to the next one.
1175 		 */
1176 		if (park->park_flags & PARKFLAG_WAITERGONE) {
1177 			KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
1178 			KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
1179 			puffs_msgpark_release(park);
1180 
1181 		} else {
1182 			opclass = park->park_preq->preq_opclass;
1183 			park->park_preq->preq_rv = ENXIO;
1184 
1185 			if (park->park_flags & PARKFLAG_CALL) {
1186 				park->park_done(pmp, park->park_preq,
1187 				    park->park_donearg);
1188 				puffs_msgpark_release1(park, 2);
1189 			} else if ((park->park_flags & PARKFLAG_WANTREPLY)==0) {
1190 				puffs_msgpark_release1(park, 2);
1191 			} else {
1192 				park->park_preq->preq_rv = ENXIO;
1193 				cv_signal(&park->park_cv);
1194 				puffs_msgpark_release(park);
1195 			}
1196 		}
1197 	}
1198 
1199 	/* signal waiters on RESPONSE FROM file server queue */
1200 	for (park=TAILQ_FIRST(&pmp->pmp_msg_replywait); park; park=park_next) {
1201 		mutex_enter(&park->park_mtx);
1202 		puffs_msgpark_reference(park);
1203 		park_next = TAILQ_NEXT(park, park_entries);
1204 
1205 		KASSERT(park->park_flags & PARKFLAG_ONQUEUE2);
1206 		KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
1207 
1208 		TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries);
1209 		park->park_flags &= ~PARKFLAG_ONQUEUE2;
1210 
1211 		if (park->park_flags & PARKFLAG_WAITERGONE) {
1212 			KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
1213 			puffs_msgpark_release(park);
1214 		} else {
1215 			park->park_preq->preq_rv = ENXIO;
1216 			if (park->park_flags & PARKFLAG_CALL) {
1217 				park->park_done(pmp, park->park_preq,
1218 				    park->park_donearg);
1219 				puffs_msgpark_release1(park, 2);
1220 			} else {
1221 				cv_signal(&park->park_cv);
1222 				puffs_msgpark_release(park);
1223 			}
1224 		}
1225 	}
1226 
1227 	cv_broadcast(&pmp->pmp_msg_waiter_cv);
1228 }
1229