puffs_msgif.c revision 1.101.10.4 1 /* $NetBSD: puffs_msgif.c,v 1.101.10.4 2018/09/18 01:15:58 pgoyette 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 1.101.10.4 2018/09/18 01:15:58 pgoyette Exp $");
34
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/atomic.h>
38 #include <sys/kmem.h>
39 #include <sys/kthread.h>
40 #include <sys/lock.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 #include <sys/compat_stub.h>
47
48 #include <uvm/uvm.h>
49
50 #include <dev/putter/putter_sys.h>
51
52 #include <fs/puffs/puffs_msgif.h>
53 #include <fs/puffs/puffs_sys.h>
54
55 /*
56 * waitq data structures
57 */
58
59 /*
60 * While a request is going to userspace, park the caller within the
61 * kernel. This is the kernel counterpart of "struct puffs_req".
62 */
63 struct puffs_msgpark {
64 struct puffs_req *park_preq; /* req followed by buf */
65
66 size_t park_copylen; /* userspace copylength */
67 size_t park_maxlen; /* max size in comeback */
68
69 struct puffs_req *park_creq; /* non-compat preq */
70 size_t park_creqlen; /* non-compat preq len */
71
72 parkdone_fn park_done; /* "biodone" a'la puffs */
73 void *park_donearg;
74
75 int park_flags;
76 int park_refcount;
77
78 kcondvar_t park_cv;
79 kmutex_t park_mtx;
80
81 TAILQ_ENTRY(puffs_msgpark) park_entries;
82 };
83 #define PARKFLAG_WAITERGONE 0x01
84 #define PARKFLAG_DONE 0x02
85 #define PARKFLAG_ONQUEUE1 0x04
86 #define PARKFLAG_ONQUEUE2 0x08
87 #define PARKFLAG_CALL 0x10
88 #define PARKFLAG_WANTREPLY 0x20
89 #define PARKFLAG_HASERROR 0x40
90
91 static pool_cache_t parkpc;
92 #ifdef PUFFSDEBUG
93 static int totalpark;
94 #endif
95
96 int puffs_sopreq_expire_timeout = PUFFS_SOPREQ_EXPIRE_TIMEOUT;
97
98 static int
99 makepark(void *arg, void *obj, int flags)
100 {
101 struct puffs_msgpark *park = obj;
102
103 mutex_init(&park->park_mtx, MUTEX_DEFAULT, IPL_NONE);
104 cv_init(&park->park_cv, "puffsrpl");
105
106 return 0;
107 }
108
109 static void
110 nukepark(void *arg, void *obj)
111 {
112 struct puffs_msgpark *park = obj;
113
114 cv_destroy(&park->park_cv);
115 mutex_destroy(&park->park_mtx);
116 }
117
118 void
119 puffs_msgif_init(void)
120 {
121
122 parkpc = pool_cache_init(sizeof(struct puffs_msgpark), 0, 0, 0,
123 "puffprkl", NULL, IPL_NONE, makepark, nukepark, NULL);
124 }
125
126 void
127 puffs_msgif_destroy(void)
128 {
129
130 pool_cache_destroy(parkpc);
131 }
132
133 static struct puffs_msgpark *
134 puffs_msgpark_alloc(int waitok)
135 {
136 struct puffs_msgpark *park;
137
138 KASSERT(curlwp != uvm.pagedaemon_lwp || !waitok);
139
140 park = pool_cache_get(parkpc, waitok ? PR_WAITOK : PR_NOWAIT);
141 if (park == NULL)
142 return park;
143
144 park->park_refcount = 1;
145 park->park_preq = park->park_creq = NULL;
146 park->park_flags = PARKFLAG_WANTREPLY;
147
148 #ifdef PUFFSDEBUG
149 totalpark++;
150 #endif
151
152 return park;
153 }
154
155 static void
156 puffs_msgpark_reference(struct puffs_msgpark *park)
157 {
158
159 KASSERT(mutex_owned(&park->park_mtx));
160 park->park_refcount++;
161 }
162
163 /*
164 * Release reference to park structure.
165 */
166 static void
167 puffs_msgpark_release1(struct puffs_msgpark *park, int howmany)
168 {
169 struct puffs_req *preq = park->park_preq;
170 struct puffs_req *creq = park->park_creq;
171 int refcnt;
172
173 KASSERT(mutex_owned(&park->park_mtx));
174 refcnt = park->park_refcount -= howmany;
175 mutex_exit(&park->park_mtx);
176
177 KASSERT(refcnt >= 0);
178
179 if (refcnt == 0) {
180 if (preq)
181 kmem_free(preq, park->park_maxlen);
182 #if 1
183 if (creq)
184 kmem_free(creq, park->park_creqlen);
185 #endif
186 pool_cache_put(parkpc, park);
187
188 #ifdef PUFFSDEBUG
189 totalpark--;
190 #endif
191 }
192 }
193 #define puffs_msgpark_release(a) puffs_msgpark_release1(a, 1)
194
195 #ifdef PUFFSDEBUG
196 static void
197 parkdump(struct puffs_msgpark *park)
198 {
199
200 DPRINTF(("park %p, preq %p, id %" PRIu64 "\n"
201 "\tcopy %zu, max %zu - done: %p/%p\n"
202 "\tflags 0x%08x, refcount %d, cv/mtx: %p/%p\n",
203 park, park->park_preq, park->park_preq->preq_id,
204 park->park_copylen, park->park_maxlen,
205 park->park_done, park->park_donearg,
206 park->park_flags, park->park_refcount,
207 &park->park_cv, &park->park_mtx));
208 }
209
210 static void
211 parkqdump(struct puffs_wq *q, int dumpall)
212 {
213 struct puffs_msgpark *park;
214 int total = 0;
215
216 TAILQ_FOREACH(park, q, park_entries) {
217 if (dumpall)
218 parkdump(park);
219 total++;
220 }
221 DPRINTF(("puffs waitqueue at %p dumped, %d total\n", q, total));
222
223 }
224 #endif /* PUFFSDEBUG */
225
226 /*
227 * A word about locking in the park structures: the lock protects the
228 * fields of the *park* structure (not preq) and acts as an interlock
229 * in cv operations. The lock is always internal to this module and
230 * callers do not need to worry about it.
231 */
232
233 int
234 puffs_msgmem_alloc(size_t len, struct puffs_msgpark **ppark, void **mem,
235 int cansleep)
236 {
237 struct puffs_msgpark *park;
238 void *m;
239
240 KASSERT(curlwp != uvm.pagedaemon_lwp || !cansleep);
241 m = kmem_zalloc(len, cansleep ? KM_SLEEP : KM_NOSLEEP);
242 if (m == NULL) {
243 KASSERT(cansleep == 0);
244 return ENOMEM;
245 }
246
247 park = puffs_msgpark_alloc(cansleep);
248 if (park == NULL) {
249 KASSERT(cansleep == 0);
250 kmem_free(m, len);
251 return ENOMEM;
252 }
253
254 park->park_preq = m;
255 park->park_maxlen = park->park_copylen = len;
256
257 *ppark = park;
258 *mem = m;
259
260 return 0;
261 }
262
263 void
264 puffs_msgmem_release(struct puffs_msgpark *park)
265 {
266
267 if (park == NULL)
268 return;
269
270 mutex_enter(&park->park_mtx);
271 puffs_msgpark_release(park);
272 }
273
274 void
275 puffs_msg_setfaf(struct puffs_msgpark *park)
276 {
277
278 KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
279 park->park_flags &= ~PARKFLAG_WANTREPLY;
280 }
281
282 void
283 puffs_msg_setdelta(struct puffs_msgpark *park, size_t delta)
284 {
285
286 KASSERT(delta < park->park_maxlen); /* "<=" wouldn't make sense */
287 park->park_copylen = park->park_maxlen - delta;
288 }
289
290 void
291 puffs_msg_setinfo(struct puffs_msgpark *park, int opclass, int type,
292 puffs_cookie_t ck)
293 {
294
295 park->park_preq->preq_opclass = PUFFSOP_OPCLASS(opclass);
296 park->park_preq->preq_optype = type;
297 park->park_preq->preq_cookie = ck;
298 }
299
300 void
301 puffs_msg_setcall(struct puffs_msgpark *park, parkdone_fn donefn, void *donearg)
302 {
303
304 KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
305 park->park_done = donefn;
306 park->park_donearg = donearg;
307 park->park_flags |= PARKFLAG_CALL;
308 }
309
310 /*
311 * kernel-user-kernel waitqueues
312 */
313
314 static uint64_t
315 puffs_getmsgid(struct puffs_mount *pmp)
316 {
317 uint64_t rv;
318
319 mutex_enter(&pmp->pmp_lock);
320 rv = pmp->pmp_nextmsgid++;
321 mutex_exit(&pmp->pmp_lock);
322
323 return rv;
324 }
325
326 /* Routines to call the compat hooks */
327 /* Out-going */
328 COMPAT_CALL_HOOK_DECL(puffs50_compat_hook, f1,
329 (struct puffs_req *oreq, struct puffs_req **creqp, ssize_t *deltap),
330 (oreq, creqp, deltap), enosys());
331 COMPAT_CALL_HOOK(puffs50_compat_hook, f1,
332 (struct puffs_req *oreq, struct puffs_req **creqp, ssize_t *deltap),
333 (oreq, creqp, deltap), enosys());
334
335 /* Incoming */
336 COMPAT_CALL_HOOK_DECL(puffs50_compat_hook, f2,
337 (struct puffs_req *oreq, struct puffs_req *creqp),
338 (oreq, creqp), enosys());
339 COMPAT_CALL_HOOK(puffs50_compat_hook, f2,
340 (struct puffs_req *oreq, struct puffs_req *creqp),
341 (oreq, creqp), enosys());
342
343 /*
344 * A word about reference counting of parks. A reference must be taken
345 * when accessing a park and additionally when it is on a queue. So
346 * when taking it off a queue and releasing the access reference, the
347 * reference count is generally decremented by 2.
348 */
349
350 void
351 puffs_msg_enqueue(struct puffs_mount *pmp, struct puffs_msgpark *park)
352 {
353 struct lwp *l = curlwp;
354 struct puffs_req *preq, *creq;
355 ssize_t delta;
356
357 /*
358 * Some clients reuse a park, so reset some flags. We might
359 * want to provide a caller-side interface for this and add
360 * a few more invariant checks here, but this will do for now.
361 */
362 park->park_flags &= ~(PARKFLAG_DONE | PARKFLAG_HASERROR);
363 KASSERT((park->park_flags & PARKFLAG_WAITERGONE) == 0);
364
365 preq = park->park_preq;
366
367 #if 1
368 /* check if we do compat adjustments */
369 if (pmp->pmp_docompat &&
370 puffs50_compat_hook_f1_call(preq, &creq, &delta) == 0) {
371 park->park_creq = park->park_preq;
372 park->park_creqlen = park->park_maxlen;
373
374 park->park_maxlen += delta;
375 park->park_copylen += delta;
376 park->park_preq = preq = creq;
377 }
378 #endif
379
380 preq->preq_buflen = park->park_maxlen;
381 KASSERT(preq->preq_id == 0
382 || (preq->preq_opclass & PUFFSOPFLAG_ISRESPONSE));
383
384 if ((park->park_flags & PARKFLAG_WANTREPLY) == 0)
385 preq->preq_opclass |= PUFFSOPFLAG_FAF;
386 else
387 preq->preq_id = puffs_getmsgid(pmp);
388
389 /* fill in caller information */
390 preq->preq_pid = l->l_proc->p_pid;
391 preq->preq_lid = l->l_lid;
392
393 /*
394 * To support cv_sig, yet another movie: check if there are signals
395 * pending and we are issueing a non-FAF. If so, return an error
396 * directly UNLESS we are issueing INACTIVE/RECLAIM. In that case,
397 * convert it to a FAF, fire off to the file server and return
398 * an error. Yes, this is bordering disgusting. Barfbags are on me.
399 */
400 if (__predict_false((park->park_flags & PARKFLAG_WANTREPLY)
401 && (park->park_flags & PARKFLAG_CALL) == 0
402 && (l->l_flag & LW_PENDSIG) != 0 && sigispending(l, 0))) {
403 sigset_t ss;
404
405 /*
406 * see the comment about signals in puffs_msg_wait.
407 */
408 sigpending1(l, &ss);
409 if (sigismember(&ss, SIGINT) ||
410 sigismember(&ss, SIGTERM) ||
411 sigismember(&ss, SIGKILL) ||
412 sigismember(&ss, SIGHUP) ||
413 sigismember(&ss, SIGQUIT)) {
414 park->park_flags |= PARKFLAG_HASERROR;
415 preq->preq_rv = EINTR;
416 if (PUFFSOP_OPCLASS(preq->preq_opclass) == PUFFSOP_VN
417 && (preq->preq_optype == PUFFS_VN_INACTIVE
418 || preq->preq_optype == PUFFS_VN_RECLAIM)) {
419 park->park_preq->preq_opclass |=
420 PUFFSOPFLAG_FAF;
421 park->park_flags &= ~PARKFLAG_WANTREPLY;
422 DPRINTF(("puffs_msg_enqueue: "
423 "converted to FAF %p\n", park));
424 } else {
425 return;
426 }
427 }
428 }
429
430 mutex_enter(&pmp->pmp_lock);
431 if (pmp->pmp_status != PUFFSTAT_RUNNING) {
432 mutex_exit(&pmp->pmp_lock);
433 park->park_flags |= PARKFLAG_HASERROR;
434 preq->preq_rv = ENXIO;
435 return;
436 }
437
438 #ifdef PUFFSDEBUG
439 parkqdump(&pmp->pmp_msg_touser, puffsdebug > 1);
440 parkqdump(&pmp->pmp_msg_replywait, puffsdebug > 1);
441 #endif
442
443 /*
444 * Note: we don't need to lock park since we have the only
445 * reference to it at this point.
446 */
447 TAILQ_INSERT_TAIL(&pmp->pmp_msg_touser, park, park_entries);
448 park->park_flags |= PARKFLAG_ONQUEUE1;
449 pmp->pmp_msg_touser_count++;
450 park->park_refcount++;
451
452 cv_broadcast(&pmp->pmp_msg_waiter_cv);
453 mutex_exit(&pmp->pmp_lock);
454 putter_notify(pmp->pmp_pi);
455
456 DPRINTF(("touser: req %" PRIu64 ", preq: %p, park: %p, "
457 "c/t: 0x%x/0x%x, f: 0x%x\n", preq->preq_id, preq, park,
458 preq->preq_opclass, preq->preq_optype, park->park_flags));
459 }
460
461 int
462 puffs_msg_wait(struct puffs_mount *pmp, struct puffs_msgpark *park)
463 {
464 lwp_t *l = curlwp;
465 proc_t *p = l->l_proc;
466 struct puffs_req *preq = park->park_preq; /* XXX: hmmm */
467 sigset_t ss;
468 sigset_t oss;
469 int error = 0;
470 int rv;
471
472 /*
473 * block unimportant signals.
474 *
475 * The set of "important" signals here was chosen to be same as
476 * nfs interruptible mount.
477 */
478 sigfillset(&ss);
479 sigdelset(&ss, SIGINT);
480 sigdelset(&ss, SIGTERM);
481 sigdelset(&ss, SIGKILL);
482 sigdelset(&ss, SIGHUP);
483 sigdelset(&ss, SIGQUIT);
484 mutex_enter(p->p_lock);
485 sigprocmask1(l, SIG_BLOCK, &ss, &oss);
486 mutex_exit(p->p_lock);
487
488 mutex_enter(&pmp->pmp_lock);
489 puffs_mp_reference(pmp);
490 mutex_exit(&pmp->pmp_lock);
491
492 mutex_enter(&park->park_mtx);
493 /* did the response beat us to the wait? */
494 if (__predict_false((park->park_flags & PARKFLAG_DONE)
495 || (park->park_flags & PARKFLAG_HASERROR))) {
496 rv = park->park_preq->preq_rv;
497 mutex_exit(&park->park_mtx);
498 goto skipwait;
499 }
500
501 if ((park->park_flags & PARKFLAG_WANTREPLY) == 0
502 || (park->park_flags & PARKFLAG_CALL)) {
503 mutex_exit(&park->park_mtx);
504 rv = 0;
505 goto skipwait;
506 }
507
508 error = cv_wait_sig(&park->park_cv, &park->park_mtx);
509 DPRINTF(("puffs_touser: waiter for %p woke up with %d\n",
510 park, error));
511 if (error) {
512 park->park_flags |= PARKFLAG_WAITERGONE;
513 if (park->park_flags & PARKFLAG_DONE) {
514 rv = preq->preq_rv;
515 mutex_exit(&park->park_mtx);
516 } else {
517 /*
518 * ok, we marked it as going away, but
519 * still need to do queue ops. take locks
520 * in correct order.
521 *
522 * We don't want to release our reference
523 * if it's on replywait queue to avoid error
524 * to file server. putop() code will DTRT.
525 */
526 mutex_exit(&park->park_mtx);
527 mutex_enter(&pmp->pmp_lock);
528 mutex_enter(&park->park_mtx);
529
530 /*
531 * Still on queue1? We can safely remove it
532 * without any consequences since the file
533 * server hasn't seen it. "else" we need to
534 * wait for the response and just ignore it
535 * to avoid signalling an incorrect error to
536 * the file server.
537 */
538 if (park->park_flags & PARKFLAG_ONQUEUE1) {
539 TAILQ_REMOVE(&pmp->pmp_msg_touser,
540 park, park_entries);
541 puffs_msgpark_release(park);
542 pmp->pmp_msg_touser_count--;
543 park->park_flags &= ~PARKFLAG_ONQUEUE1;
544 } else {
545 mutex_exit(&park->park_mtx);
546 }
547 mutex_exit(&pmp->pmp_lock);
548
549 rv = EINTR;
550 }
551 } else {
552 rv = preq->preq_rv;
553 mutex_exit(&park->park_mtx);
554 }
555
556 skipwait:
557 mutex_enter(&pmp->pmp_lock);
558 puffs_mp_release(pmp);
559 mutex_exit(&pmp->pmp_lock);
560
561 mutex_enter(p->p_lock);
562 sigprocmask1(l, SIG_SETMASK, &oss, NULL);
563 mutex_exit(p->p_lock);
564
565 return rv;
566 }
567
568 /*
569 * XXX: this suuuucks. Hopefully I'll get rid of this lossage once
570 * the whole setback-nonsense gets fixed.
571 */
572 int
573 puffs_msg_wait2(struct puffs_mount *pmp, struct puffs_msgpark *park,
574 struct puffs_node *pn1, struct puffs_node *pn2)
575 {
576 struct puffs_req *preq;
577 int rv;
578
579 rv = puffs_msg_wait(pmp, park);
580
581 preq = park->park_preq;
582 if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N1)
583 pn1->pn_stat |= PNODE_DOINACT;
584 if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_INACT_N2)
585 pn2->pn_stat |= PNODE_DOINACT;
586
587 if (pn1 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N1)
588 pn1->pn_stat |= PNODE_NOREFS;
589 if (pn2 && preq->preq_setbacks & PUFFS_SETBACK_NOREF_N2)
590 pn2->pn_stat |= PNODE_NOREFS;
591
592 return rv;
593
594 }
595
596 /*
597 * XXX: lazy bum. please, for the love of foie gras, fix me.
598 * This should *NOT* depend on setfaf. Also "memcpy" could
599 * be done more nicely.
600 */
601 void
602 puffs_msg_sendresp(struct puffs_mount *pmp, struct puffs_req *origpreq, int rv)
603 {
604 struct puffs_msgpark *park;
605 struct puffs_req *preq;
606
607 puffs_msgmem_alloc(sizeof(struct puffs_req), &park, (void *)&preq, 1);
608 puffs_msg_setfaf(park); /* XXXXXX: avoids reqid override */
609
610 memcpy(preq, origpreq, sizeof(struct puffs_req));
611 preq->preq_rv = rv;
612 preq->preq_opclass |= PUFFSOPFLAG_ISRESPONSE;
613
614 puffs_msg_enqueue(pmp, park);
615 puffs_msgmem_release(park);
616 }
617
618 /*
619 * Get next request in the outgoing queue. "maxsize" controls the
620 * size the caller can accommodate and "nonblock" signals if this
621 * should block while waiting for input. Handles all locking internally.
622 */
623 int
624 puffs_msgif_getout(void *ctx, size_t maxsize, int nonblock,
625 uint8_t **data, size_t *dlen, void **parkptr)
626 {
627 struct puffs_mount *pmp = ctx;
628 struct puffs_msgpark *park = NULL;
629 struct puffs_req *preq = NULL;
630 int error;
631
632 error = 0;
633 mutex_enter(&pmp->pmp_lock);
634 puffs_mp_reference(pmp);
635 for (;;) {
636 /* RIP? */
637 if (pmp->pmp_status != PUFFSTAT_RUNNING) {
638 error = ENXIO;
639 break;
640 }
641
642 /* need platinum yendorian express card? */
643 if (TAILQ_EMPTY(&pmp->pmp_msg_touser)) {
644 DPRINTF(("puffs_getout: no outgoing op, "));
645 if (nonblock) {
646 DPRINTF(("returning EWOULDBLOCK\n"));
647 error = EWOULDBLOCK;
648 break;
649 }
650 DPRINTF(("waiting ...\n"));
651
652 error = cv_wait_sig(&pmp->pmp_msg_waiter_cv,
653 &pmp->pmp_lock);
654 if (error)
655 break;
656 else
657 continue;
658 }
659
660 park = TAILQ_FIRST(&pmp->pmp_msg_touser);
661 if (park == NULL)
662 continue;
663
664 mutex_enter(&park->park_mtx);
665 puffs_msgpark_reference(park);
666
667 DPRINTF(("puffs_getout: found park at %p, ", park));
668
669 /* If it's a goner, don't process any furher */
670 if (park->park_flags & PARKFLAG_WAITERGONE) {
671 DPRINTF(("waitergone!\n"));
672 puffs_msgpark_release(park);
673 continue;
674 }
675 preq = park->park_preq;
676
677 #if 0
678 /* check size */
679 /*
680 * XXX: this check is not valid for now, we don't know
681 * the size of the caller's input buffer. i.e. this
682 * will most likely go away
683 */
684 if (maxsize < preq->preq_frhdr.pfr_len) {
685 DPRINTF(("buffer too small\n"));
686 puffs_msgpark_release(park);
687 error = E2BIG;
688 break;
689 }
690 #endif
691
692 DPRINTF(("returning\n"));
693
694 /*
695 * Ok, we found what we came for. Release it from the
696 * outgoing queue but do not unlock. We will unlock
697 * only after we "releaseout" it to avoid complications:
698 * otherwise it is (theoretically) possible for userland
699 * to race us into "put" before we have a change to put
700 * this baby on the receiving queue.
701 */
702 TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries);
703 KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
704 park->park_flags &= ~PARKFLAG_ONQUEUE1;
705 mutex_exit(&park->park_mtx);
706
707 pmp->pmp_msg_touser_count--;
708 KASSERT(pmp->pmp_msg_touser_count >= 0);
709
710 break;
711 }
712 puffs_mp_release(pmp);
713 mutex_exit(&pmp->pmp_lock);
714
715 if (error == 0) {
716 *data = (uint8_t *)preq;
717 preq->preq_pth.pth_framelen = park->park_copylen;
718 *dlen = preq->preq_pth.pth_framelen;
719 *parkptr = park;
720 }
721
722 return error;
723 }
724
725 /*
726 * Release outgoing structure. Now, depending on the success of the
727 * outgoing send, it is either going onto the result waiting queue
728 * or the death chamber.
729 */
730 void
731 puffs_msgif_releaseout(void *ctx, void *parkptr, int status)
732 {
733 struct puffs_mount *pmp = ctx;
734 struct puffs_msgpark *park = parkptr;
735
736 DPRINTF(("puffs_releaseout: returning park %p, errno %d: " ,
737 park, status));
738 mutex_enter(&pmp->pmp_lock);
739 mutex_enter(&park->park_mtx);
740 if (park->park_flags & PARKFLAG_WANTREPLY) {
741 if (status == 0) {
742 DPRINTF(("enqueue replywait\n"));
743 TAILQ_INSERT_TAIL(&pmp->pmp_msg_replywait, park,
744 park_entries);
745 park->park_flags |= PARKFLAG_ONQUEUE2;
746 } else {
747 DPRINTF(("error path!\n"));
748 park->park_preq->preq_rv = status;
749 park->park_flags |= PARKFLAG_DONE;
750 cv_signal(&park->park_cv);
751 }
752 puffs_msgpark_release(park);
753 } else {
754 DPRINTF(("release\n"));
755 puffs_msgpark_release1(park, 2);
756 }
757 mutex_exit(&pmp->pmp_lock);
758 }
759
760 size_t
761 puffs_msgif_waitcount(void *ctx)
762 {
763 struct puffs_mount *pmp = ctx;
764 size_t rv;
765
766 mutex_enter(&pmp->pmp_lock);
767 rv = pmp->pmp_msg_touser_count;
768 mutex_exit(&pmp->pmp_lock);
769
770 return rv;
771 }
772
773 /*
774 * XXX: locking with this one?
775 */
776 static void
777 puffsop_msg(void *ctx, struct puffs_req *preq)
778 {
779 struct puffs_mount *pmp = ctx;
780 struct putter_hdr *pth = &preq->preq_pth;
781 struct puffs_msgpark *park;
782 int wgone;
783
784 mutex_enter(&pmp->pmp_lock);
785
786 /* Locate waiter */
787 TAILQ_FOREACH(park, &pmp->pmp_msg_replywait, park_entries) {
788 if (park->park_preq->preq_id == preq->preq_id)
789 break;
790 }
791 if (park == NULL) {
792 DPRINTF(("puffsop_msg: no request: %" PRIu64 "\n",
793 preq->preq_id));
794 mutex_exit(&pmp->pmp_lock);
795 return; /* XXX send error */
796 }
797
798 mutex_enter(&park->park_mtx);
799 puffs_msgpark_reference(park);
800 if (pth->pth_framelen > park->park_maxlen) {
801 DPRINTF(("puffsop_msg: invalid buffer length: "
802 "%" PRIu64 " (req %" PRIu64 ", \n", pth->pth_framelen,
803 preq->preq_id));
804 park->park_preq->preq_rv = EPROTO;
805 cv_signal(&park->park_cv);
806 puffs_msgpark_release1(park, 2);
807 mutex_exit(&pmp->pmp_lock);
808 return; /* XXX: error */
809 }
810 wgone = park->park_flags & PARKFLAG_WAITERGONE;
811
812 KASSERT(park->park_flags & PARKFLAG_ONQUEUE2);
813 TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries);
814 park->park_flags &= ~PARKFLAG_ONQUEUE2;
815 mutex_exit(&pmp->pmp_lock);
816
817 if (wgone) {
818 DPRINTF(("puffsop_msg: bad service - waiter gone for "
819 "park %p\n", park));
820 } else {
821 #if 1
822 if (park->park_creq) {
823 struct puffs_req *creq;
824 size_t csize;
825
826 KASSERT(pmp->pmp_docompat);
827 (void)puffs50_compat_hook_f2_call(preq,
828 park->park_creq);
829 creq = park->park_creq;
830 csize = park->park_creqlen;
831 park->park_creq = park->park_preq;
832 park->park_creqlen = park->park_maxlen;
833
834 park->park_preq = creq;
835 park->park_maxlen = csize;
836
837 memcpy(park->park_creq, preq, pth->pth_framelen);
838 } else {
839 #endif
840 memcpy(park->park_preq, preq, pth->pth_framelen);
841 }
842
843 if (park->park_flags & PARKFLAG_CALL) {
844 DPRINTF(("puffsop_msg: call for %p, arg %p\n",
845 park->park_preq, park->park_donearg));
846 park->park_done(pmp, preq, park->park_donearg);
847 }
848 }
849
850 if (!wgone) {
851 DPRINTF(("puffs_putop: flagging done for "
852 "park %p\n", park));
853 cv_signal(&park->park_cv);
854 }
855
856 park->park_flags |= PARKFLAG_DONE;
857 puffs_msgpark_release1(park, 2);
858 }
859
860 /*
861 * Node expiry. We come here after an inactive on an unexpired node.
862 * The expiry has been queued and is done in sop thread.
863 */
864 static void
865 puffsop_expire(struct puffs_mount *pmp, puffs_cookie_t cookie)
866 {
867 struct vnode *vp;
868
869 KASSERT(PUFFS_USE_FS_TTL(pmp));
870
871 /*
872 * If it still exists and has no reference,
873 * vrele should cause it to be reclaimed.
874 * Otherwise, we have nothing to do.
875 */
876 if (puffs_cookie2vnode(pmp, cookie, &vp) == 0) {
877 VPTOPP(vp)->pn_stat &= ~PNODE_SOPEXP;
878 vrele(vp);
879 }
880
881 return;
882 }
883
884 static void
885 puffsop_flush(struct puffs_mount *pmp, struct puffs_flush *pf)
886 {
887 struct vnode *vp;
888 voff_t offlo, offhi;
889 int rv, flags = 0;
890
891 KASSERT(pf->pf_req.preq_pth.pth_framelen == sizeof(struct puffs_flush));
892
893 /* XXX: slurry */
894 if (pf->pf_op == PUFFS_INVAL_NAMECACHE_ALL) {
895 cache_purgevfs(PMPTOMP(pmp));
896 rv = 0;
897 goto out;
898 }
899
900 /*
901 * Get vnode, don't lock it. Namecache is protected by its own lock
902 * and we have a reference to protect against premature harvesting.
903 *
904 * The node we want here might be locked and the op is in
905 * userspace waiting for us to complete ==> deadlock. Another
906 * reason we need to eventually bump locking to userspace, as we
907 * will need to lock the node if we wish to do flushes.
908 */
909 rv = puffs_cookie2vnode(pmp, pf->pf_cookie, &vp);
910 if (rv) {
911 if (rv == PUFFS_NOSUCHCOOKIE)
912 rv = ENOENT;
913 goto out;
914 }
915
916 switch (pf->pf_op) {
917 #if 0
918 /* not quite ready, yet */
919 case PUFFS_INVAL_NAMECACHE_NODE:
920 struct componentname *pf_cn;
921 char *name;
922 /* get comfortab^Wcomponentname */
923 pf_cn = kmem_alloc(componentname);
924 memset(pf_cn, 0, sizeof(struct componentname));
925 break;
926
927 #endif
928 case PUFFS_INVAL_NAMECACHE_DIR:
929 if (vp->v_type != VDIR) {
930 rv = EINVAL;
931 break;
932 }
933 cache_purge1(vp, NULL, 0, PURGE_CHILDREN);
934 break;
935
936 case PUFFS_INVAL_PAGECACHE_NODE_RANGE:
937 flags = PGO_FREE;
938 /*FALLTHROUGH*/
939 case PUFFS_FLUSH_PAGECACHE_NODE_RANGE:
940 if (flags == 0)
941 flags = PGO_CLEANIT;
942
943 if (pf->pf_end > vp->v_size || vp->v_type != VREG) {
944 rv = EINVAL;
945 break;
946 }
947
948 offlo = trunc_page(pf->pf_start);
949 offhi = round_page(pf->pf_end);
950 if (offhi != 0 && offlo >= offhi) {
951 rv = EINVAL;
952 break;
953 }
954
955 mutex_enter(vp->v_uobj.vmobjlock);
956 rv = VOP_PUTPAGES(vp, offlo, offhi, flags);
957 break;
958
959 default:
960 rv = EINVAL;
961 }
962
963 vrele(vp);
964
965 out:
966 puffs_msg_sendresp(pmp, &pf->pf_req, rv);
967 }
968
969 int
970 puffs_msgif_dispatch(void *ctx, struct putter_hdr *pth)
971 {
972 struct puffs_mount *pmp = ctx;
973 struct puffs_req *preq = (struct puffs_req *)pth;
974 struct puffs_sopreq *psopr;
975
976 if (pth->pth_framelen < sizeof(struct puffs_req)) {
977 puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */
978 return 0;
979 }
980
981 switch (PUFFSOP_OPCLASS(preq->preq_opclass)) {
982 case PUFFSOP_VN:
983 case PUFFSOP_VFS:
984 DPRINTF(("dispatch: vn/vfs message 0x%x\n", preq->preq_optype));
985 puffsop_msg(pmp, preq);
986 break;
987
988 case PUFFSOP_FLUSH: /* process in sop thread */
989 {
990 struct puffs_flush *pf;
991
992 DPRINTF(("dispatch: flush 0x%x\n", preq->preq_optype));
993
994 if (preq->preq_pth.pth_framelen != sizeof(struct puffs_flush)) {
995 puffs_msg_sendresp(pmp, preq, EINVAL); /* E2SMALL */
996 break;
997 }
998 pf = (struct puffs_flush *)preq;
999
1000 KASSERT(curlwp != uvm.pagedaemon_lwp);
1001 psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP);
1002 memcpy(&psopr->psopr_pf, pf, sizeof(*pf));
1003 psopr->psopr_sopreq = PUFFS_SOPREQ_FLUSH;
1004
1005 mutex_enter(&pmp->pmp_sopmtx);
1006 if (pmp->pmp_sopthrcount == 0) {
1007 mutex_exit(&pmp->pmp_sopmtx);
1008 kmem_free(psopr, sizeof(*psopr));
1009 puffs_msg_sendresp(pmp, preq, ENXIO);
1010 } else {
1011 TAILQ_INSERT_TAIL(&pmp->pmp_sopfastreqs,
1012 psopr, psopr_entries);
1013 cv_signal(&pmp->pmp_sopcv);
1014 mutex_exit(&pmp->pmp_sopmtx);
1015 }
1016 break;
1017 }
1018
1019 case PUFFSOP_UNMOUNT: /* process in sop thread */
1020 {
1021
1022 DPRINTF(("dispatch: unmount 0x%x\n", preq->preq_optype));
1023
1024 KASSERT(curlwp != uvm.pagedaemon_lwp);
1025 psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP);
1026 psopr->psopr_preq = *preq;
1027 psopr->psopr_sopreq = PUFFS_SOPREQ_UNMOUNT;
1028
1029 mutex_enter(&pmp->pmp_sopmtx);
1030 if (pmp->pmp_sopthrcount == 0) {
1031 mutex_exit(&pmp->pmp_sopmtx);
1032 kmem_free(psopr, sizeof(*psopr));
1033 puffs_msg_sendresp(pmp, preq, ENXIO);
1034 } else {
1035 TAILQ_INSERT_TAIL(&pmp->pmp_sopfastreqs,
1036 psopr, psopr_entries);
1037 cv_signal(&pmp->pmp_sopcv);
1038 mutex_exit(&pmp->pmp_sopmtx);
1039 }
1040 break;
1041 }
1042
1043 default:
1044 DPRINTF(("dispatch: invalid opclass 0x%x\n", preq->preq_opclass));
1045 puffs_msg_sendresp(pmp, preq, EOPNOTSUPP);
1046 break;
1047 }
1048
1049 return 0;
1050 }
1051
1052 /*
1053 * Work loop for thread processing all ops from server which
1054 * cannot safely be handled in caller context. This includes
1055 * everything which might need a lock currently "held" by the file
1056 * server, i.e. a long-term kernel lock which will be released only
1057 * once the file server acknowledges a request
1058 */
1059 #define TIMED_OUT(expire) \
1060 ((int)((unsigned int)hardclock_ticks - (unsigned int)expire) > 0)
1061 void
1062 puffs_sop_thread(void *arg)
1063 {
1064 struct puffs_mount *pmp = arg;
1065 struct mount *mp = PMPTOMP(pmp);
1066 struct puffs_sopreq *psopr;
1067 bool keeprunning;
1068 bool unmountme = false;
1069 int timeo;
1070
1071 timeo = PUFFS_USE_FS_TTL(pmp) ? puffs_sopreq_expire_timeout : 0;
1072
1073 mutex_enter(&pmp->pmp_sopmtx);
1074 for (keeprunning = true; keeprunning; ) {
1075 /*
1076 * We have a fast queue for flush and umount, and a node
1077 * queue for delayes node reclaims. Requests on node queue * are not honoured before clock reaches psopr_at. This
1078 * code assumes that requests are ordered by psopr_at.
1079 */
1080 do {
1081 psopr = TAILQ_FIRST(&pmp->pmp_sopfastreqs);
1082 if (psopr != NULL) {
1083 TAILQ_REMOVE(&pmp->pmp_sopfastreqs,
1084 psopr, psopr_entries);
1085 break;
1086 }
1087
1088 psopr = TAILQ_FIRST(&pmp->pmp_sopnodereqs);
1089 if ((psopr != NULL) && TIMED_OUT(psopr->psopr_at)) {
1090 TAILQ_REMOVE(&pmp->pmp_sopnodereqs,
1091 psopr, psopr_entries);
1092 break;
1093 }
1094
1095 cv_timedwait(&pmp->pmp_sopcv, &pmp->pmp_sopmtx, timeo);
1096 } while (1 /* CONSTCOND */);
1097
1098 mutex_exit(&pmp->pmp_sopmtx);
1099
1100 switch (psopr->psopr_sopreq) {
1101 case PUFFS_SOPREQSYS_EXIT:
1102 keeprunning = false;
1103 break;
1104 case PUFFS_SOPREQ_FLUSH:
1105 puffsop_flush(pmp, &psopr->psopr_pf);
1106 break;
1107 case PUFFS_SOPREQ_EXPIRE:
1108 puffsop_expire(pmp, psopr->psopr_ck);
1109 break;
1110 case PUFFS_SOPREQ_UNMOUNT:
1111 puffs_msg_sendresp(pmp, &psopr->psopr_preq, 0);
1112
1113 unmountme = true;
1114 keeprunning = false;
1115
1116 /*
1117 * We know the mountpoint is still alive because
1118 * the thread that is us (poetic?) is still alive.
1119 */
1120 vfs_ref(mp);
1121 break;
1122 }
1123
1124 kmem_free(psopr, sizeof(*psopr));
1125 mutex_enter(&pmp->pmp_sopmtx);
1126 }
1127
1128 /*
1129 * Purge remaining ops.
1130 */
1131 while ((psopr = TAILQ_FIRST(&pmp->pmp_sopfastreqs)) != NULL) {
1132 TAILQ_REMOVE(&pmp->pmp_sopfastreqs, psopr, psopr_entries);
1133 mutex_exit(&pmp->pmp_sopmtx);
1134 puffs_msg_sendresp(pmp, &psopr->psopr_preq, ENXIO);
1135 kmem_free(psopr, sizeof(*psopr));
1136 mutex_enter(&pmp->pmp_sopmtx);
1137 }
1138
1139 while ((psopr = TAILQ_FIRST(&pmp->pmp_sopnodereqs)) != NULL) {
1140 TAILQ_REMOVE(&pmp->pmp_sopnodereqs, psopr, psopr_entries);
1141 mutex_exit(&pmp->pmp_sopmtx);
1142 KASSERT(psopr->psopr_sopreq == PUFFS_SOPREQ_EXPIRE);
1143 kmem_free(psopr, sizeof(*psopr));
1144 mutex_enter(&pmp->pmp_sopmtx);
1145 }
1146
1147 pmp->pmp_sopthrcount--;
1148 cv_broadcast(&pmp->pmp_sopcv);
1149 mutex_exit(&pmp->pmp_sopmtx); /* not allowed to access fs after this */
1150
1151 /*
1152 * If unmount was requested, we can now safely do it here, since
1153 * our context is dead from the point-of-view of puffs_unmount()
1154 * and we are just another thread. dounmount() makes internally
1155 * sure that VFS_UNMOUNT() isn't called reentrantly and that it
1156 * is eventually completed.
1157 */
1158 if (unmountme) {
1159 (void)dounmount(mp, MNT_FORCE, curlwp);
1160 vfs_rele(mp);
1161 }
1162
1163 kthread_exit(0);
1164 }
1165
1166 int
1167 puffs_msgif_close(void *ctx)
1168 {
1169 struct puffs_mount *pmp = ctx;
1170 struct mount *mp = PMPTOMP(pmp);
1171
1172 mutex_enter(&pmp->pmp_lock);
1173 puffs_mp_reference(pmp);
1174
1175 /*
1176 * Free the waiting callers before proceeding any further.
1177 * The syncer might be jogging around in this file system
1178 * currently. If we allow it to go to the userspace of no
1179 * return while trying to get the syncer lock, well ...
1180 */
1181 puffs_userdead(pmp);
1182
1183 /*
1184 * Make sure someone from puffs_unmount() isn't currently in
1185 * userspace. If we don't take this precautionary step,
1186 * they might notice that the mountpoint has disappeared
1187 * from under them once they return. Especially note that we
1188 * cannot simply test for an unmounter before calling
1189 * dounmount(), since it might be possible that that particular
1190 * invocation of unmount was called without MNT_FORCE. Here we
1191 * *must* make sure unmount succeeds. Also, restart is necessary
1192 * since pmp isn't locked. We might end up with PUTTER_DEAD after
1193 * restart and exit from there.
1194 */
1195 if (pmp->pmp_unmounting) {
1196 cv_wait(&pmp->pmp_unmounting_cv, &pmp->pmp_lock);
1197 puffs_mp_release(pmp);
1198 mutex_exit(&pmp->pmp_lock);
1199 DPRINTF(("puffs_fop_close: unmount was in progress for pmp %p, "
1200 "restart\n", pmp));
1201 return ERESTART;
1202 }
1203
1204 /* Won't access pmp from here anymore */
1205 vfs_ref(mp);
1206 puffs_mp_release(pmp);
1207 mutex_exit(&pmp->pmp_lock);
1208
1209 /* Detach from VFS. */
1210 (void)dounmount(mp, MNT_FORCE, curlwp);
1211 vfs_rele(mp);
1212
1213 return 0;
1214 }
1215
1216 /*
1217 * We're dead, kaput, RIP, slightly more than merely pining for the
1218 * fjords, belly-up, fallen, lifeless, finished, expired, gone to meet
1219 * our maker, ceased to be, etcetc. YASD. It's a dead FS!
1220 *
1221 * Caller must hold puffs mutex.
1222 */
1223 void
1224 puffs_userdead(struct puffs_mount *pmp)
1225 {
1226 struct puffs_msgpark *park, *park_next;
1227
1228 /*
1229 * Mark filesystem status as dying so that operations don't
1230 * attempt to march to userspace any longer.
1231 */
1232 pmp->pmp_status = PUFFSTAT_DYING;
1233
1234 /* signal waiters on REQUEST TO file server queue */
1235 for (park = TAILQ_FIRST(&pmp->pmp_msg_touser); park; park = park_next) {
1236
1237 mutex_enter(&park->park_mtx);
1238 puffs_msgpark_reference(park);
1239 park_next = TAILQ_NEXT(park, park_entries);
1240
1241 KASSERT(park->park_flags & PARKFLAG_ONQUEUE1);
1242 TAILQ_REMOVE(&pmp->pmp_msg_touser, park, park_entries);
1243 park->park_flags &= ~PARKFLAG_ONQUEUE1;
1244 pmp->pmp_msg_touser_count--;
1245
1246 /*
1247 * Even though waiters on QUEUE1 are removed in touser()
1248 * in case of WAITERGONE, it is still possible for us to
1249 * get raced here due to having to retake locks in said
1250 * touser(). In the race case simply "ignore" the item
1251 * on the queue and move on to the next one.
1252 */
1253 if (park->park_flags & PARKFLAG_WAITERGONE) {
1254 KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
1255 KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
1256 puffs_msgpark_release(park);
1257
1258 } else {
1259 park->park_preq->preq_rv = ENXIO;
1260
1261 if (park->park_flags & PARKFLAG_CALL) {
1262 park->park_done(pmp, park->park_preq,
1263 park->park_donearg);
1264 puffs_msgpark_release1(park, 2);
1265 } else if ((park->park_flags & PARKFLAG_WANTREPLY)==0) {
1266 puffs_msgpark_release1(park, 2);
1267 } else {
1268 park->park_preq->preq_rv = ENXIO;
1269 cv_signal(&park->park_cv);
1270 puffs_msgpark_release(park);
1271 }
1272 }
1273 }
1274
1275 /* signal waiters on RESPONSE FROM file server queue */
1276 for (park=TAILQ_FIRST(&pmp->pmp_msg_replywait); park; park=park_next) {
1277 mutex_enter(&park->park_mtx);
1278 puffs_msgpark_reference(park);
1279 park_next = TAILQ_NEXT(park, park_entries);
1280
1281 KASSERT(park->park_flags & PARKFLAG_ONQUEUE2);
1282 KASSERT(park->park_flags & PARKFLAG_WANTREPLY);
1283
1284 TAILQ_REMOVE(&pmp->pmp_msg_replywait, park, park_entries);
1285 park->park_flags &= ~PARKFLAG_ONQUEUE2;
1286
1287 if (park->park_flags & PARKFLAG_WAITERGONE) {
1288 KASSERT((park->park_flags & PARKFLAG_CALL) == 0);
1289 puffs_msgpark_release(park);
1290 } else {
1291 park->park_preq->preq_rv = ENXIO;
1292 if (park->park_flags & PARKFLAG_CALL) {
1293 park->park_done(pmp, park->park_preq,
1294 park->park_donearg);
1295 puffs_msgpark_release1(park, 2);
1296 } else {
1297 cv_signal(&park->park_cv);
1298 puffs_msgpark_release(park);
1299 }
1300 }
1301 }
1302
1303 cv_broadcast(&pmp->pmp_msg_waiter_cv);
1304 }
1305