xref: /netbsd-src/sys/fs/puffs/puffs_vfsops.c (revision b7b7574d3bf8eeb51a1fa3977b59142ec6434a55)
1 /*	$NetBSD: puffs_vfsops.c,v 1.113 2014/05/25 19:32:36 christos Exp $	*/
2 
3 /*
4  * Copyright (c) 2005, 2006  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_vfsops.c,v 1.113 2014/05/25 19:32:36 christos Exp $");
34 
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/mount.h>
38 #include <sys/malloc.h>
39 #include <sys/extattr.h>
40 #include <sys/queue.h>
41 #include <sys/vnode.h>
42 #include <sys/dirent.h>
43 #include <sys/kauth.h>
44 #include <sys/proc.h>
45 #include <sys/module.h>
46 #include <sys/kthread.h>
47 
48 #include <uvm/uvm.h>
49 
50 #include <dev/putter/putter_sys.h>
51 
52 #include <miscfs/genfs/genfs.h>
53 
54 #include <fs/puffs/puffs_msgif.h>
55 #include <fs/puffs/puffs_sys.h>
56 
57 #include <lib/libkern/libkern.h>
58 
59 #include <nfs/nfsproto.h> /* for fh sizes */
60 
61 MODULE(MODULE_CLASS_VFS, puffs, "putter");
62 
63 VFS_PROTOS(puffs_vfsop);
64 
65 #ifndef PUFFS_PNODEBUCKETS
66 #define PUFFS_PNODEBUCKETS 256
67 #endif
68 #ifndef PUFFS_MAXPNODEBUCKETS
69 #define PUFFS_MAXPNODEBUCKETS 8192
70 #endif
71 int puffs_pnodebuckets_default = PUFFS_PNODEBUCKETS;
72 int puffs_maxpnodebuckets = PUFFS_MAXPNODEBUCKETS;
73 
74 #define BUCKETALLOC(a) (sizeof(struct puffs_pnode_hashlist *) * (a))
75 
76 static struct putter_ops puffs_putter = {
77 	.pop_getout	= puffs_msgif_getout,
78 	.pop_releaseout	= puffs_msgif_releaseout,
79 	.pop_waitcount	= puffs_msgif_waitcount,
80 	.pop_dispatch	= puffs_msgif_dispatch,
81 	.pop_close	= puffs_msgif_close,
82 };
83 
84 /*
85  * Try to ensure data structures used by the puffs protocol
86  * do not unexpectedly change.
87  */
88 #if defined(__i386__) && defined(__ELF__)
89 CTASSERT(sizeof(struct puffs_kargs) == 3928);
90 CTASSERT(sizeof(struct vattr) == 136);
91 CTASSERT(sizeof(struct puffs_req) == 44);
92 #endif
93 
94 int
95 puffs_vfsop_mount(struct mount *mp, const char *path, void *data,
96 	size_t *data_len)
97 {
98 	struct puffs_mount *pmp = NULL;
99 	struct puffs_kargs *args;
100 	char fstype[_VFS_NAMELEN];
101 	char *p;
102 	int error = 0, i;
103 	pid_t mntpid = curlwp->l_proc->p_pid;
104 
105 	if (data == NULL)
106 		return EINVAL;
107 	if (*data_len < sizeof *args)
108 		return EINVAL;
109 
110 	if (mp->mnt_flag & MNT_GETARGS) {
111 		pmp = MPTOPUFFSMP(mp);
112 		*(struct puffs_kargs *)data = pmp->pmp_args;
113 		*data_len = sizeof *args;
114 		return 0;
115 	}
116 
117 	/* update is not supported currently */
118 	if (mp->mnt_flag & MNT_UPDATE)
119 		return EOPNOTSUPP;
120 
121 	args = (struct puffs_kargs *)data;
122 
123 	if (args->pa_vers != PUFFSVERSION) {
124 		printf("puffs_mount: development version mismatch: "
125 		    "kernel %d, lib %d\n", PUFFSVERSION, args->pa_vers);
126 		error = EINVAL;
127 		goto out;
128 	}
129 
130 	if ((args->pa_flags & ~PUFFS_KFLAG_MASK) != 0) {
131 		printf("puffs_mount: invalid KFLAGs 0x%x\n", args->pa_flags);
132 		error = EINVAL;
133 		goto out;
134 	}
135 	if ((args->pa_fhflags & ~PUFFS_FHFLAG_MASK) != 0) {
136 		printf("puffs_mount: invalid FHFLAGs 0x%x\n", args->pa_fhflags);
137 		error = EINVAL;
138 		goto out;
139 	}
140 
141 	for (i = 0; i < __arraycount(args->pa_spare); i++) {
142 		if (args->pa_spare[i] != 0) {
143 			printf("puffs_mount: pa_spare[%d] = 0x%x\n",
144 			    i, args->pa_spare[i]);
145 			error = EINVAL;
146 			goto out;
147 		}
148 	}
149 
150 	/* use dummy value for passthrough */
151 	if (args->pa_fhflags & PUFFS_FHFLAG_PASSTHROUGH)
152 		args->pa_fhsize = sizeof(struct fid);
153 
154 	/* sanitize file handle length */
155 	if (PUFFS_TOFHSIZE(args->pa_fhsize) > FHANDLE_SIZE_MAX) {
156 		printf("puffs_mount: handle size %zu too large\n",
157 		    args->pa_fhsize);
158 		error = EINVAL;
159 		goto out;
160 	}
161 	/* sanity check file handle max sizes */
162 	if (args->pa_fhsize && args->pa_fhflags & PUFFS_FHFLAG_PROTOMASK) {
163 		size_t kfhsize = PUFFS_TOFHSIZE(args->pa_fhsize);
164 
165 		if (args->pa_fhflags & PUFFS_FHFLAG_NFSV2) {
166 			if (NFSX_FHTOOBIG_P(kfhsize, 0)) {
167 				printf("puffs_mount: fhsize larger than "
168 				    "NFSv2 max %d\n",
169 				    PUFFS_FROMFHSIZE(NFSX_V2FH));
170 				error = EINVAL;
171 				goto out;
172 			}
173 		}
174 
175 		if (args->pa_fhflags & PUFFS_FHFLAG_NFSV3) {
176 			if (NFSX_FHTOOBIG_P(kfhsize, 1)) {
177 				printf("puffs_mount: fhsize larger than "
178 				    "NFSv3 max %d\n",
179 				    PUFFS_FROMFHSIZE(NFSX_V3FHMAX));
180 				error = EINVAL;
181 				goto out;
182 			}
183 		}
184 	}
185 
186 	/* don't allow non-printing characters (like my sweet umlauts.. snif) */
187 	args->pa_typename[sizeof(args->pa_typename)-1] = '\0';
188 	for (p = args->pa_typename; *p; p++)
189 		if (*p < ' ' || *p > '~')
190 			*p = '.';
191 
192 	args->pa_mntfromname[sizeof(args->pa_mntfromname)-1] = '\0';
193 	for (p = args->pa_mntfromname; *p; p++)
194 		if (*p < ' ' || *p > '~')
195 			*p = '.';
196 
197 	/* build real name */
198 	(void)strlcpy(fstype, PUFFS_TYPEPREFIX, sizeof(fstype));
199 	(void)strlcat(fstype, args->pa_typename, sizeof(fstype));
200 
201 	/* inform user server if it got the max request size it wanted */
202 	if (args->pa_maxmsglen == 0 || args->pa_maxmsglen > PUFFS_MSG_MAXSIZE)
203 		args->pa_maxmsglen = PUFFS_MSG_MAXSIZE;
204 	else if (args->pa_maxmsglen < 2*PUFFS_MSGSTRUCT_MAX)
205 		args->pa_maxmsglen = 2*PUFFS_MSGSTRUCT_MAX;
206 
207 	(void)strlcpy(args->pa_typename, fstype, sizeof(args->pa_typename));
208 
209 	if (args->pa_nhashbuckets == 0)
210 		args->pa_nhashbuckets = puffs_pnodebuckets_default;
211 	if (args->pa_nhashbuckets < 1)
212 		args->pa_nhashbuckets = 1;
213 	if (args->pa_nhashbuckets > PUFFS_MAXPNODEBUCKETS) {
214 		args->pa_nhashbuckets = puffs_maxpnodebuckets;
215 		printf("puffs_mount: using %d hash buckets. "
216 		    "adjust puffs_maxpnodebuckets for more\n",
217 		    puffs_maxpnodebuckets);
218 	}
219 
220 	error = set_statvfs_info(path, UIO_USERSPACE, args->pa_mntfromname,
221 	    UIO_SYSSPACE, fstype, mp, curlwp);
222 	if (error)
223 		goto out;
224 	mp->mnt_stat.f_iosize = DEV_BSIZE;
225 	mp->mnt_stat.f_namemax = args->pa_svfsb.f_namemax;
226 
227 	/*
228 	 * We can't handle the VFS_STATVFS() mount_domount() does
229 	 * after VFS_MOUNT() because we'd deadlock, so handle it
230 	 * here already.
231 	 */
232 	copy_statvfs_info(&args->pa_svfsb, mp);
233 	(void)memcpy(&mp->mnt_stat, &args->pa_svfsb, sizeof(mp->mnt_stat));
234 
235 	KASSERT(curlwp != uvm.pagedaemon_lwp);
236 	pmp = kmem_zalloc(sizeof(struct puffs_mount), KM_SLEEP);
237 
238 	mp->mnt_fs_bshift = DEV_BSHIFT;
239 	mp->mnt_dev_bshift = DEV_BSHIFT;
240 	mp->mnt_flag &= ~MNT_LOCAL; /* we don't really know, so ... */
241 	mp->mnt_data = pmp;
242 
243 #if 0
244 	/*
245 	 * XXX: puffs code is MPSAFE.  However, VFS really isn't.
246 	 * Currently, there is nothing which protects an inode from
247 	 * reclaim while there are threads inside the file system.
248 	 * This means that in the event of a server crash, an MPSAFE
249 	 * mount is likely to end up accessing invalid memory.  For the
250 	 * non-mpsafe case, the kernel lock, general structure of
251 	 * puffs and pmp_refcount protect the threads during escape.
252 	 *
253 	 * Fixing this will require:
254 	 *  a) fixing vfs
255 	 * OR
256 	 *  b) adding a small sleep to puffs_msgif_close() between
257 	 *     userdead() and dounmount().
258 	 *     (well, this isn't really a fix, but would solve
259 	 *     99.999% of the race conditions).
260 	 *
261 	 * Also, in the event of "b", unmount -f should be used,
262 	 * like with any other file system, sparingly and only when
263 	 * it is "known" to be safe.
264 	 */
265 	mp->mnt_iflags |= IMNT_MPSAFE;
266 #endif
267 
268 	pmp->pmp_status = PUFFSTAT_MOUNTING;
269 	pmp->pmp_mp = mp;
270 	pmp->pmp_msg_maxsize = args->pa_maxmsglen;
271 	pmp->pmp_args = *args;
272 
273 	pmp->pmp_npnodehash = args->pa_nhashbuckets;
274 	pmp->pmp_pnodehash = kmem_alloc(BUCKETALLOC(pmp->pmp_npnodehash),
275 	    KM_SLEEP);
276 	for (i = 0; i < pmp->pmp_npnodehash; i++)
277 		LIST_INIT(&pmp->pmp_pnodehash[i]);
278 	LIST_INIT(&pmp->pmp_newcookie);
279 
280 	/*
281 	 * Inform the fileops processing code that we have a mountpoint.
282 	 * If it doesn't know about anyone with our pid/fd having the
283 	 * device open, punt
284 	 */
285 	if ((pmp->pmp_pi
286 	    = putter_attach(mntpid, args->pa_fd, pmp, &puffs_putter)) == NULL) {
287 		error = ENOENT;
288 		goto out;
289 	}
290 
291 	/* XXX: check parameters */
292 	pmp->pmp_root_cookie = args->pa_root_cookie;
293 	pmp->pmp_root_vtype = args->pa_root_vtype;
294 	pmp->pmp_root_vsize = args->pa_root_vsize;
295 	pmp->pmp_root_rdev = args->pa_root_rdev;
296 	pmp->pmp_docompat = args->pa_time32;
297 
298 	mutex_init(&pmp->pmp_lock, MUTEX_DEFAULT, IPL_NONE);
299 	mutex_init(&pmp->pmp_sopmtx, MUTEX_DEFAULT, IPL_NONE);
300 	cv_init(&pmp->pmp_msg_waiter_cv, "puffsget");
301 	cv_init(&pmp->pmp_refcount_cv, "puffsref");
302 	cv_init(&pmp->pmp_unmounting_cv, "puffsum");
303 	cv_init(&pmp->pmp_sopcv, "puffsop");
304 	TAILQ_INIT(&pmp->pmp_msg_touser);
305 	TAILQ_INIT(&pmp->pmp_msg_replywait);
306 	TAILQ_INIT(&pmp->pmp_sopfastreqs);
307 	TAILQ_INIT(&pmp->pmp_sopnodereqs);
308 
309 	if ((error = kthread_create(PRI_NONE, KTHREAD_MPSAFE, NULL,
310 	    puffs_sop_thread, pmp, NULL, "puffsop")) != 0)
311 		goto out;
312 	pmp->pmp_sopthrcount = 1;
313 
314 	DPRINTF(("puffs_mount: mount point at %p, puffs specific at %p\n",
315 	    mp, MPTOPUFFSMP(mp)));
316 
317 	vfs_getnewfsid(mp);
318 
319  out:
320 	if (error && pmp && pmp->pmp_pi)
321 		putter_detach(pmp->pmp_pi);
322 	if (error && pmp && pmp->pmp_pnodehash)
323 		kmem_free(pmp->pmp_pnodehash, BUCKETALLOC(pmp->pmp_npnodehash));
324 	if (error && pmp)
325 		kmem_free(pmp, sizeof(struct puffs_mount));
326 	return error;
327 }
328 
329 int
330 puffs_vfsop_start(struct mount *mp, int flags)
331 {
332 	struct puffs_mount *pmp = MPTOPUFFSMP(mp);
333 
334 	KASSERT(pmp->pmp_status == PUFFSTAT_MOUNTING);
335 	pmp->pmp_status = PUFFSTAT_RUNNING;
336 
337 	return 0;
338 }
339 
340 int
341 puffs_vfsop_unmount(struct mount *mp, int mntflags)
342 {
343 	PUFFS_MSG_VARS(vfs, unmount);
344 	struct puffs_mount *pmp;
345 	int error, force;
346 
347 	error = 0;
348 	force = mntflags & MNT_FORCE;
349 	pmp = MPTOPUFFSMP(mp);
350 
351 	DPRINTF(("puffs_unmount: detach filesystem from vfs, current "
352 	    "status 0x%x\n", pmp->pmp_status));
353 
354 	/*
355 	 * flush all the vnodes.  VOP_RECLAIM() takes care that the
356 	 * root vnode does not get flushed until unmount.  The
357 	 * userspace root node cookie is stored in the mount
358 	 * structure, so we can always re-instantiate a root vnode,
359 	 * should userspace unmount decide it doesn't want to
360 	 * cooperate.
361 	 */
362 	error = vflush(mp, NULLVP, force ? FORCECLOSE : 0);
363 	if (error)
364 		goto out;
365 
366 	/*
367 	 * If we are not DYING, we should ask userspace's opinion
368 	 * about the situation
369 	 */
370 	mutex_enter(&pmp->pmp_lock);
371 	if (pmp->pmp_status != PUFFSTAT_DYING) {
372 		pmp->pmp_unmounting = 1;
373 		mutex_exit(&pmp->pmp_lock);
374 
375 		PUFFS_MSG_ALLOC(vfs, unmount);
376 		puffs_msg_setinfo(park_unmount,
377 		    PUFFSOP_VFS, PUFFS_VFS_UNMOUNT, NULL);
378 		unmount_msg->pvfsr_flags = mntflags;
379 
380 		PUFFS_MSG_ENQUEUEWAIT(pmp, park_unmount, error);
381 		PUFFS_MSG_RELEASE(unmount);
382 
383 		error = checkerr(pmp, error, __func__);
384 		DPRINTF(("puffs_unmount: error %d force %d\n", error, force));
385 
386 		mutex_enter(&pmp->pmp_lock);
387 		pmp->pmp_unmounting = 0;
388 		cv_broadcast(&pmp->pmp_unmounting_cv);
389 	}
390 
391 	/*
392 	 * if userspace cooperated or we really need to die,
393 	 * screw what userland thinks and just die.
394 	 */
395 	if (error == 0 || force) {
396 		struct puffs_sopreq *psopr;
397 
398 		/* tell waiters & other resources to go unwait themselves */
399 		puffs_userdead(pmp);
400 		putter_detach(pmp->pmp_pi);
401 
402 		/*
403 		 * Wait until there are no more users for the mount resource.
404 		 * Notice that this is hooked against transport_close
405 		 * and return from touser.  In an ideal world, it would
406 		 * be hooked against final return from all operations.
407 		 * But currently it works well enough, since nobody
408 		 * does weird blocking voodoo after return from touser().
409 		 */
410 		while (pmp->pmp_refcount != 0)
411 			cv_wait(&pmp->pmp_refcount_cv, &pmp->pmp_lock);
412 		mutex_exit(&pmp->pmp_lock);
413 
414 		/*
415 		 * Release kernel thread now that there is nothing
416 		 * it would be wanting to lock.
417 		 */
418 		KASSERT(curlwp != uvm.pagedaemon_lwp);
419 		psopr = kmem_alloc(sizeof(*psopr), KM_SLEEP);
420 		psopr->psopr_sopreq = PUFFS_SOPREQSYS_EXIT;
421 		mutex_enter(&pmp->pmp_sopmtx);
422 		if (pmp->pmp_sopthrcount == 0) {
423 			mutex_exit(&pmp->pmp_sopmtx);
424 			kmem_free(psopr, sizeof(*psopr));
425 			mutex_enter(&pmp->pmp_sopmtx);
426 			KASSERT(pmp->pmp_sopthrcount == 0);
427 		} else {
428 			TAILQ_INSERT_TAIL(&pmp->pmp_sopfastreqs,
429 			    psopr, psopr_entries);
430 			cv_signal(&pmp->pmp_sopcv);
431 		}
432 		while (pmp->pmp_sopthrcount > 0)
433 			cv_wait(&pmp->pmp_sopcv, &pmp->pmp_sopmtx);
434 		mutex_exit(&pmp->pmp_sopmtx);
435 
436 		/* free resources now that we hopefully have no waiters left */
437 		cv_destroy(&pmp->pmp_unmounting_cv);
438 		cv_destroy(&pmp->pmp_refcount_cv);
439 		cv_destroy(&pmp->pmp_msg_waiter_cv);
440 		cv_destroy(&pmp->pmp_sopcv);
441 		mutex_destroy(&pmp->pmp_lock);
442 		mutex_destroy(&pmp->pmp_sopmtx);
443 
444 		kmem_free(pmp->pmp_pnodehash, BUCKETALLOC(pmp->pmp_npnodehash));
445 		kmem_free(pmp, sizeof(struct puffs_mount));
446 		error = 0;
447 	} else {
448 		mutex_exit(&pmp->pmp_lock);
449 	}
450 
451  out:
452 	DPRINTF(("puffs_unmount: return %d\n", error));
453 	return error;
454 }
455 
456 /*
457  * This doesn't need to travel to userspace
458  */
459 int
460 puffs_vfsop_root(struct mount *mp, struct vnode **vpp)
461 {
462 	struct puffs_mount *pmp = MPTOPUFFSMP(mp);
463 	int rv;
464 
465 	rv = puffs_cookie2vnode(pmp, pmp->pmp_root_cookie, 1, 1, vpp);
466 	KASSERT(rv != PUFFS_NOSUCHCOOKIE);
467 	return rv;
468 }
469 
470 int
471 puffs_vfsop_statvfs(struct mount *mp, struct statvfs *sbp)
472 {
473 	PUFFS_MSG_VARS(vfs, statvfs);
474 	struct puffs_mount *pmp;
475 	int error = 0;
476 
477 	pmp = MPTOPUFFSMP(mp);
478 
479 	/*
480 	 * If we are mounting, it means that the userspace counterpart
481 	 * is calling mount(2), but mount(2) also calls statvfs.  So
482 	 * requesting statvfs from userspace would mean a deadlock.
483 	 * Compensate.
484 	 */
485 	if (__predict_false(pmp->pmp_status == PUFFSTAT_MOUNTING))
486 		return EINPROGRESS;
487 
488 	PUFFS_MSG_ALLOC(vfs, statvfs);
489 	puffs_msg_setinfo(park_statvfs, PUFFSOP_VFS, PUFFS_VFS_STATVFS, NULL);
490 
491 	PUFFS_MSG_ENQUEUEWAIT(pmp, park_statvfs, error);
492 	error = checkerr(pmp, error, __func__);
493 	statvfs_msg->pvfsr_sb.f_iosize = DEV_BSIZE;
494 
495 	/*
496 	 * Try to produce a sensible result even in the event
497 	 * of userspace error.
498 	 *
499 	 * XXX: cache the copy in non-error case
500 	 */
501 	if (!error) {
502 		copy_statvfs_info(&statvfs_msg->pvfsr_sb, mp);
503 		(void)memcpy(sbp, &statvfs_msg->pvfsr_sb,
504 		    sizeof(struct statvfs));
505 	} else {
506 		copy_statvfs_info(sbp, mp);
507 	}
508 
509 	PUFFS_MSG_RELEASE(statvfs);
510 	return error;
511 }
512 
513 static bool
514 pageflush_selector(void *cl, struct vnode *vp)
515 {
516 	return vp->v_type == VREG &&
517 	    !(LIST_EMPTY(&vp->v_dirtyblkhd) && UVM_OBJ_IS_CLEAN(&vp->v_uobj));
518 }
519 
520 static int
521 pageflush(struct mount *mp, kauth_cred_t cred, int waitfor)
522 {
523 	struct puffs_node *pn;
524 	struct vnode *vp;
525 	struct vnode_iterator *marker;
526 	int error, rv, fsyncwait;
527 
528 	error = 0;
529 	fsyncwait = (waitfor == MNT_WAIT) ? FSYNC_WAIT : 0;
530 
531 	/*
532 	 * Sync all cached data from regular vnodes (which are not
533 	 * currently locked, see below).  After this we call VFS_SYNC
534 	 * for the fs server, which should handle data and metadata for
535 	 * all the nodes it knows to exist.
536 	 */
537 	vfs_vnode_iterator_init(mp, &marker);
538 	while ((vp = vfs_vnode_iterator_next(marker, pageflush_selector,
539 	    NULL)))
540 	{
541 		/*
542 		 * Here we try to get a reference to the vnode and to
543 		 * lock it.  This is mostly cargo-culted, but I will
544 		 * offer an explanation to why I believe this might
545 		 * actually do the right thing.
546 		 *
547 		 * If the vnode is a goner, we quite obviously don't need
548 		 * to sync it.
549 		 *
550 		 * If the vnode was busy, we don't need to sync it because
551 		 * this is never called with MNT_WAIT except from
552 		 * dounmount(), when we are wait-flushing all the dirty
553 		 * vnodes through other routes in any case.  So there,
554 		 * sync() doesn't actually sync.  Happy now?
555 		 */
556 		error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
557 		if (error) {
558 			vrele(vp);
559 			continue;
560 		}
561 		pn = VPTOPP(vp);
562 		/* hmm.. is the FAF thing entirely sensible? */
563 		if (waitfor == MNT_LAZY) {
564 			mutex_enter(vp->v_interlock);
565 			pn->pn_stat |= PNODE_FAF;
566 			mutex_exit(vp->v_interlock);
567 		}
568 		rv = VOP_FSYNC(vp, cred, fsyncwait, 0, 0);
569 		if (waitfor == MNT_LAZY) {
570 			mutex_enter(vp->v_interlock);
571 			pn->pn_stat &= ~PNODE_FAF;
572 			mutex_exit(vp->v_interlock);
573 		}
574 		if (rv)
575 			error = rv;
576 		vput(vp);
577 	}
578 	vfs_vnode_iterator_destroy(marker);
579 
580 	return error;
581 }
582 
583 int
584 puffs_vfsop_sync(struct mount *mp, int waitfor, struct kauth_cred *cred)
585 {
586 	PUFFS_MSG_VARS(vfs, sync);
587 	struct puffs_mount *pmp = MPTOPUFFSMP(mp);
588 	int error, rv;
589 
590 	error = pageflush(mp, cred, waitfor);
591 
592 	/* sync fs */
593 	PUFFS_MSG_ALLOC(vfs, sync);
594 	sync_msg->pvfsr_waitfor = waitfor;
595 	puffs_credcvt(&sync_msg->pvfsr_cred, cred);
596 	puffs_msg_setinfo(park_sync, PUFFSOP_VFS, PUFFS_VFS_SYNC, NULL);
597 
598 	PUFFS_MSG_ENQUEUEWAIT(pmp, park_sync, rv);
599 	rv = checkerr(pmp, rv, __func__);
600 	if (rv)
601 		error = rv;
602 
603 	PUFFS_MSG_RELEASE(sync);
604 	return error;
605 }
606 
607 int
608 puffs_vfsop_fhtovp(struct mount *mp, struct fid *fhp, struct vnode **vpp)
609 {
610 	PUFFS_MSG_VARS(vfs, fhtonode);
611 	struct puffs_mount *pmp = MPTOPUFFSMP(mp);
612 	struct vnode *vp;
613 	void *fhdata;
614 	size_t argsize, fhlen;
615 	int error;
616 
617 	if (pmp->pmp_args.pa_fhsize == 0)
618 		return EOPNOTSUPP;
619 
620 	if (pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_PASSTHROUGH) {
621 		fhlen = fhp->fid_len;
622 		fhdata = fhp;
623 	} else {
624 		fhlen = PUFFS_FROMFHSIZE(fhp->fid_len);
625 		fhdata = fhp->fid_data;
626 
627 		if (pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_DYNAMIC) {
628 			if (pmp->pmp_args.pa_fhsize < fhlen)
629 				return EINVAL;
630 		} else {
631 			if (pmp->pmp_args.pa_fhsize != fhlen)
632 				return EINVAL;
633 		}
634 	}
635 
636 	argsize = sizeof(struct puffs_vfsmsg_fhtonode) + fhlen;
637 	puffs_msgmem_alloc(argsize, &park_fhtonode, (void *)&fhtonode_msg, 1);
638 	fhtonode_msg->pvfsr_dsize = fhlen;
639 	memcpy(fhtonode_msg->pvfsr_data, fhdata, fhlen);
640 	puffs_msg_setinfo(park_fhtonode, PUFFSOP_VFS, PUFFS_VFS_FHTOVP, NULL);
641 
642 	PUFFS_MSG_ENQUEUEWAIT(pmp, park_fhtonode, error);
643 	error = checkerr(pmp, error, __func__);
644 	if (error)
645 		goto out;
646 
647 	error = puffs_cookie2vnode(pmp, fhtonode_msg->pvfsr_fhcookie, 1,1,&vp);
648 	DPRINTF(("puffs_fhtovp: got cookie %p, existing vnode %p\n",
649 	    fhtonode_msg->pvfsr_fhcookie, vp));
650 	if (error == PUFFS_NOSUCHCOOKIE) {
651 		error = puffs_getvnode(mp, fhtonode_msg->pvfsr_fhcookie,
652 		    fhtonode_msg->pvfsr_vtype, fhtonode_msg->pvfsr_size,
653 		    fhtonode_msg->pvfsr_rdev, &vp);
654 		if (error)
655 			goto out;
656 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
657 	} else if (error) {
658 		goto out;
659 	}
660 
661 	*vpp = vp;
662  out:
663 	puffs_msgmem_release(park_fhtonode);
664 	return error;
665 }
666 
667 int
668 puffs_vfsop_vptofh(struct vnode *vp, struct fid *fhp, size_t *fh_size)
669 {
670 	PUFFS_MSG_VARS(vfs, nodetofh);
671 	struct puffs_mount *pmp = MPTOPUFFSMP(vp->v_mount);
672 	size_t argsize, fhlen;
673 	int error;
674 
675 	if (pmp->pmp_args.pa_fhsize == 0)
676 		return EOPNOTSUPP;
677 
678 	/* if file handles are static len, we can test len immediately */
679 	if (((pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_DYNAMIC) == 0)
680 	    && ((pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_PASSTHROUGH) == 0)
681 	    && (PUFFS_FROMFHSIZE(*fh_size) < pmp->pmp_args.pa_fhsize)) {
682 		*fh_size = PUFFS_TOFHSIZE(pmp->pmp_args.pa_fhsize);
683 		return E2BIG;
684 	}
685 
686 	if (pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_PASSTHROUGH)
687 		fhlen = *fh_size;
688 	else
689 		fhlen = PUFFS_FROMFHSIZE(*fh_size);
690 
691 	argsize = sizeof(struct puffs_vfsmsg_nodetofh) + fhlen;
692 	puffs_msgmem_alloc(argsize, &park_nodetofh, (void *)&nodetofh_msg, 1);
693 	nodetofh_msg->pvfsr_fhcookie = VPTOPNC(vp);
694 	nodetofh_msg->pvfsr_dsize = fhlen;
695 	puffs_msg_setinfo(park_nodetofh, PUFFSOP_VFS, PUFFS_VFS_VPTOFH, NULL);
696 
697 	PUFFS_MSG_ENQUEUEWAIT(pmp, park_nodetofh, error);
698 	error = checkerr(pmp, error, __func__);
699 
700 	if (pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_PASSTHROUGH)
701 		fhlen = nodetofh_msg->pvfsr_dsize;
702 	else if (pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_DYNAMIC)
703 		fhlen = PUFFS_TOFHSIZE(nodetofh_msg->pvfsr_dsize);
704 	else
705 		fhlen = PUFFS_TOFHSIZE(pmp->pmp_args.pa_fhsize);
706 
707 	if (error) {
708 		if (error == E2BIG)
709 			*fh_size = fhlen;
710 		goto out;
711 	}
712 
713 	if (fhlen > FHANDLE_SIZE_MAX) {
714 		puffs_senderr(pmp, PUFFS_ERR_VPTOFH, E2BIG,
715 		    "file handle too big", VPTOPNC(vp));
716 		error = EPROTO;
717 		goto out;
718 	}
719 
720 	if (*fh_size < fhlen) {
721 		*fh_size = fhlen;
722 		error = E2BIG;
723 		goto out;
724 	}
725 	*fh_size = fhlen;
726 
727 	if (fhp) {
728 		if (pmp->pmp_args.pa_fhflags & PUFFS_FHFLAG_PASSTHROUGH) {
729 			memcpy(fhp, nodetofh_msg->pvfsr_data, fhlen);
730 		} else {
731 			fhp->fid_len = *fh_size;
732 			memcpy(fhp->fid_data, nodetofh_msg->pvfsr_data,
733 			    nodetofh_msg->pvfsr_dsize);
734 		}
735 	}
736 
737  out:
738 	puffs_msgmem_release(park_nodetofh);
739 	return error;
740 }
741 
742 void
743 puffs_vfsop_init(void)
744 {
745 
746 	/* some checks depend on this */
747 	KASSERT(VNOVAL == VSIZENOTSET);
748 
749 	pool_init(&puffs_pnpool, sizeof(struct puffs_node), 0, 0, 0,
750 	    "puffpnpl", &pool_allocator_nointr, IPL_NONE);
751 	pool_init(&puffs_vapool, sizeof(struct vattr), 0, 0, 0,
752 	    "puffvapl", &pool_allocator_nointr, IPL_NONE);
753 	puffs_msgif_init();
754 }
755 
756 void
757 puffs_vfsop_done(void)
758 {
759 
760 	puffs_msgif_destroy();
761 	pool_destroy(&puffs_pnpool);
762 	pool_destroy(&puffs_vapool);
763 }
764 
765 int
766 puffs_vfsop_snapshot(struct mount *mp, struct vnode *vp, struct timespec *ts)
767 {
768 
769 	return EOPNOTSUPP;
770 }
771 
772 int
773 puffs_vfsop_extattrctl(struct mount *mp, int cmd, struct vnode *vp,
774 	int attrnamespace, const char *attrname)
775 {
776 	PUFFS_MSG_VARS(vfs, extattrctl);
777 	struct puffs_mount *pmp = MPTOPUFFSMP(mp);
778 	struct puffs_node *pnp;
779 	puffs_cookie_t pnc;
780 	int error, flags;
781 
782 	if (vp) {
783 		/* doesn't make sense for puffs servers */
784 		if (vp->v_mount != mp)
785 			return EXDEV;
786 		pnp = vp->v_data;
787 		pnc = pnp->pn_cookie;
788 		flags = PUFFS_EXTATTRCTL_HASNODE;
789 	} else {
790 		pnp = pnc = NULL;
791 		flags = 0;
792 	}
793 
794 	PUFFS_MSG_ALLOC(vfs, extattrctl);
795 	extattrctl_msg->pvfsr_cmd = cmd;
796 	extattrctl_msg->pvfsr_attrnamespace = attrnamespace;
797 	extattrctl_msg->pvfsr_flags = flags;
798 	if (attrname) {
799 		strlcpy(extattrctl_msg->pvfsr_attrname, attrname,
800 		    sizeof(extattrctl_msg->pvfsr_attrname));
801 		extattrctl_msg->pvfsr_flags |= PUFFS_EXTATTRCTL_HASATTRNAME;
802 	}
803 	puffs_msg_setinfo(park_extattrctl,
804 	    PUFFSOP_VFS, PUFFS_VFS_EXTATTRCTL, pnc);
805 
806 	puffs_msg_enqueue(pmp, park_extattrctl);
807 	if (vp) {
808 		mutex_enter(&pnp->pn_mtx);
809 		puffs_referencenode(pnp);
810 		mutex_exit(&pnp->pn_mtx);
811 		VOP_UNLOCK(vp);
812 	}
813 	error = puffs_msg_wait2(pmp, park_extattrctl, pnp, NULL);
814 	PUFFS_MSG_RELEASE(extattrctl);
815 	if (vp) {
816 		puffs_releasenode(pnp);
817 	}
818 
819 	return checkerr(pmp, error, __func__);
820 }
821 
822 const struct vnodeopv_desc * const puffs_vnodeopv_descs[] = {
823 	&puffs_vnodeop_opv_desc,
824 	&puffs_specop_opv_desc,
825 	&puffs_fifoop_opv_desc,
826 	&puffs_msgop_opv_desc,
827 	NULL,
828 };
829 
830 struct vfsops puffs_vfsops = {
831 	.vfs_name = MOUNT_PUFFS,
832 	.vfs_min_mount_data = sizeof (struct puffs_kargs),
833 	.vfs_mount = puffs_vfsop_mount,
834 	.vfs_start = puffs_vfsop_start,
835 	.vfs_unmount = puffs_vfsop_unmount,
836 	.vfs_root = puffs_vfsop_root,
837 	.vfs_quotactl = (void *)eopnotsupp,
838 	.vfs_statvfs = puffs_vfsop_statvfs,
839 	.vfs_sync = puffs_vfsop_sync,
840 	.vfs_vget = (void *)eopnotsupp,
841 	.vfs_fhtovp = puffs_vfsop_fhtovp,
842 	.vfs_vptofh = puffs_vfsop_vptofh,
843 	.vfs_init = puffs_vfsop_init,
844 	.vfs_done = puffs_vfsop_done,
845 	.vfs_snapshot = puffs_vfsop_snapshot,
846 	.vfs_extattrctl = puffs_vfsop_extattrctl,
847 	.vfs_suspendctl = (void *)eopnotsupp,
848 	.vfs_renamelock_enter = genfs_renamelock_enter,
849 	.vfs_renamelock_exit = genfs_renamelock_exit,
850 	.vfs_fsync = (void *)eopnotsupp,
851 	.vfs_opv_descs = puffs_vnodeopv_descs
852 };
853 
854 static int
855 puffs_modcmd(modcmd_t cmd, void *arg)
856 {
857 
858 	switch (cmd) {
859 	case MODULE_CMD_INIT:
860 		return vfs_attach(&puffs_vfsops);
861 	case MODULE_CMD_FINI:
862 		return vfs_detach(&puffs_vfsops);
863 	default:
864 		return ENOTTY;
865 	}
866 }
867