xref: /netbsd-src/sys/miscfs/specfs/spec_vnops.c (revision 5f2f42719cd62ff11fd913b40b7ce19f07c4fd25)
1 /*	$NetBSD: spec_vnops.c,v 1.215 2022/09/21 10:59:10 riastradh Exp $	*/
2 
3 /*-
4  * Copyright (c) 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
17  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
18  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
19  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
20  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
21  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
22  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
23  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
24  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
25  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
26  * POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * Copyright (c) 1989, 1993
31  *	The Regents of the University of California.  All rights reserved.
32  *
33  * Redistribution and use in source and binary forms, with or without
34  * modification, are permitted provided that the following conditions
35  * are met:
36  * 1. Redistributions of source code must retain the above copyright
37  *    notice, this list of conditions and the following disclaimer.
38  * 2. Redistributions in binary form must reproduce the above copyright
39  *    notice, this list of conditions and the following disclaimer in the
40  *    documentation and/or other materials provided with the distribution.
41  * 3. Neither the name of the University nor the names of its contributors
42  *    may be used to endorse or promote products derived from this software
43  *    without specific prior written permission.
44  *
45  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
46  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
47  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
48  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
49  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
50  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
51  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
52  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
53  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
54  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
55  * SUCH DAMAGE.
56  *
57  *	@(#)spec_vnops.c	8.15 (Berkeley) 7/14/95
58  */
59 
60 #include <sys/cdefs.h>
61 __KERNEL_RCSID(0, "$NetBSD: spec_vnops.c,v 1.215 2022/09/21 10:59:10 riastradh Exp $");
62 
63 #include <sys/param.h>
64 #include <sys/proc.h>
65 #include <sys/systm.h>
66 #include <sys/kernel.h>
67 #include <sys/conf.h>
68 #include <sys/buf.h>
69 #include <sys/mount.h>
70 #include <sys/namei.h>
71 #include <sys/vnode_impl.h>
72 #include <sys/stat.h>
73 #include <sys/errno.h>
74 #include <sys/ioctl.h>
75 #include <sys/poll.h>
76 #include <sys/file.h>
77 #include <sys/disklabel.h>
78 #include <sys/disk.h>
79 #include <sys/lockf.h>
80 #include <sys/tty.h>
81 #include <sys/kauth.h>
82 #include <sys/fstrans.h>
83 #include <sys/module.h>
84 #include <sys/atomic.h>
85 
86 #include <miscfs/genfs/genfs.h>
87 #include <miscfs/specfs/specdev.h>
88 
89 /*
90  * Lock order:
91  *
92  *	vnode lock
93  *	-> device_lock
94  *	-> struct vnode::v_interlock
95  */
96 
97 /* symbolic sleep message strings for devices */
98 const char	devopn[] = "devopn";
99 const char	devio[] = "devio";
100 const char	devwait[] = "devwait";
101 const char	devin[] = "devin";
102 const char	devout[] = "devout";
103 const char	devioc[] = "devioc";
104 const char	devcls[] = "devcls";
105 
106 #define	SPECHSZ	64
107 #if	((SPECHSZ&(SPECHSZ-1)) == 0)
108 #define	SPECHASH(rdev)	(((rdev>>5)+(rdev))&(SPECHSZ-1))
109 #else
110 #define	SPECHASH(rdev)	(((unsigned)((rdev>>5)+(rdev)))%SPECHSZ)
111 #endif
112 
113 static vnode_t	*specfs_hash[SPECHSZ];
114 extern struct mount *dead_rootmount;
115 
116 /*
117  * This vnode operations vector is used for special device nodes
118  * created from whole cloth by the kernel.  For the ops vector for
119  * vnodes built from special devices found in a filesystem, see (e.g)
120  * ffs_specop_entries[] in ffs_vnops.c or the equivalent for other
121  * filesystems.
122  */
123 
124 int (**spec_vnodeop_p)(void *);
125 const struct vnodeopv_entry_desc spec_vnodeop_entries[] = {
126 	{ &vop_default_desc, vn_default_error },
127 	{ &vop_parsepath_desc, genfs_parsepath },	/* parsepath */
128 	{ &vop_lookup_desc, spec_lookup },		/* lookup */
129 	{ &vop_create_desc, genfs_badop },		/* create */
130 	{ &vop_mknod_desc, genfs_badop },		/* mknod */
131 	{ &vop_open_desc, spec_open },			/* open */
132 	{ &vop_close_desc, spec_close },		/* close */
133 	{ &vop_access_desc, genfs_ebadf },		/* access */
134 	{ &vop_accessx_desc, genfs_ebadf },		/* accessx */
135 	{ &vop_getattr_desc, genfs_ebadf },		/* getattr */
136 	{ &vop_setattr_desc, genfs_ebadf },		/* setattr */
137 	{ &vop_read_desc, spec_read },			/* read */
138 	{ &vop_write_desc, spec_write },		/* write */
139 	{ &vop_fallocate_desc, genfs_eopnotsupp },	/* fallocate */
140 	{ &vop_fdiscard_desc, spec_fdiscard },		/* fdiscard */
141 	{ &vop_fcntl_desc, genfs_fcntl },		/* fcntl */
142 	{ &vop_ioctl_desc, spec_ioctl },		/* ioctl */
143 	{ &vop_poll_desc, spec_poll },			/* poll */
144 	{ &vop_kqfilter_desc, spec_kqfilter },		/* kqfilter */
145 	{ &vop_revoke_desc, genfs_revoke },		/* revoke */
146 	{ &vop_mmap_desc, spec_mmap },			/* mmap */
147 	{ &vop_fsync_desc, spec_fsync },		/* fsync */
148 	{ &vop_seek_desc, spec_seek },			/* seek */
149 	{ &vop_remove_desc, genfs_badop },		/* remove */
150 	{ &vop_link_desc, genfs_badop },		/* link */
151 	{ &vop_rename_desc, genfs_badop },		/* rename */
152 	{ &vop_mkdir_desc, genfs_badop },		/* mkdir */
153 	{ &vop_rmdir_desc, genfs_badop },		/* rmdir */
154 	{ &vop_symlink_desc, genfs_badop },		/* symlink */
155 	{ &vop_readdir_desc, genfs_badop },		/* readdir */
156 	{ &vop_readlink_desc, genfs_badop },		/* readlink */
157 	{ &vop_abortop_desc, genfs_badop },		/* abortop */
158 	{ &vop_inactive_desc, spec_inactive },		/* inactive */
159 	{ &vop_reclaim_desc, spec_reclaim },		/* reclaim */
160 	{ &vop_lock_desc, genfs_lock },			/* lock */
161 	{ &vop_unlock_desc, genfs_unlock },		/* unlock */
162 	{ &vop_bmap_desc, spec_bmap },			/* bmap */
163 	{ &vop_strategy_desc, spec_strategy },		/* strategy */
164 	{ &vop_print_desc, spec_print },		/* print */
165 	{ &vop_islocked_desc, genfs_islocked },		/* islocked */
166 	{ &vop_pathconf_desc, spec_pathconf },		/* pathconf */
167 	{ &vop_advlock_desc, spec_advlock },		/* advlock */
168 	{ &vop_bwrite_desc, vn_bwrite },		/* bwrite */
169 	{ &vop_getpages_desc, genfs_getpages },		/* getpages */
170 	{ &vop_putpages_desc, genfs_putpages },		/* putpages */
171 	{ NULL, NULL }
172 };
173 const struct vnodeopv_desc spec_vnodeop_opv_desc =
174 	{ &spec_vnodeop_p, spec_vnodeop_entries };
175 
176 static kauth_listener_t rawio_listener;
177 static struct kcondvar specfs_iocv;
178 
179 /* Returns true if vnode is /dev/mem or /dev/kmem. */
180 bool
181 iskmemvp(struct vnode *vp)
182 {
183 	return ((vp->v_type == VCHR) && iskmemdev(vp->v_rdev));
184 }
185 
186 /*
187  * Returns true if dev is /dev/mem or /dev/kmem.
188  */
189 int
190 iskmemdev(dev_t dev)
191 {
192 	/* mem_no is emitted by config(8) to generated devsw.c */
193 	extern const int mem_no;
194 
195 	/* minor 14 is /dev/io on i386 with COMPAT_10 */
196 	return (major(dev) == mem_no && (minor(dev) < 2 || minor(dev) == 14));
197 }
198 
199 static int
200 rawio_listener_cb(kauth_cred_t cred, kauth_action_t action, void *cookie,
201     void *arg0, void *arg1, void *arg2, void *arg3)
202 {
203 	int result;
204 
205 	result = KAUTH_RESULT_DEFER;
206 
207 	if ((action != KAUTH_DEVICE_RAWIO_SPEC) &&
208 	    (action != KAUTH_DEVICE_RAWIO_PASSTHRU))
209 		return result;
210 
211 	/* Access is mandated by permissions. */
212 	result = KAUTH_RESULT_ALLOW;
213 
214 	return result;
215 }
216 
217 void
218 spec_init(void)
219 {
220 
221 	rawio_listener = kauth_listen_scope(KAUTH_SCOPE_DEVICE,
222 	    rawio_listener_cb, NULL);
223 	cv_init(&specfs_iocv, "specio");
224 }
225 
226 /*
227  * spec_io_enter(vp, &sn, &dev)
228  *
229  *	Enter an operation that may not hold vp's vnode lock or an
230  *	fstrans on vp's mount.  Until spec_io_exit, the vnode will not
231  *	be revoked.
232  *
233  *	On success, set sn to the specnode pointer and dev to the dev_t
234  *	number and return zero.  Caller must later call spec_io_exit
235  *	when done.
236  *
237  *	On failure, return ENXIO -- the device has been revoked and no
238  *	longer exists.
239  */
240 static int
241 spec_io_enter(struct vnode *vp, struct specnode **snp, dev_t *devp)
242 {
243 	dev_t dev;
244 	struct specnode *sn;
245 	unsigned iocnt;
246 	int error = 0;
247 
248 	mutex_enter(vp->v_interlock);
249 
250 	/*
251 	 * Extract all the info we need from the vnode, unless the
252 	 * vnode has already been reclaimed.  This can happen if the
253 	 * underlying device has been removed and all the device nodes
254 	 * for it have been revoked.  The caller may not hold a vnode
255 	 * lock or fstrans to prevent this from happening before it has
256 	 * had an opportunity to notice the vnode is dead.
257 	 */
258 	if (vdead_check(vp, VDEAD_NOWAIT) != 0 ||
259 	    (sn = vp->v_specnode) == NULL ||
260 	    (dev = vp->v_rdev) == NODEV) {
261 		error = ENXIO;
262 		goto out;
263 	}
264 
265 	/*
266 	 * Notify spec_close that we are doing an I/O operation which
267 	 * may not be not bracketed by fstrans(9) and thus is not
268 	 * blocked by vfs suspension.
269 	 *
270 	 * We could hold this reference with psref(9) instead, but we
271 	 * already have to take the interlock for vdead_check, so
272 	 * there's not much more cost here to another atomic operation.
273 	 */
274 	do {
275 		iocnt = atomic_load_relaxed(&sn->sn_dev->sd_iocnt);
276 		if (__predict_false(iocnt == UINT_MAX)) {
277 			/*
278 			 * The I/O count is limited by the number of
279 			 * LWPs (which will never overflow this) --
280 			 * unless one driver uses another driver via
281 			 * specfs, which is rather unusual, but which
282 			 * could happen via pud(4) userspace drivers.
283 			 * We could use a 64-bit count, but can't use
284 			 * atomics for that on all platforms.
285 			 * (Probably better to switch to psref or
286 			 * localcount instead.)
287 			 */
288 			error = EBUSY;
289 			goto out;
290 		}
291 	} while (atomic_cas_uint(&sn->sn_dev->sd_iocnt, iocnt, iocnt + 1)
292 	    != iocnt);
293 
294 	/* Success!  */
295 	*snp = sn;
296 	*devp = dev;
297 	error = 0;
298 
299 out:	mutex_exit(vp->v_interlock);
300 	return error;
301 }
302 
303 /*
304  * spec_io_exit(vp, sn)
305  *
306  *	Exit an operation entered with a successful spec_io_enter --
307  *	allow concurrent spec_node_revoke to proceed.  The argument sn
308  *	must match the struct specnode pointer returned by spec_io_exit
309  *	for vp.
310  */
311 static void
312 spec_io_exit(struct vnode *vp, struct specnode *sn)
313 {
314 	struct specdev *sd = sn->sn_dev;
315 	unsigned iocnt;
316 
317 	KASSERT(vp->v_specnode == sn);
318 
319 	/*
320 	 * We are done.  Notify spec_close if appropriate.  The
321 	 * transition of 1 -> 0 must happen under device_lock so
322 	 * spec_close doesn't miss a wakeup.
323 	 */
324 	do {
325 		iocnt = atomic_load_relaxed(&sd->sd_iocnt);
326 		KASSERT(iocnt > 0);
327 		if (iocnt == 1) {
328 			mutex_enter(&device_lock);
329 			if (atomic_dec_uint_nv(&sd->sd_iocnt) == 0)
330 				cv_broadcast(&specfs_iocv);
331 			mutex_exit(&device_lock);
332 			break;
333 		}
334 	} while (atomic_cas_uint(&sd->sd_iocnt, iocnt, iocnt - 1) != iocnt);
335 }
336 
337 /*
338  * spec_io_drain(sd)
339  *
340  *	Wait for all existing spec_io_enter/exit sections to complete.
341  *	Caller must ensure spec_io_enter will fail at this point.
342  */
343 static void
344 spec_io_drain(struct specdev *sd)
345 {
346 
347 	/*
348 	 * I/O at the same time as closing is unlikely -- it often
349 	 * indicates an application bug.
350 	 */
351 	if (__predict_true(atomic_load_relaxed(&sd->sd_iocnt) == 0))
352 		return;
353 
354 	mutex_enter(&device_lock);
355 	while (atomic_load_relaxed(&sd->sd_iocnt) > 0)
356 		cv_wait(&specfs_iocv, &device_lock);
357 	mutex_exit(&device_lock);
358 }
359 
360 /*
361  * Initialize a vnode that represents a device.
362  */
363 void
364 spec_node_init(vnode_t *vp, dev_t rdev)
365 {
366 	specnode_t *sn;
367 	specdev_t *sd;
368 	vnode_t *vp2;
369 	vnode_t **vpp;
370 
371 	KASSERT(vp->v_type == VBLK || vp->v_type == VCHR);
372 	KASSERT(vp->v_specnode == NULL);
373 
374 	/*
375 	 * Search the hash table for this device.  If known, add a
376 	 * reference to the device structure.  If not known, create
377 	 * a new entry to represent the device.  In all cases add
378 	 * the vnode to the hash table.
379 	 */
380 	sn = kmem_alloc(sizeof(*sn), KM_SLEEP);
381 	sd = kmem_alloc(sizeof(*sd), KM_SLEEP);
382 	mutex_enter(&device_lock);
383 	vpp = &specfs_hash[SPECHASH(rdev)];
384 	for (vp2 = *vpp; vp2 != NULL; vp2 = vp2->v_specnext) {
385 		KASSERT(vp2->v_specnode != NULL);
386 		if (rdev == vp2->v_rdev && vp->v_type == vp2->v_type) {
387 			break;
388 		}
389 	}
390 	if (vp2 == NULL) {
391 		/* No existing record, create a new one. */
392 		sd->sd_rdev = rdev;
393 		sd->sd_mountpoint = NULL;
394 		sd->sd_lockf = NULL;
395 		sd->sd_refcnt = 1;
396 		sd->sd_opencnt = 0;
397 		sd->sd_bdevvp = NULL;
398 		sd->sd_iocnt = 0;
399 		sd->sd_opened = false;
400 		sd->sd_closing = false;
401 		sn->sn_dev = sd;
402 		sd = NULL;
403 	} else {
404 		/* Use the existing record. */
405 		sn->sn_dev = vp2->v_specnode->sn_dev;
406 		sn->sn_dev->sd_refcnt++;
407 	}
408 	/* Insert vnode into the hash chain. */
409 	sn->sn_opencnt = 0;
410 	sn->sn_rdev = rdev;
411 	sn->sn_gone = false;
412 	vp->v_specnode = sn;
413 	vp->v_specnext = *vpp;
414 	*vpp = vp;
415 	mutex_exit(&device_lock);
416 
417 	/* Free the record we allocated if unused. */
418 	if (sd != NULL) {
419 		kmem_free(sd, sizeof(*sd));
420 	}
421 }
422 
423 /*
424  * Lookup a vnode by device number and return it referenced.
425  */
426 int
427 spec_node_lookup_by_dev(enum vtype type, dev_t dev, int flags, vnode_t **vpp)
428 {
429 	int error;
430 	vnode_t *vp;
431 
432 top:	mutex_enter(&device_lock);
433 	for (vp = specfs_hash[SPECHASH(dev)]; vp; vp = vp->v_specnext) {
434 		if (type == vp->v_type && dev == vp->v_rdev) {
435 			mutex_enter(vp->v_interlock);
436 			/* If clean or being cleaned, then ignore it. */
437 			if (vdead_check(vp, VDEAD_NOWAIT) == 0)
438 				break;
439 			if ((flags & VDEAD_NOWAIT) == 0) {
440 				mutex_exit(&device_lock);
441 				/*
442 				 * It may be being revoked as we speak,
443 				 * and the caller wants to wait until
444 				 * all revocation has completed.  Let
445 				 * vcache_vget wait for it to finish
446 				 * dying; as a side effect, vcache_vget
447 				 * releases vp->v_interlock.  Note that
448 				 * vcache_vget cannot succeed at this
449 				 * point because vdead_check already
450 				 * failed.
451 				 */
452 				error = vcache_vget(vp);
453 				KASSERT(error);
454 				goto top;
455 			}
456 			mutex_exit(vp->v_interlock);
457 		}
458 	}
459 	KASSERT(vp == NULL || mutex_owned(vp->v_interlock));
460 	if (vp == NULL) {
461 		mutex_exit(&device_lock);
462 		return ENOENT;
463 	}
464 	/*
465 	 * If it is an opened block device return the opened vnode.
466 	 */
467 	if (type == VBLK && vp->v_specnode->sn_dev->sd_bdevvp != NULL) {
468 		mutex_exit(vp->v_interlock);
469 		vp = vp->v_specnode->sn_dev->sd_bdevvp;
470 		mutex_enter(vp->v_interlock);
471 	}
472 	mutex_exit(&device_lock);
473 	error = vcache_vget(vp);
474 	if (error != 0)
475 		return error;
476 	*vpp = vp;
477 
478 	return 0;
479 }
480 
481 /*
482  * Lookup a vnode by file system mounted on and return it referenced.
483  */
484 int
485 spec_node_lookup_by_mount(struct mount *mp, vnode_t **vpp)
486 {
487 	int i, error;
488 	vnode_t *vp, *vq;
489 
490 	mutex_enter(&device_lock);
491 	for (i = 0, vq = NULL; i < SPECHSZ && vq == NULL; i++) {
492 		for (vp = specfs_hash[i]; vp; vp = vp->v_specnext) {
493 			if (vp->v_type != VBLK)
494 				continue;
495 			vq = vp->v_specnode->sn_dev->sd_bdevvp;
496 			if (vq != NULL &&
497 			    vq->v_specnode->sn_dev->sd_mountpoint == mp)
498 				break;
499 			vq = NULL;
500 		}
501 	}
502 	if (vq == NULL) {
503 		mutex_exit(&device_lock);
504 		return ENOENT;
505 	}
506 	mutex_enter(vq->v_interlock);
507 	mutex_exit(&device_lock);
508 	error = vcache_vget(vq);
509 	if (error != 0)
510 		return error;
511 	*vpp = vq;
512 
513 	return 0;
514 
515 }
516 
517 /*
518  * Get the file system mounted on this block device.
519  *
520  * XXX Caller should hold the vnode lock -- shared or exclusive -- so
521  * that this can't changed, and the vnode can't be revoked while we
522  * examine it.  But not all callers do, and they're scattered through a
523  * lot of file systems, so we can't assert this yet.
524  */
525 struct mount *
526 spec_node_getmountedfs(vnode_t *devvp)
527 {
528 	struct mount *mp;
529 
530 	KASSERT(devvp->v_type == VBLK);
531 	mp = devvp->v_specnode->sn_dev->sd_mountpoint;
532 
533 	return mp;
534 }
535 
536 /*
537  * Set the file system mounted on this block device.
538  *
539  * XXX Caller should hold the vnode lock exclusively so this can't be
540  * changed or assumed by spec_node_getmountedfs while we change it, and
541  * the vnode can't be revoked while we handle it.  But not all callers
542  * do, and they're scattered through a lot of file systems, so we can't
543  * assert this yet.  Instead, for now, we'll take an I/O reference so
544  * at least the ioctl doesn't race with revoke/detach.
545  *
546  * If you do change this to assert an exclusive vnode lock, you must
547  * also do vdead_check before trying bdev_ioctl, because the vnode may
548  * have been revoked by the time the caller locked it, and this is
549  * _not_ a vop -- calls to spec_node_setmountedfs don't go through
550  * v_op, so revoking the vnode doesn't prevent further calls.
551  *
552  * XXX Caller should additionally have the vnode open, at least if mp
553  * is nonnull, but I'm not sure all callers do that -- need to audit.
554  * Currently udf closes the vnode before clearing the mount.
555  */
556 void
557 spec_node_setmountedfs(vnode_t *devvp, struct mount *mp)
558 {
559 	struct dkwedge_info dkw;
560 	struct specnode *sn;
561 	dev_t dev;
562 	int error;
563 
564 	KASSERT(devvp->v_type == VBLK);
565 
566 	error = spec_io_enter(devvp, &sn, &dev);
567 	if (error)
568 		return;
569 
570 	KASSERT(sn->sn_dev->sd_mountpoint == NULL || mp == NULL);
571 	sn->sn_dev->sd_mountpoint = mp;
572 	if (mp == NULL)
573 		goto out;
574 
575 	error = bdev_ioctl(dev, DIOCGWEDGEINFO, &dkw, FREAD, curlwp);
576 	if (error)
577 		goto out;
578 
579 	strlcpy(mp->mnt_stat.f_mntfromlabel, dkw.dkw_wname,
580 	    sizeof(mp->mnt_stat.f_mntfromlabel));
581 
582 out:	spec_io_exit(devvp, sn);
583 }
584 
585 /*
586  * A vnode representing a special device is going away.  Close
587  * the device if the vnode holds it open.
588  */
589 void
590 spec_node_revoke(vnode_t *vp)
591 {
592 	specnode_t *sn;
593 	specdev_t *sd;
594 	struct vnode **vpp;
595 
596 	KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
597 
598 	sn = vp->v_specnode;
599 	sd = sn->sn_dev;
600 
601 	KASSERT(vp->v_type == VBLK || vp->v_type == VCHR);
602 	KASSERT(vp->v_specnode != NULL);
603 	KASSERT(sn->sn_gone == false);
604 
605 	mutex_enter(&device_lock);
606 	KASSERTMSG(sn->sn_opencnt <= sd->sd_opencnt,
607 	    "sn_opencnt=%u > sd_opencnt=%u",
608 	    sn->sn_opencnt, sd->sd_opencnt);
609 	sn->sn_gone = true;
610 	if (sn->sn_opencnt != 0) {
611 		sd->sd_opencnt -= (sn->sn_opencnt - 1);
612 		sn->sn_opencnt = 1;
613 		mutex_exit(&device_lock);
614 
615 		VOP_CLOSE(vp, FNONBLOCK, NOCRED);
616 
617 		mutex_enter(&device_lock);
618 		KASSERT(sn->sn_opencnt == 0);
619 	}
620 
621 	/*
622 	 * We may have revoked the vnode in this thread while another
623 	 * thread was in the middle of spec_close, in the window when
624 	 * spec_close releases the vnode lock to call .d_close for the
625 	 * last close.  In that case, wait for the concurrent
626 	 * spec_close to complete.
627 	 */
628 	while (sd->sd_closing)
629 		cv_wait(&specfs_iocv, &device_lock);
630 
631 	/*
632 	 * Remove from the hash so lookups stop returning this
633 	 * specnode.  We will dissociate it from the specdev -- and
634 	 * possibly free the specdev -- in spec_node_destroy.
635 	 */
636 	KASSERT(sn->sn_gone);
637 	KASSERT(sn->sn_opencnt == 0);
638 	for (vpp = &specfs_hash[SPECHASH(vp->v_rdev)];;
639 	     vpp = &(*vpp)->v_specnext) {
640 		if (*vpp == vp) {
641 			*vpp = vp->v_specnext;
642 			vp->v_specnext = NULL;
643 			break;
644 		}
645 	}
646 	mutex_exit(&device_lock);
647 }
648 
649 /*
650  * A vnode representing a special device is being recycled.
651  * Destroy the specfs component.
652  */
653 void
654 spec_node_destroy(vnode_t *vp)
655 {
656 	specnode_t *sn;
657 	specdev_t *sd;
658 	int refcnt;
659 
660 	sn = vp->v_specnode;
661 	sd = sn->sn_dev;
662 
663 	KASSERT(vp->v_type == VBLK || vp->v_type == VCHR);
664 	KASSERT(vp->v_specnode != NULL);
665 	KASSERT(sn->sn_opencnt == 0);
666 
667 	mutex_enter(&device_lock);
668 	sn = vp->v_specnode;
669 	vp->v_specnode = NULL;
670 	refcnt = sd->sd_refcnt--;
671 	KASSERT(refcnt > 0);
672 	mutex_exit(&device_lock);
673 
674 	/* If the device is no longer in use, destroy our record. */
675 	if (refcnt == 1) {
676 		KASSERT(sd->sd_iocnt == 0);
677 		KASSERT(sd->sd_opencnt == 0);
678 		KASSERT(sd->sd_bdevvp == NULL);
679 		kmem_free(sd, sizeof(*sd));
680 	}
681 	kmem_free(sn, sizeof(*sn));
682 }
683 
684 /*
685  * Trivial lookup routine that always fails.
686  */
687 int
688 spec_lookup(void *v)
689 {
690 	struct vop_lookup_v2_args /* {
691 		struct vnode *a_dvp;
692 		struct vnode **a_vpp;
693 		struct componentname *a_cnp;
694 	} */ *ap = v;
695 
696 	*ap->a_vpp = NULL;
697 	return (ENOTDIR);
698 }
699 
700 typedef int (*spec_ioctl_t)(dev_t, u_long, void *, int, struct lwp *);
701 
702 /*
703  * Open a special file.
704  */
705 /* ARGSUSED */
706 int
707 spec_open(void *v)
708 {
709 	struct vop_open_args /* {
710 		struct vnode *a_vp;
711 		int  a_mode;
712 		kauth_cred_t a_cred;
713 	} */ *ap = v;
714 	struct lwp *l = curlwp;
715 	struct vnode *vp = ap->a_vp;
716 	dev_t dev, dev1;
717 	int error;
718 	enum kauth_device_req req;
719 	specnode_t *sn, *sn1;
720 	specdev_t *sd;
721 	spec_ioctl_t ioctl;
722 	u_int gen = 0;
723 	const char *name = NULL;
724 	bool needclose = false;
725 	struct partinfo pi;
726 
727 	KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
728 	KASSERTMSG(vp->v_type == VBLK || vp->v_type == VCHR, "type=%d",
729 	    vp->v_type);
730 
731 	dev = vp->v_rdev;
732 	sn = vp->v_specnode;
733 	sd = sn->sn_dev;
734 
735 	/*
736 	 * Don't allow open if fs is mounted -nodev.
737 	 */
738 	if (vp->v_mount && (vp->v_mount->mnt_flag & MNT_NODEV))
739 		return (ENXIO);
740 
741 	switch (ap->a_mode & (FREAD | FWRITE)) {
742 	case FREAD | FWRITE:
743 		req = KAUTH_REQ_DEVICE_RAWIO_SPEC_RW;
744 		break;
745 	case FWRITE:
746 		req = KAUTH_REQ_DEVICE_RAWIO_SPEC_WRITE;
747 		break;
748 	default:
749 		req = KAUTH_REQ_DEVICE_RAWIO_SPEC_READ;
750 		break;
751 	}
752 	error = kauth_authorize_device_spec(ap->a_cred, req, vp);
753 	if (error != 0)
754 		return (error);
755 
756 	/*
757 	 * Acquire an open reference -- as long as we hold onto it, and
758 	 * the vnode isn't revoked, it can't be closed, and the vnode
759 	 * can't be revoked until we release the vnode lock.
760 	 */
761 	mutex_enter(&device_lock);
762 	KASSERT(!sn->sn_gone);
763 	switch (vp->v_type) {
764 	case VCHR:
765 		/*
766 		 * Character devices can accept opens from multiple
767 		 * vnodes.  But first, wait for any close to finish.
768 		 * Wait under the vnode lock so we don't have to worry
769 		 * about the vnode being revoked while we wait.
770 		 */
771 		while (sd->sd_closing) {
772 			error = cv_wait_sig(&specfs_iocv, &device_lock);
773 			if (error)
774 				break;
775 		}
776 		if (error)
777 			break;
778 		sd->sd_opencnt++;
779 		sn->sn_opencnt++;
780 		KASSERTMSG(sn->sn_opencnt <= sd->sd_opencnt,
781 		    "sn_opencnt=%u > sd_opencnt=%u",
782 		    sn->sn_opencnt, sd->sd_opencnt);
783 		break;
784 	case VBLK:
785 		/*
786 		 * For block devices, permit only one open.  The buffer
787 		 * cache cannot remain self-consistent with multiple
788 		 * vnodes holding a block device open.
789 		 *
790 		 * Treat zero opencnt with non-NULL mountpoint as open.
791 		 * This may happen after forced detach of a mounted device.
792 		 *
793 		 * Also treat sd_closing, meaning there is a concurrent
794 		 * close in progress, as still open.
795 		 */
796 		if (sd->sd_opencnt != 0 ||
797 		    sd->sd_mountpoint != NULL ||
798 		    sd->sd_closing) {
799 			error = EBUSY;
800 			break;
801 		}
802 		KASSERTMSG(sn->sn_opencnt == 0, "sn_opencnt=%u",
803 		    sn->sn_opencnt);
804 		sn->sn_opencnt = 1;
805 		sd->sd_opencnt = 1;
806 		sd->sd_bdevvp = vp;
807 		break;
808 	default:
809 		panic("invalid specfs vnode type: %d", vp->v_type);
810 	}
811 	mutex_exit(&device_lock);
812 	if (error)
813 		return error;
814 
815 	/*
816 	 * Set VV_ISTTY if this is a tty cdev.
817 	 *
818 	 * XXX This does the wrong thing if the module has to be
819 	 * autoloaded.  We should maybe set this after autoloading
820 	 * modules and calling .d_open successfully, except (a) we need
821 	 * the vnode lock to touch it, and (b) once we acquire the
822 	 * vnode lock again, the vnode may have been revoked, and
823 	 * deadfs's dead_read needs VV_ISTTY to be already set in order
824 	 * to return the right answer.  So this needs some additional
825 	 * synchronization to be made to work correctly with tty driver
826 	 * module autoload.  For now, let's just hope it doesn't cause
827 	 * too much trouble for a tty from an autoloaded driver module
828 	 * to fail with EIO instead of returning EOF.
829 	 */
830 	if (vp->v_type == VCHR) {
831 		if (cdev_type(dev) == D_TTY)
832 			vp->v_vflag |= VV_ISTTY;
833 	}
834 
835 	/*
836 	 * Because opening the device may block indefinitely, e.g. when
837 	 * opening a tty, and loading a module may cross into many
838 	 * other subsystems, we must not hold the vnode lock while
839 	 * calling .d_open, so release it now and reacquire it when
840 	 * done.
841 	 *
842 	 * Take an I/O reference so that any concurrent spec_close via
843 	 * spec_node_revoke will wait for us to finish calling .d_open.
844 	 * The vnode can't be dead at this point because we have it
845 	 * locked.  Note that if revoked, the driver must interrupt
846 	 * .d_open before spec_close starts waiting for I/O to drain so
847 	 * this doesn't deadlock.
848 	 */
849 	VOP_UNLOCK(vp);
850 	error = spec_io_enter(vp, &sn1, &dev1);
851 	if (error) {
852 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
853 		return error;
854 	}
855 	KASSERT(sn1 == sn);
856 	KASSERT(dev1 == dev);
857 
858 	/*
859 	 * Open the device.  If .d_open returns ENXIO (device not
860 	 * configured), the driver may not be loaded, so try
861 	 * autoloading a module and then try .d_open again if anything
862 	 * got loaded.
863 	 */
864 	switch (vp->v_type) {
865 	case VCHR:
866 		do {
867 			const struct cdevsw *cdev;
868 
869 			gen = module_gen;
870 			error = cdev_open(dev, ap->a_mode, S_IFCHR, l);
871 			if (error != ENXIO)
872 				break;
873 
874 			/* Check if we already have a valid driver */
875 			mutex_enter(&device_lock);
876 			cdev = cdevsw_lookup(dev);
877 			mutex_exit(&device_lock);
878 			if (cdev != NULL)
879 				break;
880 
881 			/* Get device name from devsw_conv array */
882 			if ((name = cdevsw_getname(major(dev))) == NULL)
883 				break;
884 
885 			/* Try to autoload device module */
886 			(void) module_autoload(name, MODULE_CLASS_DRIVER);
887 		} while (gen != module_gen);
888 		break;
889 
890 	case VBLK:
891 		do {
892 			const struct bdevsw *bdev;
893 
894 			gen = module_gen;
895 			error = bdev_open(dev, ap->a_mode, S_IFBLK, l);
896 			if (error != ENXIO)
897 				break;
898 
899 			/* Check if we already have a valid driver */
900 			mutex_enter(&device_lock);
901 			bdev = bdevsw_lookup(dev);
902 			mutex_exit(&device_lock);
903 			if (bdev != NULL)
904 				break;
905 
906 			/* Get device name from devsw_conv array */
907 			if ((name = bdevsw_getname(major(dev))) == NULL)
908 				break;
909 
910                         /* Try to autoload device module */
911 			(void) module_autoload(name, MODULE_CLASS_DRIVER);
912 		} while (gen != module_gen);
913 		break;
914 
915 	default:
916 		__unreachable();
917 	}
918 
919 	/*
920 	 * Release the I/O reference now that we have called .d_open,
921 	 * and reacquire the vnode lock.  At this point, the device may
922 	 * have been revoked, so we must tread carefully.  However, sn
923 	 * and sd remain valid pointers until we drop our reference.
924 	 */
925 	spec_io_exit(vp, sn);
926 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
927 	KASSERT(vp->v_specnode == sn);
928 
929 	/*
930 	 * If it has been revoked since we released the vnode lock and
931 	 * reacquired it, then spec_node_revoke has closed it, and we
932 	 * must fail with EBADF.
933 	 *
934 	 * Otherwise, if opening it failed, back out and release the
935 	 * open reference.  If it was ever successfully opened and we
936 	 * got the last reference this way, it's now our job to close
937 	 * it.  This might happen in the following scenario:
938 	 *
939 	 *	Thread 1		Thread 2
940 	 *	VOP_OPEN
941 	 *	  ...
942 	 *	  .d_open -> 0 (success)
943 	 *	  acquire vnode lock
944 	 *	  do stuff		VOP_OPEN
945 	 *	  release vnode lock	...
946 	 *				  .d_open -> EBUSY
947 	 *	VOP_CLOSE
948 	 *	  acquire vnode lock
949 	 *	  --sd_opencnt != 0
950 	 *	  => no .d_close
951 	 *	  release vnode lock
952 	 *				  acquire vnode lock
953 	 *				  --sd_opencnt == 0
954 	 *
955 	 * We can't resolve this by making spec_close wait for .d_open
956 	 * to complete before examining sd_opencnt, because .d_open can
957 	 * hang indefinitely, e.g. for a tty.
958 	 */
959 	mutex_enter(&device_lock);
960 	if (sn->sn_gone) {
961 		if (error == 0)
962 			error = EBADF;
963 	} else if (error == 0) {
964 		/*
965 		 * Device has not been revoked, so our opencnt can't
966 		 * have gone away at this point -- transition to
967 		 * sn_gone=true happens before transition to
968 		 * sn_opencnt=0 in spec_node_revoke.
969 		 */
970 		KASSERT(sd->sd_opencnt);
971 		KASSERT(sn->sn_opencnt);
972 		KASSERTMSG(sn->sn_opencnt <= sd->sd_opencnt,
973 		    "sn_opencnt=%u > sd_opencnt=%u",
974 		    sn->sn_opencnt, sd->sd_opencnt);
975 		KASSERT(!sd->sd_closing);
976 		sd->sd_opened = true;
977 	} else if (sd->sd_opencnt == 1 && sd->sd_opened) {
978 		/*
979 		 * We're the last reference to a _previous_ open even
980 		 * though this one failed, so we have to close it.
981 		 * Don't decrement the reference count here --
982 		 * spec_close will do that.
983 		 */
984 		KASSERT(sn->sn_opencnt == 1);
985 		needclose = true;
986 	} else {
987 		KASSERT(sd->sd_opencnt);
988 		KASSERT(sn->sn_opencnt);
989 		KASSERTMSG(sn->sn_opencnt <= sd->sd_opencnt,
990 		    "sn_opencnt=%u > sd_opencnt=%u",
991 		    sn->sn_opencnt, sd->sd_opencnt);
992 		sd->sd_opencnt--;
993 		sn->sn_opencnt--;
994 		if (vp->v_type == VBLK)
995 			sd->sd_bdevvp = NULL;
996 	}
997 	mutex_exit(&device_lock);
998 
999 	/*
1000 	 * If this open failed, but the device was previously opened,
1001 	 * and another thread concurrently closed the vnode while we
1002 	 * were in the middle of reopening it, the other thread will
1003 	 * see sd_opencnt > 0 and thus decide not to call .d_close --
1004 	 * it is now our responsibility to do so.
1005 	 *
1006 	 * XXX The flags passed to VOP_CLOSE here are wrong, but
1007 	 * drivers can't rely on FREAD|FWRITE anyway -- e.g., consider
1008 	 * a device opened by thread 0 with O_READ, then opened by
1009 	 * thread 1 with O_WRITE, then closed by thread 0, and finally
1010 	 * closed by thread 1; the last .d_close call will have FWRITE
1011 	 * but not FREAD.  We should just eliminate the FREAD/FWRITE
1012 	 * parameter to .d_close altogether.
1013 	 */
1014 	if (needclose) {
1015 		KASSERT(error);
1016 		VOP_CLOSE(vp, FNONBLOCK, NOCRED);
1017 	}
1018 
1019 	/* If anything went wrong, we're done.  */
1020 	if (error)
1021 		return error;
1022 
1023 	/*
1024 	 * For disk devices, automagically set the vnode size to the
1025 	 * partition size, if we can.  This applies to block devices
1026 	 * and character devices alike -- every block device must have
1027 	 * a corresponding character device.  And if the module is
1028 	 * loaded it will remain loaded until we're done here (it is
1029 	 * forbidden to devsw_detach until closed).  So it is safe to
1030 	 * query cdev_type unconditionally here.
1031 	 */
1032 	if (cdev_type(dev) == D_DISK) {
1033 		ioctl = vp->v_type == VCHR ? cdev_ioctl : bdev_ioctl;
1034 		if ((*ioctl)(dev, DIOCGPARTINFO, &pi, FREAD, curlwp) == 0)
1035 			uvm_vnp_setsize(vp,
1036 			    (voff_t)pi.pi_secsize * pi.pi_size);
1037 	}
1038 
1039 	/* Success!  */
1040 	return 0;
1041 }
1042 
1043 /*
1044  * Vnode op for read
1045  */
1046 /* ARGSUSED */
1047 int
1048 spec_read(void *v)
1049 {
1050 	struct vop_read_args /* {
1051 		struct vnode *a_vp;
1052 		struct uio *a_uio;
1053 		int  a_ioflag;
1054 		kauth_cred_t a_cred;
1055 	} */ *ap = v;
1056 	struct vnode *vp = ap->a_vp;
1057 	struct uio *uio = ap->a_uio;
1058  	struct lwp *l = curlwp;
1059 	struct specnode *sn;
1060 	dev_t dev;
1061 	struct buf *bp;
1062 	daddr_t bn;
1063 	int bsize, bscale;
1064 	struct partinfo pi;
1065 	int n, on;
1066 	int error = 0;
1067 	int i, nra;
1068 	daddr_t lastbn, *rablks;
1069 	int *rasizes;
1070 	int nrablks, ratogo;
1071 
1072 	KASSERT(uio->uio_rw == UIO_READ);
1073 	KASSERTMSG(VMSPACE_IS_KERNEL_P(uio->uio_vmspace) ||
1074 		   uio->uio_vmspace == curproc->p_vmspace,
1075 		"vmspace belongs to neither kernel nor curproc");
1076 
1077 	if (uio->uio_resid == 0)
1078 		return (0);
1079 
1080 	switch (vp->v_type) {
1081 
1082 	case VCHR:
1083 		/*
1084 		 * Release the lock while we sleep -- possibly
1085 		 * indefinitely, if this is, e.g., a tty -- in
1086 		 * cdev_read, so we don't hold up everything else that
1087 		 * might want access to the vnode.
1088 		 *
1089 		 * But before we issue the read, take an I/O reference
1090 		 * to the specnode so close will know when we're done
1091 		 * reading.  Note that the moment we release the lock,
1092 		 * the vnode's identity may change; hence spec_io_enter
1093 		 * may fail, and the caller may have a dead vnode on
1094 		 * their hands, if the file system on which vp lived
1095 		 * has been unmounted.
1096 		 */
1097 		VOP_UNLOCK(vp);
1098 		error = spec_io_enter(vp, &sn, &dev);
1099 		if (error)
1100 			goto out;
1101 		error = cdev_read(dev, uio, ap->a_ioflag);
1102 		spec_io_exit(vp, sn);
1103 out:		/* XXX What if the caller held an exclusive lock?  */
1104 		vn_lock(vp, LK_SHARED | LK_RETRY);
1105 		return (error);
1106 
1107 	case VBLK:
1108 		KASSERT(vp == vp->v_specnode->sn_dev->sd_bdevvp);
1109 		if (uio->uio_offset < 0)
1110 			return (EINVAL);
1111 
1112 		if (bdev_ioctl(vp->v_rdev, DIOCGPARTINFO, &pi, FREAD, l) == 0)
1113 			bsize = imin(imax(pi.pi_bsize, DEV_BSIZE), MAXBSIZE);
1114 		else
1115 			bsize = BLKDEV_IOSIZE;
1116 
1117 		bscale = bsize >> DEV_BSHIFT;
1118 
1119 		nra = uimax(16 * MAXPHYS / bsize - 1, 511);
1120 		rablks = kmem_alloc(nra * sizeof(*rablks), KM_SLEEP);
1121 		rasizes = kmem_alloc(nra * sizeof(*rasizes), KM_SLEEP);
1122 		lastbn = ((uio->uio_offset + uio->uio_resid - 1) >> DEV_BSHIFT)
1123 		    &~ (bscale - 1);
1124 		nrablks = ratogo = 0;
1125 		do {
1126 			bn = (uio->uio_offset >> DEV_BSHIFT) &~ (bscale - 1);
1127 			on = uio->uio_offset % bsize;
1128 			n = uimin((unsigned)(bsize - on), uio->uio_resid);
1129 
1130 			if (ratogo == 0) {
1131 				nrablks = uimin((lastbn - bn) / bscale, nra);
1132 				ratogo = nrablks;
1133 
1134 				for (i = 0; i < nrablks; ++i) {
1135 					rablks[i] = bn + (i+1) * bscale;
1136 					rasizes[i] = bsize;
1137 				}
1138 
1139 				error = breadn(vp, bn, bsize,
1140 					       rablks, rasizes, nrablks,
1141 					       0, &bp);
1142 			} else {
1143 				if (ratogo > 0)
1144 					--ratogo;
1145 				error = bread(vp, bn, bsize, 0, &bp);
1146 			}
1147 			if (error)
1148 				break;
1149 			n = uimin(n, bsize - bp->b_resid);
1150 			error = uiomove((char *)bp->b_data + on, n, uio);
1151 			brelse(bp, 0);
1152 		} while (error == 0 && uio->uio_resid > 0 && n != 0);
1153 
1154 		kmem_free(rablks, nra * sizeof(*rablks));
1155 		kmem_free(rasizes, nra * sizeof(*rasizes));
1156 
1157 		return (error);
1158 
1159 	default:
1160 		panic("spec_read type");
1161 	}
1162 	/* NOTREACHED */
1163 }
1164 
1165 /*
1166  * Vnode op for write
1167  */
1168 /* ARGSUSED */
1169 int
1170 spec_write(void *v)
1171 {
1172 	struct vop_write_args /* {
1173 		struct vnode *a_vp;
1174 		struct uio *a_uio;
1175 		int  a_ioflag;
1176 		kauth_cred_t a_cred;
1177 	} */ *ap = v;
1178 	struct vnode *vp = ap->a_vp;
1179 	struct uio *uio = ap->a_uio;
1180 	struct lwp *l = curlwp;
1181 	struct specnode *sn;
1182 	dev_t dev;
1183 	struct buf *bp;
1184 	daddr_t bn;
1185 	int bsize, bscale;
1186 	struct partinfo pi;
1187 	int n, on;
1188 	int error = 0;
1189 
1190 	KASSERT(uio->uio_rw == UIO_WRITE);
1191 	KASSERTMSG(VMSPACE_IS_KERNEL_P(uio->uio_vmspace) ||
1192 		   uio->uio_vmspace == curproc->p_vmspace,
1193 		"vmspace belongs to neither kernel nor curproc");
1194 
1195 	switch (vp->v_type) {
1196 
1197 	case VCHR:
1198 		/*
1199 		 * Release the lock while we sleep -- possibly
1200 		 * indefinitely, if this is, e.g., a tty -- in
1201 		 * cdev_write, so we don't hold up everything else that
1202 		 * might want access to the vnode.
1203 		 *
1204 		 * But before we issue the write, take an I/O reference
1205 		 * to the specnode so close will know when we're done
1206 		 * writing.  Note that the moment we release the lock,
1207 		 * the vnode's identity may change; hence spec_io_enter
1208 		 * may fail, and the caller may have a dead vnode on
1209 		 * their hands, if the file system on which vp lived
1210 		 * has been unmounted.
1211 		 */
1212 		VOP_UNLOCK(vp);
1213 		error = spec_io_enter(vp, &sn, &dev);
1214 		if (error)
1215 			goto out;
1216 		error = cdev_write(dev, uio, ap->a_ioflag);
1217 		spec_io_exit(vp, sn);
1218 out:		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1219 		return (error);
1220 
1221 	case VBLK:
1222 		KASSERT(vp == vp->v_specnode->sn_dev->sd_bdevvp);
1223 		if (uio->uio_resid == 0)
1224 			return (0);
1225 		if (uio->uio_offset < 0)
1226 			return (EINVAL);
1227 
1228 		if (bdev_ioctl(vp->v_rdev, DIOCGPARTINFO, &pi, FREAD, l) == 0)
1229 			bsize = imin(imax(pi.pi_bsize, DEV_BSIZE), MAXBSIZE);
1230 		else
1231 			bsize = BLKDEV_IOSIZE;
1232 
1233 		bscale = bsize >> DEV_BSHIFT;
1234 		do {
1235 			bn = (uio->uio_offset >> DEV_BSHIFT) &~ (bscale - 1);
1236 			on = uio->uio_offset % bsize;
1237 			n = uimin((unsigned)(bsize - on), uio->uio_resid);
1238 			if (n == bsize)
1239 				bp = getblk(vp, bn, bsize, 0, 0);
1240 			else
1241 				error = bread(vp, bn, bsize, B_MODIFY, &bp);
1242 			if (error) {
1243 				return (error);
1244 			}
1245 			n = uimin(n, bsize - bp->b_resid);
1246 			error = uiomove((char *)bp->b_data + on, n, uio);
1247 			if (error)
1248 				brelse(bp, 0);
1249 			else {
1250 				if (n + on == bsize)
1251 					bawrite(bp);
1252 				else
1253 					bdwrite(bp);
1254 				error = bp->b_error;
1255 			}
1256 		} while (error == 0 && uio->uio_resid > 0 && n != 0);
1257 		return (error);
1258 
1259 	default:
1260 		panic("spec_write type");
1261 	}
1262 	/* NOTREACHED */
1263 }
1264 
1265 /*
1266  * fdiscard, which on disk devices becomes TRIM.
1267  */
1268 int
1269 spec_fdiscard(void *v)
1270 {
1271 	struct vop_fdiscard_args /* {
1272 		struct vnode *a_vp;
1273 		off_t a_pos;
1274 		off_t a_len;
1275 	} */ *ap = v;
1276 	struct vnode *vp = ap->a_vp;
1277 	dev_t dev;
1278 
1279 	KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
1280 
1281 	dev = vp->v_rdev;
1282 
1283 	switch (vp->v_type) {
1284 	    case VCHR:
1285 		// this is not stored for character devices
1286 		//KASSERT(vp == vp->v_specnode->sn_dev->sd_cdevvp);
1287 		return cdev_discard(dev, ap->a_pos, ap->a_len);
1288 	    case VBLK:
1289 		KASSERT(vp == vp->v_specnode->sn_dev->sd_bdevvp);
1290 		return bdev_discard(dev, ap->a_pos, ap->a_len);
1291 	    default:
1292 		panic("spec_fdiscard: not a device\n");
1293 	}
1294 }
1295 
1296 /*
1297  * Device ioctl operation.
1298  */
1299 /* ARGSUSED */
1300 int
1301 spec_ioctl(void *v)
1302 {
1303 	struct vop_ioctl_args /* {
1304 		struct vnode *a_vp;
1305 		u_long a_command;
1306 		void  *a_data;
1307 		int  a_fflag;
1308 		kauth_cred_t a_cred;
1309 	} */ *ap = v;
1310 	struct vnode *vp = ap->a_vp;
1311 	struct specnode *sn;
1312 	dev_t dev;
1313 	int error;
1314 
1315 	error = spec_io_enter(vp, &sn, &dev);
1316 	if (error)
1317 		return error;
1318 
1319 	switch (vp->v_type) {
1320 	case VCHR:
1321 		error = cdev_ioctl(dev, ap->a_command, ap->a_data,
1322 		    ap->a_fflag, curlwp);
1323 		break;
1324 	case VBLK:
1325 		KASSERT(vp == vp->v_specnode->sn_dev->sd_bdevvp);
1326 		error = bdev_ioctl(dev, ap->a_command, ap->a_data,
1327 		   ap->a_fflag, curlwp);
1328 		break;
1329 	default:
1330 		panic("spec_ioctl");
1331 		/* NOTREACHED */
1332 	}
1333 
1334 	spec_io_exit(vp, sn);
1335 	return error;
1336 }
1337 
1338 /* ARGSUSED */
1339 int
1340 spec_poll(void *v)
1341 {
1342 	struct vop_poll_args /* {
1343 		struct vnode *a_vp;
1344 		int a_events;
1345 	} */ *ap = v;
1346 	struct vnode *vp = ap->a_vp;
1347 	struct specnode *sn;
1348 	dev_t dev;
1349 	int revents;
1350 
1351 	if (spec_io_enter(vp, &sn, &dev) != 0)
1352 		return POLLERR;
1353 
1354 	switch (vp->v_type) {
1355 	case VCHR:
1356 		revents = cdev_poll(dev, ap->a_events, curlwp);
1357 		break;
1358 	default:
1359 		revents = genfs_poll(v);
1360 		break;
1361 	}
1362 
1363 	spec_io_exit(vp, sn);
1364 	return revents;
1365 }
1366 
1367 /* ARGSUSED */
1368 int
1369 spec_kqfilter(void *v)
1370 {
1371 	struct vop_kqfilter_args /* {
1372 		struct vnode	*a_vp;
1373 		struct proc	*a_kn;
1374 	} */ *ap = v;
1375 	struct vnode *vp = ap->a_vp;
1376 	struct specnode *sn;
1377 	dev_t dev;
1378 	int error;
1379 
1380 	error = spec_io_enter(vp, &sn, &dev);
1381 	if (error)
1382 		return error;
1383 
1384 	switch (vp->v_type) {
1385 	case VCHR:
1386 		error = cdev_kqfilter(dev, ap->a_kn);
1387 		break;
1388 	default:
1389 		/*
1390 		 * Block devices don't support kqfilter, and refuse it
1391 		 * for any other files (like those vflush()ed) too.
1392 		 */
1393 		error = EOPNOTSUPP;
1394 		break;
1395 	}
1396 
1397 	spec_io_exit(vp, sn);
1398 	return error;
1399 }
1400 
1401 /*
1402  * Allow mapping of only D_DISK.  This is called only for VBLK.
1403  */
1404 int
1405 spec_mmap(void *v)
1406 {
1407 	struct vop_mmap_args /* {
1408 		struct vnode *a_vp;
1409 		vm_prot_t a_prot;
1410 		kauth_cred_t a_cred;
1411 	} */ *ap = v;
1412 	struct vnode *vp = ap->a_vp;
1413 	struct specnode *sn;
1414 	dev_t dev;
1415 	int error;
1416 
1417 	KASSERT(vp->v_type == VBLK);
1418 
1419 	error = spec_io_enter(vp, &sn, &dev);
1420 	if (error)
1421 		return error;
1422 
1423 	error = bdev_type(dev) == D_DISK ? 0 : EINVAL;
1424 
1425 	spec_io_exit(vp, sn);
1426 	return 0;
1427 }
1428 
1429 /*
1430  * Synch buffers associated with a block device
1431  */
1432 /* ARGSUSED */
1433 int
1434 spec_fsync(void *v)
1435 {
1436 	struct vop_fsync_args /* {
1437 		struct vnode *a_vp;
1438 		kauth_cred_t a_cred;
1439 		int  a_flags;
1440 		off_t offlo;
1441 		off_t offhi;
1442 	} */ *ap = v;
1443 	struct vnode *vp = ap->a_vp;
1444 	struct mount *mp;
1445 	int error;
1446 
1447 	if (vp->v_type == VBLK) {
1448 		if ((mp = spec_node_getmountedfs(vp)) != NULL) {
1449 			error = VFS_FSYNC(mp, vp, ap->a_flags);
1450 			if (error != EOPNOTSUPP)
1451 				return error;
1452 		}
1453 		return vflushbuf(vp, ap->a_flags);
1454 	}
1455 	return (0);
1456 }
1457 
1458 /*
1459  * Just call the device strategy routine
1460  */
1461 int
1462 spec_strategy(void *v)
1463 {
1464 	struct vop_strategy_args /* {
1465 		struct vnode *a_vp;
1466 		struct buf *a_bp;
1467 	} */ *ap = v;
1468 	struct vnode *vp = ap->a_vp;
1469 	struct buf *bp = ap->a_bp;
1470 	struct specnode *sn = NULL;
1471 	dev_t dev;
1472 	int error;
1473 
1474 	error = spec_io_enter(vp, &sn, &dev);
1475 	if (error)
1476 		goto out;
1477 
1478 	bp->b_dev = dev;
1479 
1480 	if (!(bp->b_flags & B_READ)) {
1481 #ifdef DIAGNOSTIC
1482 		if (bp->b_vp && bp->b_vp->v_type == VBLK) {
1483 			struct mount *mp = spec_node_getmountedfs(bp->b_vp);
1484 
1485 			if (mp && (mp->mnt_flag & MNT_RDONLY)) {
1486 				printf("%s blk %"PRId64" written while ro!\n",
1487 				    mp->mnt_stat.f_mntonname, bp->b_blkno);
1488 			}
1489 		}
1490 #endif /* DIAGNOSTIC */
1491 		error = fscow_run(bp, false);
1492 		if (error)
1493 			goto out;
1494 	}
1495 	bdev_strategy(bp);
1496 
1497 	error = 0;
1498 
1499 out:	if (sn)
1500 		spec_io_exit(vp, sn);
1501 	if (error) {
1502 		bp->b_error = error;
1503 		bp->b_resid = bp->b_bcount;
1504 		biodone(bp);
1505 	}
1506 	return error;
1507 }
1508 
1509 int
1510 spec_inactive(void *v)
1511 {
1512 	struct vop_inactive_v2_args /* {
1513 		struct vnode *a_vp;
1514 		struct bool *a_recycle;
1515 	} */ *ap = v;
1516 
1517 	KASSERT(ap->a_vp->v_mount == dead_rootmount);
1518 	*ap->a_recycle = true;
1519 
1520 	return 0;
1521 }
1522 
1523 int
1524 spec_reclaim(void *v)
1525 {
1526 	struct vop_reclaim_v2_args /* {
1527 		struct vnode *a_vp;
1528 	} */ *ap = v;
1529 	struct vnode *vp = ap->a_vp;
1530 
1531 	KASSERT(vp->v_specnode->sn_opencnt == 0);
1532 
1533 	VOP_UNLOCK(vp);
1534 
1535 	KASSERT(vp->v_mount == dead_rootmount);
1536 	return 0;
1537 }
1538 
1539 /*
1540  * This is a noop, simply returning what one has been given.
1541  */
1542 int
1543 spec_bmap(void *v)
1544 {
1545 	struct vop_bmap_args /* {
1546 		struct vnode *a_vp;
1547 		daddr_t  a_bn;
1548 		struct vnode **a_vpp;
1549 		daddr_t *a_bnp;
1550 		int *a_runp;
1551 	} */ *ap = v;
1552 
1553 	if (ap->a_vpp != NULL)
1554 		*ap->a_vpp = ap->a_vp;
1555 	if (ap->a_bnp != NULL)
1556 		*ap->a_bnp = ap->a_bn;
1557 	if (ap->a_runp != NULL)
1558 		*ap->a_runp = (MAXBSIZE >> DEV_BSHIFT) - 1;
1559 	return (0);
1560 }
1561 
1562 /*
1563  * Device close routine
1564  */
1565 /* ARGSUSED */
1566 int
1567 spec_close(void *v)
1568 {
1569 	struct vop_close_args /* {
1570 		struct vnode *a_vp;
1571 		int  a_fflag;
1572 		kauth_cred_t a_cred;
1573 	} */ *ap = v;
1574 	struct vnode *vp = ap->a_vp;
1575 	struct session *sess;
1576 	dev_t dev;
1577 	int flags = ap->a_fflag;
1578 	int mode, error, count;
1579 	specnode_t *sn;
1580 	specdev_t *sd;
1581 
1582 	KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
1583 
1584 	mutex_enter(vp->v_interlock);
1585 	sn = vp->v_specnode;
1586 	dev = vp->v_rdev;
1587 	sd = sn->sn_dev;
1588 	/*
1589 	 * If we're going away soon, make this non-blocking.
1590 	 * Also ensures that we won't wedge in vn_lock below.
1591 	 */
1592 	if (vdead_check(vp, VDEAD_NOWAIT) != 0)
1593 		flags |= FNONBLOCK;
1594 	mutex_exit(vp->v_interlock);
1595 
1596 	switch (vp->v_type) {
1597 
1598 	case VCHR:
1599 		/*
1600 		 * Hack: a tty device that is a controlling terminal
1601 		 * has a reference from the session structure.  We
1602 		 * cannot easily tell that a character device is a
1603 		 * controlling terminal, unless it is the closing
1604 		 * process' controlling terminal.  In that case, if the
1605 		 * open count is 1 release the reference from the
1606 		 * session.  Also, remove the link from the tty back to
1607 		 * the session and pgrp.
1608 		 *
1609 		 * XXX V. fishy.
1610 		 */
1611 		mutex_enter(&proc_lock);
1612 		sess = curlwp->l_proc->p_session;
1613 		if (sn->sn_opencnt == 1 && vp == sess->s_ttyvp) {
1614 			mutex_spin_enter(&tty_lock);
1615 			sess->s_ttyvp = NULL;
1616 			if (sess->s_ttyp->t_session != NULL) {
1617 				sess->s_ttyp->t_pgrp = NULL;
1618 				sess->s_ttyp->t_session = NULL;
1619 				mutex_spin_exit(&tty_lock);
1620 				/* Releases proc_lock. */
1621 				proc_sessrele(sess);
1622 			} else {
1623 				mutex_spin_exit(&tty_lock);
1624 				if (sess->s_ttyp->t_pgrp != NULL)
1625 					panic("spec_close: spurious pgrp ref");
1626 				mutex_exit(&proc_lock);
1627 			}
1628 			vrele(vp);
1629 		} else
1630 			mutex_exit(&proc_lock);
1631 
1632 		/*
1633 		 * If the vnode is locked, then we are in the midst
1634 		 * of forcably closing the device, otherwise we only
1635 		 * close on last reference.
1636 		 */
1637 		mode = S_IFCHR;
1638 		break;
1639 
1640 	case VBLK:
1641 		KASSERT(vp == vp->v_specnode->sn_dev->sd_bdevvp);
1642 		/*
1643 		 * On last close of a block device (that isn't mounted)
1644 		 * we must invalidate any in core blocks, so that
1645 		 * we can, for instance, change floppy disks.
1646 		 */
1647 		error = vinvalbuf(vp, V_SAVE, ap->a_cred, curlwp, 0, 0);
1648 		if (error)
1649 			return (error);
1650 		/*
1651 		 * We do not want to really close the device if it
1652 		 * is still in use unless we are trying to close it
1653 		 * forcibly. Since every use (buffer, vnode, swap, cmap)
1654 		 * holds a reference to the vnode, and because we mark
1655 		 * any other vnodes that alias this device, when the
1656 		 * sum of the reference counts on all the aliased
1657 		 * vnodes descends to one, we are on last close.
1658 		 */
1659 		mode = S_IFBLK;
1660 		break;
1661 
1662 	default:
1663 		panic("spec_close: not special");
1664 	}
1665 
1666 	/*
1667 	 * Decrement the open reference count of this node and the
1668 	 * device.  For block devices, the open reference count must be
1669 	 * 1 at this point.  If the device's open reference count goes
1670 	 * to zero, we're the last one out so get the lights.
1671 	 *
1672 	 * We may find --sd->sd_opencnt gives zero, and yet
1673 	 * sd->sd_opened is false.  This happens if the vnode is
1674 	 * revoked at the same time as it is being opened, which can
1675 	 * happen when opening a tty blocks indefinitely.  In that
1676 	 * case, we still must call close -- it is the job of close to
1677 	 * interrupt the open.  Either way, the device will be no
1678 	 * longer opened, so we have to clear sd->sd_opened; subsequent
1679 	 * opens will have responsibility for issuing close.
1680 	 *
1681 	 * This has the side effect that the sequence of opens might
1682 	 * happen out of order -- we might end up doing open, open,
1683 	 * close, close, instead of open, close, open, close.  This is
1684 	 * unavoidable with the current devsw API, where open is
1685 	 * allowed to block and close must be able to run concurrently
1686 	 * to interrupt it.  It is the driver's responsibility to
1687 	 * ensure that close is idempotent so that this works.  Drivers
1688 	 * requiring per-open state and exact 1:1 correspondence
1689 	 * between open and close can use fd_clone.
1690 	 */
1691 	mutex_enter(&device_lock);
1692 	KASSERT(sn->sn_opencnt);
1693 	KASSERT(sd->sd_opencnt);
1694 	KASSERTMSG(sn->sn_opencnt <= sd->sd_opencnt,
1695 	    "sn_opencnt=%u > sd_opencnt=%u",
1696 	    sn->sn_opencnt, sd->sd_opencnt);
1697 	sn->sn_opencnt--;
1698 	count = --sd->sd_opencnt;
1699 	if (vp->v_type == VBLK) {
1700 		KASSERTMSG(count == 0, "block device with %u opens",
1701 		    count + 1);
1702 		sd->sd_bdevvp = NULL;
1703 	}
1704 	if (count == 0) {
1705 		KASSERTMSG(sn->sn_opencnt == 0, "sn_opencnt=%u",
1706 		    sn->sn_opencnt);
1707 		KASSERT(!sd->sd_closing);
1708 		sd->sd_opened = false;
1709 		sd->sd_closing = true;
1710 	}
1711 	mutex_exit(&device_lock);
1712 
1713 	if (count != 0)
1714 		return 0;
1715 
1716 	/*
1717 	 * If we're able to block, release the vnode lock & reacquire. We
1718 	 * might end up sleeping for someone else who wants our queues. They
1719 	 * won't get them if we hold the vnode locked.
1720 	 */
1721 	if (!(flags & FNONBLOCK))
1722 		VOP_UNLOCK(vp);
1723 
1724 	/*
1725 	 * If we can cancel all outstanding I/O, then wait for it to
1726 	 * drain before we call .d_close.  Drivers that split up
1727 	 * .d_cancel and .d_close this way need not have any internal
1728 	 * mechanism for waiting in .d_close for I/O to drain.
1729 	 */
1730 	if (vp->v_type == VBLK)
1731 		error = bdev_cancel(dev, flags, mode, curlwp);
1732 	else
1733 		error = cdev_cancel(dev, flags, mode, curlwp);
1734 	if (error == 0)
1735 		spec_io_drain(sd);
1736 	else
1737 		KASSERTMSG(error == ENODEV, "cancel dev=0x%lx failed with %d",
1738 		    (unsigned long)dev, error);
1739 
1740 	if (vp->v_type == VBLK)
1741 		error = bdev_close(dev, flags, mode, curlwp);
1742 	else
1743 		error = cdev_close(dev, flags, mode, curlwp);
1744 
1745 	/*
1746 	 * Wait for all other devsw operations to drain.  After this
1747 	 * point, no bdev/cdev_* can be active for this specdev.
1748 	 */
1749 	spec_io_drain(sd);
1750 
1751 	/*
1752 	 * Wake any spec_open calls waiting for close to finish -- do
1753 	 * this before reacquiring the vnode lock, because spec_open
1754 	 * holds the vnode lock while waiting, so doing this after
1755 	 * reacquiring the lock would deadlock.
1756 	 */
1757 	mutex_enter(&device_lock);
1758 	KASSERT(!sd->sd_opened);
1759 	KASSERT(sd->sd_closing);
1760 	sd->sd_closing = false;
1761 	cv_broadcast(&specfs_iocv);
1762 	mutex_exit(&device_lock);
1763 
1764 	if (!(flags & FNONBLOCK))
1765 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1766 
1767 	return (error);
1768 }
1769 
1770 /*
1771  * Print out the contents of a special device vnode.
1772  */
1773 int
1774 spec_print(void *v)
1775 {
1776 	struct vop_print_args /* {
1777 		struct vnode *a_vp;
1778 	} */ *ap = v;
1779 
1780 	printf("dev %llu, %llu\n", (unsigned long long)major(ap->a_vp->v_rdev),
1781 	    (unsigned long long)minor(ap->a_vp->v_rdev));
1782 	return 0;
1783 }
1784 
1785 /*
1786  * Return POSIX pathconf information applicable to special devices.
1787  */
1788 int
1789 spec_pathconf(void *v)
1790 {
1791 	struct vop_pathconf_args /* {
1792 		struct vnode *a_vp;
1793 		int a_name;
1794 		register_t *a_retval;
1795 	} */ *ap = v;
1796 
1797 	switch (ap->a_name) {
1798 	case _PC_LINK_MAX:
1799 		*ap->a_retval = LINK_MAX;
1800 		return (0);
1801 	case _PC_MAX_CANON:
1802 		*ap->a_retval = MAX_CANON;
1803 		return (0);
1804 	case _PC_MAX_INPUT:
1805 		*ap->a_retval = MAX_INPUT;
1806 		return (0);
1807 	case _PC_PIPE_BUF:
1808 		*ap->a_retval = PIPE_BUF;
1809 		return (0);
1810 	case _PC_CHOWN_RESTRICTED:
1811 		*ap->a_retval = 1;
1812 		return (0);
1813 	case _PC_VDISABLE:
1814 		*ap->a_retval = _POSIX_VDISABLE;
1815 		return (0);
1816 	case _PC_SYNC_IO:
1817 		*ap->a_retval = 1;
1818 		return (0);
1819 	default:
1820 		return genfs_pathconf(ap);
1821 	}
1822 	/* NOTREACHED */
1823 }
1824 
1825 /*
1826  * Advisory record locking support.
1827  */
1828 int
1829 spec_advlock(void *v)
1830 {
1831 	struct vop_advlock_args /* {
1832 		struct vnode *a_vp;
1833 		void *a_id;
1834 		int a_op;
1835 		struct flock *a_fl;
1836 		int a_flags;
1837 	} */ *ap = v;
1838 	struct vnode *vp = ap->a_vp;
1839 
1840 	return lf_advlock(ap, &vp->v_speclockf, (off_t)0);
1841 }
1842