xref: /netbsd-src/sys/kern/vfs_trans.c (revision deb6f0161a9109e7de9b519dc8dfb9478668dcdd)
1 /*	$NetBSD: vfs_trans.c,v 1.51 2018/10/05 09:51:55 hannken Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Juergen Hannken-Illjes.
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 NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __KERNEL_RCSID(0, "$NetBSD: vfs_trans.c,v 1.51 2018/10/05 09:51:55 hannken Exp $");
34 
35 /*
36  * File system transaction operations.
37  */
38 
39 #ifdef _KERNEL_OPT
40 #include "opt_ddb.h"
41 #endif
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/atomic.h>
46 #include <sys/buf.h>
47 #include <sys/kmem.h>
48 #include <sys/mount.h>
49 #include <sys/pserialize.h>
50 #include <sys/vnode.h>
51 #include <sys/fstrans.h>
52 #include <sys/proc.h>
53 
54 #include <miscfs/specfs/specdev.h>
55 
56 enum fstrans_lock_type {
57 	FSTRANS_LAZY,			/* Granted while not suspended */
58 	FSTRANS_SHARED,			/* Granted while not suspending */
59 	FSTRANS_EXCL			/* Internal: exclusive lock */
60 };
61 
62 struct fscow_handler {
63 	LIST_ENTRY(fscow_handler) ch_list;
64 	int (*ch_func)(void *, struct buf *, bool);
65 	void *ch_arg;
66 };
67 struct fstrans_lwp_info {
68 	struct fstrans_lwp_info *fli_succ;
69 	struct lwp *fli_self;
70 	struct mount *fli_mount;
71 	int fli_trans_cnt;
72 	int fli_cow_cnt;
73 	enum fstrans_lock_type fli_lock_type;
74 	LIST_ENTRY(fstrans_lwp_info) fli_list;
75 };
76 struct fstrans_mount_info {
77 	enum fstrans_state fmi_state;
78 	unsigned int fmi_ref_cnt;
79 	bool fmi_cow_change;
80 	LIST_HEAD(, fscow_handler) fmi_cow_handler;
81 };
82 
83 static specificdata_key_t lwp_data_key;	/* Our specific data key. */
84 static kmutex_t vfs_suspend_lock;	/* Serialize suspensions. */
85 static kmutex_t fstrans_lock;		/* Fstrans big lock. */
86 static kmutex_t fstrans_mount_lock;	/* Fstrans mount big lock. */
87 static kcondvar_t fstrans_state_cv;	/* Fstrans or cow state changed. */
88 static kcondvar_t fstrans_count_cv;	/* Fstrans or cow count changed. */
89 static pserialize_t fstrans_psz;	/* Pserialize state. */
90 static LIST_HEAD(fstrans_lwp_head, fstrans_lwp_info) fstrans_fli_head;
91 					/* List of all fstrans_lwp_info. */
92 
93 static inline struct mount *fstrans_normalize_mount(struct mount *);
94 static void fstrans_lwp_dtor(void *);
95 static void fstrans_mount_dtor(struct mount *);
96 static void fstrans_clear_lwp_info(void);
97 static inline struct fstrans_lwp_info *
98     fstrans_get_lwp_info(struct mount *, bool);
99 static struct fstrans_lwp_info *fstrans_alloc_lwp_info(struct mount *);
100 static inline int _fstrans_start(struct mount *, enum fstrans_lock_type, int);
101 static bool grant_lock(const enum fstrans_state, const enum fstrans_lock_type);
102 static bool state_change_done(const struct mount *);
103 static bool cow_state_change_done(const struct mount *);
104 static void cow_change_enter(const struct mount *);
105 static void cow_change_done(const struct mount *);
106 
107 /*
108  * Initialize.
109  */
110 void
111 fstrans_init(void)
112 {
113 	int error __diagused;
114 
115 	error = lwp_specific_key_create(&lwp_data_key, fstrans_lwp_dtor);
116 	KASSERT(error == 0);
117 
118 	mutex_init(&vfs_suspend_lock, MUTEX_DEFAULT, IPL_NONE);
119 	mutex_init(&fstrans_lock, MUTEX_DEFAULT, IPL_NONE);
120 	mutex_init(&fstrans_mount_lock, MUTEX_DEFAULT, IPL_NONE);
121 	cv_init(&fstrans_state_cv, "fstchg");
122 	cv_init(&fstrans_count_cv, "fstcnt");
123 	fstrans_psz = pserialize_create();
124 	LIST_INIT(&fstrans_fli_head);
125 }
126 
127 /*
128  * Normalize mount.
129  * Return mount if file system supports fstrans, NULL otherwise.
130  */
131 static inline struct mount *
132 fstrans_normalize_mount(struct mount *mp)
133 {
134 
135 	while (mp && mp->mnt_lower)
136 		mp = mp->mnt_lower;
137 	if (mp == NULL)
138 		return NULL;
139 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
140 		return NULL;
141 	return mp;
142 }
143 
144 /*
145  * Deallocate lwp state.
146  */
147 static void
148 fstrans_lwp_dtor(void *arg)
149 {
150 	struct fstrans_lwp_info *fli, *fli_next;
151 
152 	for (fli = arg; fli; fli = fli_next) {
153 		KASSERT(fli->fli_trans_cnt == 0);
154 		KASSERT(fli->fli_cow_cnt == 0);
155 		if (fli->fli_mount != NULL)
156 			fstrans_mount_dtor(fli->fli_mount);
157 		fli_next = fli->fli_succ;
158 		fli->fli_mount = NULL;
159 		membar_sync();
160 		fli->fli_self = NULL;
161 	}
162 }
163 
164 /*
165  * Dereference mount state.
166  */
167 static void
168 fstrans_mount_dtor(struct mount *mp)
169 {
170 	struct fstrans_mount_info *fmi;
171 
172 	mutex_enter(&fstrans_mount_lock);
173 
174 	fmi = mp->mnt_transinfo;
175 	KASSERT(fmi != NULL);
176 	fmi->fmi_ref_cnt -= 1;
177 	if (fmi->fmi_ref_cnt > 0) {
178 		mutex_exit(&fstrans_mount_lock);
179 		return;
180 	}
181 
182 	KASSERT(fmi->fmi_state == FSTRANS_NORMAL);
183 	KASSERT(LIST_FIRST(&fmi->fmi_cow_handler) == NULL);
184 
185 	mp->mnt_iflag &= ~IMNT_HAS_TRANS;
186 	mp->mnt_transinfo = NULL;
187 
188 	mutex_exit(&fstrans_mount_lock);
189 
190 	kmem_free(fmi, sizeof(*fmi));
191 	vfs_rele(mp);
192 }
193 
194 /*
195  * Allocate mount state.
196  */
197 int
198 fstrans_mount(struct mount *mp)
199 {
200 	struct fstrans_mount_info *newfmi;
201 
202 	newfmi = kmem_alloc(sizeof(*newfmi), KM_SLEEP);
203 	newfmi->fmi_state = FSTRANS_NORMAL;
204 	newfmi->fmi_ref_cnt = 1;
205 	LIST_INIT(&newfmi->fmi_cow_handler);
206 	newfmi->fmi_cow_change = false;
207 
208 	mutex_enter(&fstrans_mount_lock);
209 	mp->mnt_transinfo = newfmi;
210 	mp->mnt_iflag |= IMNT_HAS_TRANS;
211 	mutex_exit(&fstrans_mount_lock);
212 
213 	vfs_ref(mp);
214 
215 	return 0;
216 }
217 
218 /*
219  * Deallocate mount state.
220  */
221 void
222 fstrans_unmount(struct mount *mp)
223 {
224 
225 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
226 		return;
227 
228 	KASSERT(mp->mnt_transinfo != NULL);
229 
230 	fstrans_mount_dtor(mp);
231 }
232 
233 /*
234  * Clear mount entries whose mount is gone.
235  */
236 static void
237 fstrans_clear_lwp_info(void)
238 {
239 	struct fstrans_lwp_info *fli;
240 
241 	/*
242 	 * Scan our list clearing entries whose mount is gone.
243 	 */
244 	for (fli = lwp_getspecific(lwp_data_key); fli; fli = fli->fli_succ) {
245 		if (fli->fli_mount != NULL &&
246 		    (fli->fli_mount->mnt_iflag & IMNT_GONE) != 0 &&
247 		    fli->fli_trans_cnt == 0 && fli->fli_cow_cnt == 0) {
248 			fstrans_mount_dtor(fli->fli_mount);
249 			fli->fli_mount = NULL;
250 		}
251 	}
252 }
253 
254 /*
255  * Allocate and return per lwp info for this mount.
256  */
257 static struct fstrans_lwp_info *
258 fstrans_alloc_lwp_info(struct mount *mp)
259 {
260 	struct fstrans_lwp_info *fli;
261 	struct fstrans_mount_info *fmi;
262 
263 	/*
264 	 * Try to reuse a cleared entry or allocate a new one.
265 	 */
266 	for (fli = lwp_getspecific(lwp_data_key); fli; fli = fli->fli_succ) {
267 		KASSERT(fli->fli_mount != mp);
268 		if (fli->fli_mount == NULL) {
269 			KASSERT(fli->fli_trans_cnt == 0);
270 			KASSERT(fli->fli_cow_cnt == 0);
271 			break;
272 		}
273 	}
274 	if (fli == NULL) {
275 		mutex_enter(&fstrans_lock);
276 		LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
277 			if (fli->fli_self == NULL) {
278 				KASSERT(fli->fli_mount == NULL);
279 				KASSERT(fli->fli_trans_cnt == 0);
280 				KASSERT(fli->fli_cow_cnt == 0);
281 				fli->fli_self = curlwp;
282 				fli->fli_succ = lwp_getspecific(lwp_data_key);
283 				lwp_setspecific(lwp_data_key, fli);
284 				break;
285 			}
286 		}
287 		mutex_exit(&fstrans_lock);
288 	}
289 	if (fli == NULL) {
290 		fli = kmem_alloc(sizeof(*fli), KM_SLEEP);
291 		mutex_enter(&fstrans_lock);
292 		memset(fli, 0, sizeof(*fli));
293 		fli->fli_self = curlwp;
294 		LIST_INSERT_HEAD(&fstrans_fli_head, fli, fli_list);
295 		mutex_exit(&fstrans_lock);
296 		fli->fli_succ = lwp_getspecific(lwp_data_key);
297 		lwp_setspecific(lwp_data_key, fli);
298 	}
299 
300 	/*
301 	 * Attach the entry to the mount if its mnt_transinfo is valid.
302 	 */
303 	mutex_enter(&fstrans_mount_lock);
304 	fmi = mp->mnt_transinfo;
305 	if (__predict_true(fmi != NULL)) {
306 		fli->fli_mount = mp;
307 		fmi->fmi_ref_cnt += 1;
308 	} else {
309 		fli = NULL;
310 	}
311 	mutex_exit(&fstrans_mount_lock);
312 
313 	return fli;
314 }
315 
316 /*
317  * Retrieve the per lwp info for this mount allocating if necessary.
318  */
319 static inline struct fstrans_lwp_info *
320 fstrans_get_lwp_info(struct mount *mp, bool do_alloc)
321 {
322 	struct fstrans_lwp_info *fli;
323 
324 	/*
325 	 * Scan our list for a match.
326 	 */
327 	for (fli = lwp_getspecific(lwp_data_key); fli; fli = fli->fli_succ) {
328 		if (fli->fli_mount == mp)
329 			return fli;
330 	}
331 
332 	return (do_alloc ? fstrans_alloc_lwp_info(mp) : NULL);
333 }
334 
335 /*
336  * Check if this lock type is granted at this state.
337  */
338 static bool
339 grant_lock(const enum fstrans_state state, const enum fstrans_lock_type type)
340 {
341 
342 	if (__predict_true(state == FSTRANS_NORMAL))
343 		return true;
344 	if (type == FSTRANS_EXCL)
345 		return true;
346 	if  (state == FSTRANS_SUSPENDING && type == FSTRANS_LAZY)
347 		return true;
348 
349 	return false;
350 }
351 
352 /*
353  * Start a transaction.  If this thread already has a transaction on this
354  * file system increment the reference counter.
355  */
356 static inline int
357 _fstrans_start(struct mount *mp, enum fstrans_lock_type lock_type, int wait)
358 {
359 	int s;
360 	struct mount *lmp;
361 	struct fstrans_lwp_info *fli;
362 	struct fstrans_mount_info *fmi;
363 
364 	if ((lmp = fstrans_normalize_mount(mp)) == NULL)
365 		return 0;
366 
367 	ASSERT_SLEEPABLE();
368 
369 	/*
370 	 * Allocate per lwp info for layered file systems to
371 	 * get a reference to the mount.  No need to increment
372 	 * the reference counter here.
373 	 */
374 	for (lmp = mp; lmp->mnt_lower; lmp = lmp->mnt_lower) {
375 		fli = fstrans_get_lwp_info(lmp, true);
376 	}
377 
378 	if ((fli = fstrans_get_lwp_info(lmp, true)) == NULL)
379 		return 0;
380 
381 	if (fli->fli_trans_cnt > 0) {
382 		KASSERT(lock_type != FSTRANS_EXCL);
383 		fli->fli_trans_cnt += 1;
384 
385 		return 0;
386 	}
387 
388 	s = pserialize_read_enter();
389 	fmi = lmp->mnt_transinfo;
390 	if (__predict_true(grant_lock(fmi->fmi_state, lock_type))) {
391 		fli->fli_trans_cnt = 1;
392 		fli->fli_lock_type = lock_type;
393 		pserialize_read_exit(s);
394 
395 		return 0;
396 	}
397 	pserialize_read_exit(s);
398 
399 	if (! wait)
400 		return EBUSY;
401 
402 	mutex_enter(&fstrans_lock);
403 	while (! grant_lock(fmi->fmi_state, lock_type))
404 		cv_wait(&fstrans_state_cv, &fstrans_lock);
405 	fli->fli_trans_cnt = 1;
406 	fli->fli_lock_type = lock_type;
407 	mutex_exit(&fstrans_lock);
408 
409 	return 0;
410 }
411 
412 void
413 fstrans_start(struct mount *mp)
414 {
415 	int error __diagused;
416 
417 	error = _fstrans_start(mp, FSTRANS_SHARED, 1);
418 	KASSERT(error == 0);
419 }
420 
421 int
422 fstrans_start_nowait(struct mount *mp)
423 {
424 
425 	return _fstrans_start(mp, FSTRANS_SHARED, 0);
426 }
427 
428 void
429 fstrans_start_lazy(struct mount *mp)
430 {
431 	int error __diagused;
432 
433 	error = _fstrans_start(mp, FSTRANS_LAZY, 1);
434 	KASSERT(error == 0);
435 }
436 
437 /*
438  * Finish a transaction.
439  */
440 void
441 fstrans_done(struct mount *mp)
442 {
443 	int s;
444 	struct fstrans_lwp_info *fli;
445 	struct fstrans_mount_info *fmi;
446 
447 	if ((mp = fstrans_normalize_mount(mp)) == NULL)
448 		return;
449 	if ((fli = fstrans_get_lwp_info(mp, false)) == NULL)
450 		return;
451 	KASSERT(fli->fli_trans_cnt > 0);
452 
453 	if (fli->fli_trans_cnt > 1) {
454 		fli->fli_trans_cnt -= 1;
455 
456 		return;
457 	}
458 
459 	fstrans_clear_lwp_info();
460 
461 	s = pserialize_read_enter();
462 	fmi = mp->mnt_transinfo;
463 	if (__predict_true(fmi->fmi_state == FSTRANS_NORMAL)) {
464 		fli->fli_trans_cnt = 0;
465 		pserialize_read_exit(s);
466 
467 		return;
468 	}
469 	pserialize_read_exit(s);
470 
471 	mutex_enter(&fstrans_lock);
472 	fli->fli_trans_cnt = 0;
473 	cv_signal(&fstrans_count_cv);
474 	mutex_exit(&fstrans_lock);
475 }
476 
477 /*
478  * Check if this thread has an exclusive lock.
479  */
480 int
481 fstrans_is_owner(struct mount *mp)
482 {
483 	struct fstrans_lwp_info *fli;
484 
485 	if ((mp = fstrans_normalize_mount(mp)) == NULL)
486 		return 0;
487 	if ((fli = fstrans_get_lwp_info(mp, false)) == NULL)
488 		return 0;
489 
490 	if (fli->fli_trans_cnt == 0)
491 		return 0;
492 
493 	KASSERT(fli->fli_mount == mp);
494 	KASSERT(fli->fli_trans_cnt > 0);
495 
496 	return (fli->fli_lock_type == FSTRANS_EXCL);
497 }
498 
499 /*
500  * True, if no thread is in a transaction not granted at the current state.
501  */
502 static bool
503 state_change_done(const struct mount *mp)
504 {
505 	struct fstrans_lwp_info *fli;
506 	struct fstrans_mount_info *fmi;
507 
508 	KASSERT(mutex_owned(&fstrans_lock));
509 
510 	fmi = mp->mnt_transinfo;
511 	LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
512 		if (fli->fli_mount != mp)
513 			continue;
514 		if (fli->fli_trans_cnt == 0)
515 			continue;
516 		if (grant_lock(fmi->fmi_state, fli->fli_lock_type))
517 			continue;
518 
519 		return false;
520 	}
521 
522 	return true;
523 }
524 
525 /*
526  * Set new file system state.
527  */
528 int
529 fstrans_setstate(struct mount *mp, enum fstrans_state new_state)
530 {
531 	int error;
532 	enum fstrans_state old_state;
533 	struct fstrans_mount_info *fmi;
534 
535 	fmi = mp->mnt_transinfo;
536 	old_state = fmi->fmi_state;
537 	if (old_state == new_state)
538 		return 0;
539 
540 	mutex_enter(&fstrans_lock);
541 	fmi->fmi_state = new_state;
542 	pserialize_perform(fstrans_psz);
543 
544 	/*
545 	 * All threads see the new state now.
546 	 * Wait for transactions invalid at this state to leave.
547 	 */
548 	error = 0;
549 	while (! state_change_done(mp)) {
550 		error = cv_wait_sig(&fstrans_count_cv, &fstrans_lock);
551 		if (error) {
552 			new_state = fmi->fmi_state = FSTRANS_NORMAL;
553 			break;
554 		}
555 	}
556 	cv_broadcast(&fstrans_state_cv);
557 	mutex_exit(&fstrans_lock);
558 
559 	if (old_state != new_state) {
560 		if (old_state == FSTRANS_NORMAL)
561 			_fstrans_start(mp, FSTRANS_EXCL, 1);
562 		if (new_state == FSTRANS_NORMAL)
563 			fstrans_done(mp);
564 	}
565 
566 	return error;
567 }
568 
569 /*
570  * Get current file system state.
571  */
572 enum fstrans_state
573 fstrans_getstate(struct mount *mp)
574 {
575 	struct fstrans_mount_info *fmi;
576 
577 	fmi = mp->mnt_transinfo;
578 	KASSERT(fmi != NULL);
579 
580 	return fmi->fmi_state;
581 }
582 
583 /*
584  * Request a filesystem to suspend all operations.
585  */
586 int
587 vfs_suspend(struct mount *mp, int nowait)
588 {
589 	int error;
590 
591 	if ((mp = fstrans_normalize_mount(mp)) == NULL)
592 		return EOPNOTSUPP;
593 	if (nowait) {
594 		if (!mutex_tryenter(&vfs_suspend_lock))
595 			return EWOULDBLOCK;
596 	} else
597 		mutex_enter(&vfs_suspend_lock);
598 
599 	if ((error = VFS_SUSPENDCTL(mp, SUSPEND_SUSPEND)) != 0)
600 		mutex_exit(&vfs_suspend_lock);
601 
602 	return error;
603 }
604 
605 /*
606  * Request a filesystem to resume all operations.
607  */
608 void
609 vfs_resume(struct mount *mp)
610 {
611 
612 	mp = fstrans_normalize_mount(mp);
613 	KASSERT(mp != NULL);
614 
615 	VFS_SUSPENDCTL(mp, SUSPEND_RESUME);
616 	mutex_exit(&vfs_suspend_lock);
617 }
618 
619 
620 /*
621  * True, if no thread is running a cow handler.
622  */
623 static bool
624 cow_state_change_done(const struct mount *mp)
625 {
626 	struct fstrans_lwp_info *fli;
627 	struct fstrans_mount_info *fmi __diagused;
628 
629 	fmi = mp->mnt_transinfo;
630 
631 	KASSERT(mutex_owned(&fstrans_lock));
632 	KASSERT(fmi->fmi_cow_change);
633 
634 	LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
635 		if (fli->fli_mount != mp)
636 			continue;
637 		if (fli->fli_cow_cnt == 0)
638 			continue;
639 
640 		return false;
641 	}
642 
643 	return true;
644 }
645 
646 /*
647  * Prepare for changing this mounts cow list.
648  * Returns with fstrans_lock locked.
649  */
650 static void
651 cow_change_enter(const struct mount *mp)
652 {
653 	struct fstrans_mount_info *fmi;
654 
655 	fmi = mp->mnt_transinfo;
656 
657 	mutex_enter(&fstrans_lock);
658 
659 	/*
660 	 * Wait for other threads changing the list.
661 	 */
662 	while (fmi->fmi_cow_change)
663 		cv_wait(&fstrans_state_cv, &fstrans_lock);
664 
665 	/*
666 	 * Wait until all threads are aware of a state change.
667 	 */
668 	fmi->fmi_cow_change = true;
669 	pserialize_perform(fstrans_psz);
670 
671 	while (! cow_state_change_done(mp))
672 		cv_wait(&fstrans_count_cv, &fstrans_lock);
673 }
674 
675 /*
676  * Done changing this mounts cow list.
677  */
678 static void
679 cow_change_done(const struct mount *mp)
680 {
681 	struct fstrans_mount_info *fmi;
682 
683 	KASSERT(mutex_owned(&fstrans_lock));
684 
685 	fmi = mp->mnt_transinfo;
686 
687 	fmi->fmi_cow_change = false;
688 	pserialize_perform(fstrans_psz);
689 
690 	cv_broadcast(&fstrans_state_cv);
691 
692 	mutex_exit(&fstrans_lock);
693 }
694 
695 /*
696  * Add a handler to this mount.
697  */
698 int
699 fscow_establish(struct mount *mp, int (*func)(void *, struct buf *, bool),
700     void *arg)
701 {
702 	struct fstrans_mount_info *fmi;
703 	struct fscow_handler *newch;
704 
705 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
706 		return EINVAL;
707 
708 	fmi = mp->mnt_transinfo;
709 	KASSERT(fmi != NULL);
710 
711 	newch = kmem_alloc(sizeof(*newch), KM_SLEEP);
712 	newch->ch_func = func;
713 	newch->ch_arg = arg;
714 
715 	cow_change_enter(mp);
716 	LIST_INSERT_HEAD(&fmi->fmi_cow_handler, newch, ch_list);
717 	cow_change_done(mp);
718 
719 	return 0;
720 }
721 
722 /*
723  * Remove a handler from this mount.
724  */
725 int
726 fscow_disestablish(struct mount *mp, int (*func)(void *, struct buf *, bool),
727     void *arg)
728 {
729 	struct fstrans_mount_info *fmi;
730 	struct fscow_handler *hp = NULL;
731 
732 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
733 		return EINVAL;
734 
735 	fmi = mp->mnt_transinfo;
736 	KASSERT(fmi != NULL);
737 
738 	cow_change_enter(mp);
739 	LIST_FOREACH(hp, &fmi->fmi_cow_handler, ch_list)
740 		if (hp->ch_func == func && hp->ch_arg == arg)
741 			break;
742 	if (hp != NULL) {
743 		LIST_REMOVE(hp, ch_list);
744 		kmem_free(hp, sizeof(*hp));
745 	}
746 	cow_change_done(mp);
747 
748 	return hp ? 0 : EINVAL;
749 }
750 
751 /*
752  * Check for need to copy block that is about to be written.
753  */
754 int
755 fscow_run(struct buf *bp, bool data_valid)
756 {
757 	int error, s;
758 	struct mount *mp;
759 	struct fstrans_lwp_info *fli;
760 	struct fstrans_mount_info *fmi;
761 	struct fscow_handler *hp;
762 
763 	/*
764 	 * First check if we need run the copy-on-write handler.
765 	 */
766 	if ((bp->b_flags & B_COWDONE))
767 		return 0;
768 	if (bp->b_vp == NULL) {
769 		bp->b_flags |= B_COWDONE;
770 		return 0;
771 	}
772 	if (bp->b_vp->v_type == VBLK)
773 		mp = spec_node_getmountedfs(bp->b_vp);
774 	else
775 		mp = bp->b_vp->v_mount;
776 	if (mp == NULL || (mp->mnt_iflag & IMNT_HAS_TRANS) == 0) {
777 		bp->b_flags |= B_COWDONE;
778 		return 0;
779 	}
780 
781 	fli = fstrans_get_lwp_info(mp, true);
782 	fmi = mp->mnt_transinfo;
783 
784 	/*
785 	 * On non-recursed run check if other threads
786 	 * want to change the list.
787 	 */
788 	if (fli->fli_cow_cnt == 0) {
789 		s = pserialize_read_enter();
790 		if (__predict_false(fmi->fmi_cow_change)) {
791 			pserialize_read_exit(s);
792 			mutex_enter(&fstrans_lock);
793 			while (fmi->fmi_cow_change)
794 				cv_wait(&fstrans_state_cv, &fstrans_lock);
795 			fli->fli_cow_cnt = 1;
796 			mutex_exit(&fstrans_lock);
797 		} else {
798 			fli->fli_cow_cnt = 1;
799 			pserialize_read_exit(s);
800 		}
801 	} else
802 		fli->fli_cow_cnt += 1;
803 
804 	/*
805 	 * Run all copy-on-write handlers, stop on error.
806 	 */
807 	error = 0;
808 	LIST_FOREACH(hp, &fmi->fmi_cow_handler, ch_list)
809 		if ((error = (*hp->ch_func)(hp->ch_arg, bp, data_valid)) != 0)
810 			break;
811  	if (error == 0)
812  		bp->b_flags |= B_COWDONE;
813 
814 	/*
815 	 * Check if other threads want to change the list.
816 	 */
817 	if (fli->fli_cow_cnt > 1) {
818 		fli->fli_cow_cnt -= 1;
819 	} else {
820 		s = pserialize_read_enter();
821 		if (__predict_false(fmi->fmi_cow_change)) {
822 			pserialize_read_exit(s);
823 			mutex_enter(&fstrans_lock);
824 			fli->fli_cow_cnt = 0;
825 			cv_signal(&fstrans_count_cv);
826 			mutex_exit(&fstrans_lock);
827 		} else {
828 			fli->fli_cow_cnt = 0;
829 			pserialize_read_exit(s);
830 		}
831 	}
832 
833 	return error;
834 }
835 
836 #if defined(DDB)
837 void fstrans_dump(int);
838 
839 static void
840 fstrans_print_lwp(struct proc *p, struct lwp *l, int verbose)
841 {
842 	char prefix[9];
843 	struct fstrans_lwp_info *fli;
844 
845 	snprintf(prefix, sizeof(prefix), "%d.%d", p->p_pid, l->l_lid);
846 	LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
847 		if (fli->fli_self != l)
848 			continue;
849 		if (fli->fli_trans_cnt == 0 && fli->fli_cow_cnt == 0) {
850 			if (! verbose)
851 				continue;
852 		}
853 		printf("%-8s", prefix);
854 		if (verbose)
855 			printf(" @%p", fli);
856 		if (fli->fli_mount != NULL)
857 			printf(" (%s)", fli->fli_mount->mnt_stat.f_mntonname);
858 		else
859 			printf(" NULL");
860 		if (fli->fli_trans_cnt == 0) {
861 			printf(" -");
862 		} else {
863 			switch (fli->fli_lock_type) {
864 			case FSTRANS_LAZY:
865 				printf(" lazy");
866 				break;
867 			case FSTRANS_SHARED:
868 				printf(" shared");
869 				break;
870 			case FSTRANS_EXCL:
871 				printf(" excl");
872 				break;
873 			default:
874 				printf(" %#x", fli->fli_lock_type);
875 				break;
876 			}
877 		}
878 		printf(" %d cow %d\n", fli->fli_trans_cnt, fli->fli_cow_cnt);
879 		prefix[0] = '\0';
880 	}
881 }
882 
883 static void
884 fstrans_print_mount(struct mount *mp, int verbose)
885 {
886 	struct fstrans_mount_info *fmi;
887 
888 	fmi = mp->mnt_transinfo;
889 	if (!verbose && (fmi == NULL || fmi->fmi_state == FSTRANS_NORMAL))
890 		return;
891 
892 	printf("%-16s ", mp->mnt_stat.f_mntonname);
893 	if (fmi == NULL) {
894 		printf("(null)\n");
895 		return;
896 	}
897 	switch (fmi->fmi_state) {
898 	case FSTRANS_NORMAL:
899 		printf("state normal\n");
900 		break;
901 	case FSTRANS_SUSPENDING:
902 		printf("state suspending\n");
903 		break;
904 	case FSTRANS_SUSPENDED:
905 		printf("state suspended\n");
906 		break;
907 	default:
908 		printf("state %#x\n", fmi->fmi_state);
909 		break;
910 	}
911 }
912 
913 void
914 fstrans_dump(int full)
915 {
916 	const struct proclist_desc *pd;
917 	struct proc *p;
918 	struct lwp *l;
919 	struct mount *mp;
920 
921 	printf("Fstrans locks by lwp:\n");
922 	for (pd = proclists; pd->pd_list != NULL; pd++)
923 		PROCLIST_FOREACH(p, pd->pd_list)
924 			LIST_FOREACH(l, &p->p_lwps, l_sibling)
925 				fstrans_print_lwp(p, l, full == 1);
926 
927 	printf("Fstrans state by mount:\n");
928 	for (mp = _mountlist_next(NULL); mp; mp = _mountlist_next(mp))
929 		fstrans_print_mount(mp, full == 1);
930 }
931 #endif /* defined(DDB) */
932