Home | History | Annotate | Line # | Download | only in kern
vfs_trans.c revision 1.47
      1 /*	$NetBSD: vfs_trans.c,v 1.47 2017/06/18 13:59:45 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 1.47 2017/06/18 13:59:45 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_SHARED,			/* Granted while not suspending */
     58 	FSTRANS_EXCL			/* Internal: exclusive lock */
     59 };
     60 
     61 struct fscow_handler {
     62 	LIST_ENTRY(fscow_handler) ch_list;
     63 	int (*ch_func)(void *, struct buf *, bool);
     64 	void *ch_arg;
     65 };
     66 struct fstrans_lwp_info {
     67 	struct fstrans_lwp_info *fli_succ;
     68 	struct lwp *fli_self;
     69 	struct mount *fli_mount;
     70 	int fli_trans_cnt;
     71 	int fli_cow_cnt;
     72 	enum fstrans_lock_type fli_lock_type;
     73 	LIST_ENTRY(fstrans_lwp_info) fli_list;
     74 };
     75 struct fstrans_mount_info {
     76 	enum fstrans_state fmi_state;
     77 	unsigned int fmi_ref_cnt;
     78 	bool fmi_cow_change;
     79 	LIST_HEAD(, fscow_handler) fmi_cow_handler;
     80 };
     81 
     82 static specificdata_key_t lwp_data_key;	/* Our specific data key. */
     83 static kmutex_t vfs_suspend_lock;	/* Serialize suspensions. */
     84 static kmutex_t fstrans_lock;		/* Fstrans big lock. */
     85 static kmutex_t fstrans_mount_lock;	/* Fstrans mount big lock. */
     86 static kcondvar_t fstrans_state_cv;	/* Fstrans or cow state changed. */
     87 static kcondvar_t fstrans_count_cv;	/* Fstrans or cow count changed. */
     88 static pserialize_t fstrans_psz;	/* Pserialize state. */
     89 static LIST_HEAD(fstrans_lwp_head, fstrans_lwp_info) fstrans_fli_head;
     90 					/* List of all fstrans_lwp_info. */
     91 
     92 static inline struct mount *fstrans_normalize_mount(struct mount *);
     93 static void fstrans_lwp_dtor(void *);
     94 static void fstrans_mount_dtor(struct mount *);
     95 static void fstrans_clear_lwp_info(void);
     96 static struct fstrans_lwp_info *fstrans_get_lwp_info(struct mount *, bool);
     97 static inline int _fstrans_start(struct mount *, enum fstrans_lock_type, int);
     98 static bool grant_lock(const enum fstrans_state, const enum fstrans_lock_type);
     99 static bool state_change_done(const struct mount *);
    100 static bool cow_state_change_done(const struct mount *);
    101 static void cow_change_enter(const struct mount *);
    102 static void cow_change_done(const struct mount *);
    103 
    104 /*
    105  * Initialize.
    106  */
    107 void
    108 fstrans_init(void)
    109 {
    110 	int error __diagused;
    111 
    112 	error = lwp_specific_key_create(&lwp_data_key, fstrans_lwp_dtor);
    113 	KASSERT(error == 0);
    114 
    115 	mutex_init(&vfs_suspend_lock, MUTEX_DEFAULT, IPL_NONE);
    116 	mutex_init(&fstrans_lock, MUTEX_DEFAULT, IPL_NONE);
    117 	mutex_init(&fstrans_mount_lock, MUTEX_DEFAULT, IPL_NONE);
    118 	cv_init(&fstrans_state_cv, "fstchg");
    119 	cv_init(&fstrans_count_cv, "fstcnt");
    120 	fstrans_psz = pserialize_create();
    121 	LIST_INIT(&fstrans_fli_head);
    122 }
    123 
    124 /*
    125  * Normalize mount.
    126  * Return mount if file system supports fstrans, NULL otherwise.
    127  */
    128 static inline struct mount *
    129 fstrans_normalize_mount(struct mount *mp)
    130 {
    131 
    132 	while (mp && mp->mnt_lower)
    133 		mp = mp->mnt_lower;
    134 	if (mp == NULL)
    135 		return NULL;
    136 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
    137 		return NULL;
    138 	return mp;
    139 }
    140 
    141 /*
    142  * Deallocate lwp state.
    143  */
    144 static void
    145 fstrans_lwp_dtor(void *arg)
    146 {
    147 	struct fstrans_lwp_info *fli, *fli_next;
    148 
    149 	for (fli = arg; fli; fli = fli_next) {
    150 		KASSERT(fli->fli_trans_cnt == 0);
    151 		KASSERT(fli->fli_cow_cnt == 0);
    152 		if (fli->fli_mount != NULL)
    153 			fstrans_mount_dtor(fli->fli_mount);
    154 		fli_next = fli->fli_succ;
    155 		fli->fli_mount = NULL;
    156 		membar_sync();
    157 		fli->fli_self = NULL;
    158 	}
    159 }
    160 
    161 /*
    162  * Dereference mount state.
    163  */
    164 static void
    165 fstrans_mount_dtor(struct mount *mp)
    166 {
    167 	struct fstrans_mount_info *fmi;
    168 
    169 	mutex_enter(&fstrans_mount_lock);
    170 
    171 	fmi = mp->mnt_transinfo;
    172 	KASSERT(fmi != NULL);
    173 	fmi->fmi_ref_cnt -= 1;
    174 	if (fmi->fmi_ref_cnt > 0) {
    175 		mutex_exit(&fstrans_mount_lock);
    176 		return;
    177 	}
    178 
    179 	KASSERT(fmi->fmi_state == FSTRANS_NORMAL);
    180 	KASSERT(LIST_FIRST(&fmi->fmi_cow_handler) == NULL);
    181 
    182 	mp->mnt_iflag &= ~IMNT_HAS_TRANS;
    183 	mp->mnt_transinfo = NULL;
    184 
    185 	mutex_exit(&fstrans_mount_lock);
    186 
    187 	kmem_free(fmi, sizeof(*fmi));
    188 	vfs_rele(mp);
    189 }
    190 
    191 /*
    192  * Allocate mount state.
    193  */
    194 int
    195 fstrans_mount(struct mount *mp)
    196 {
    197 	struct fstrans_mount_info *newfmi;
    198 
    199 	newfmi = kmem_alloc(sizeof(*newfmi), KM_SLEEP);
    200 	newfmi->fmi_state = FSTRANS_NORMAL;
    201 	newfmi->fmi_ref_cnt = 1;
    202 	LIST_INIT(&newfmi->fmi_cow_handler);
    203 	newfmi->fmi_cow_change = false;
    204 
    205 	mutex_enter(&fstrans_mount_lock);
    206 	mp->mnt_transinfo = newfmi;
    207 	mp->mnt_iflag |= IMNT_HAS_TRANS;
    208 	mutex_exit(&fstrans_mount_lock);
    209 
    210 	vfs_ref(mp);
    211 
    212 	return 0;
    213 }
    214 
    215 /*
    216  * Deallocate mount state.
    217  */
    218 void
    219 fstrans_unmount(struct mount *mp)
    220 {
    221 
    222 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
    223 		return;
    224 
    225 	KASSERT(mp->mnt_transinfo != NULL);
    226 
    227 	fstrans_mount_dtor(mp);
    228 }
    229 
    230 /*
    231  * Clear mount entries whose mount is gone.
    232  */
    233 static void
    234 fstrans_clear_lwp_info(void)
    235 {
    236 	struct fstrans_lwp_info *fli;
    237 
    238 	/*
    239 	 * Scan our list clearing entries whose mount is gone.
    240 	 */
    241 	for (fli = lwp_getspecific(lwp_data_key); fli; fli = fli->fli_succ) {
    242 		if (fli->fli_mount != NULL &&
    243 		    (fli->fli_mount->mnt_iflag & IMNT_GONE) != 0 &&
    244 		    fli->fli_trans_cnt == 0 && fli->fli_cow_cnt == 0) {
    245 			fstrans_mount_dtor(fli->fli_mount);
    246 			fli->fli_mount = NULL;
    247 		}
    248 	}
    249 }
    250 
    251 /*
    252  * Retrieve the per lwp info for this mount allocating if necessary.
    253  */
    254 static struct fstrans_lwp_info *
    255 fstrans_get_lwp_info(struct mount *mp, bool do_alloc)
    256 {
    257 	struct fstrans_lwp_info *fli;
    258 	struct fstrans_mount_info *fmi;
    259 
    260 	/*
    261 	 * Scan our list for a match.
    262 	 */
    263 	for (fli = lwp_getspecific(lwp_data_key); fli; fli = fli->fli_succ) {
    264 		if (fli->fli_mount == mp)
    265 			return fli;
    266 	}
    267 
    268 	if (! do_alloc)
    269 		return NULL;
    270 
    271 	/*
    272 	 * Try to reuse a cleared entry or allocate a new one.
    273 	 */
    274 	for (fli = lwp_getspecific(lwp_data_key); fli; fli = fli->fli_succ) {
    275 		if (fli->fli_mount == NULL) {
    276 			KASSERT(fli->fli_trans_cnt == 0);
    277 			KASSERT(fli->fli_cow_cnt == 0);
    278 			break;
    279 		}
    280 	}
    281 	if (fli == NULL) {
    282 		mutex_enter(&fstrans_lock);
    283 		LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
    284 			if (fli->fli_self == NULL) {
    285 				KASSERT(fli->fli_mount == NULL);
    286 				KASSERT(fli->fli_trans_cnt == 0);
    287 				KASSERT(fli->fli_cow_cnt == 0);
    288 				fli->fli_self = curlwp;
    289 				fli->fli_succ = lwp_getspecific(lwp_data_key);
    290 				lwp_setspecific(lwp_data_key, fli);
    291 				break;
    292 			}
    293 		}
    294 		mutex_exit(&fstrans_lock);
    295 	}
    296 	if (fli == NULL) {
    297 		fli = kmem_alloc(sizeof(*fli), KM_SLEEP);
    298 		mutex_enter(&fstrans_lock);
    299 		memset(fli, 0, sizeof(*fli));
    300 		fli->fli_self = curlwp;
    301 		LIST_INSERT_HEAD(&fstrans_fli_head, fli, fli_list);
    302 		mutex_exit(&fstrans_lock);
    303 		fli->fli_succ = lwp_getspecific(lwp_data_key);
    304 		lwp_setspecific(lwp_data_key, fli);
    305 	}
    306 
    307 	/*
    308 	 * Attach the entry to the mount if its mnt_transinfo is valid.
    309 	 */
    310 	mutex_enter(&fstrans_mount_lock);
    311 	fmi = mp->mnt_transinfo;
    312 	if (__predict_true(fmi != NULL)) {
    313 		fli->fli_mount = mp;
    314 		fmi->fmi_ref_cnt += 1;
    315 	} else {
    316 		fli = NULL;
    317 	}
    318 	mutex_exit(&fstrans_mount_lock);
    319 
    320 	return fli;
    321 }
    322 
    323 /*
    324  * Check if this lock type is granted at this state.
    325  */
    326 static bool
    327 grant_lock(const enum fstrans_state state, const enum fstrans_lock_type type)
    328 {
    329 
    330 	if (__predict_true(state == FSTRANS_NORMAL))
    331 		return true;
    332 	if (type == FSTRANS_EXCL)
    333 		return true;
    334 
    335 	return false;
    336 }
    337 
    338 /*
    339  * Start a transaction.  If this thread already has a transaction on this
    340  * file system increment the reference counter.
    341  */
    342 static inline int
    343 _fstrans_start(struct mount *mp, enum fstrans_lock_type lock_type, int wait)
    344 {
    345 	int s;
    346 	struct mount *lmp;
    347 	struct fstrans_lwp_info *fli;
    348 	struct fstrans_mount_info *fmi;
    349 
    350 	if ((lmp = fstrans_normalize_mount(mp)) == NULL)
    351 		return 0;
    352 
    353 	ASSERT_SLEEPABLE();
    354 
    355 	/*
    356 	 * Allocate per lwp info for layered file systems to
    357 	 * get a reference to the mount.  No need to increment
    358 	 * the reference counter here.
    359 	 */
    360 	for (lmp = mp; lmp->mnt_lower; lmp = lmp->mnt_lower) {
    361 		fli = fstrans_get_lwp_info(lmp, true);
    362 	}
    363 
    364 	if ((fli = fstrans_get_lwp_info(lmp, true)) == NULL)
    365 		return 0;
    366 
    367 	if (fli->fli_trans_cnt > 0) {
    368 		KASSERT(lock_type != FSTRANS_EXCL);
    369 		fli->fli_trans_cnt += 1;
    370 
    371 		return 0;
    372 	}
    373 
    374 	s = pserialize_read_enter();
    375 	fmi = lmp->mnt_transinfo;
    376 	if (__predict_true(grant_lock(fmi->fmi_state, lock_type))) {
    377 		fli->fli_trans_cnt = 1;
    378 		fli->fli_lock_type = lock_type;
    379 		pserialize_read_exit(s);
    380 
    381 		return 0;
    382 	}
    383 	pserialize_read_exit(s);
    384 
    385 	if (! wait)
    386 		return EBUSY;
    387 
    388 	mutex_enter(&fstrans_lock);
    389 	while (! grant_lock(fmi->fmi_state, lock_type))
    390 		cv_wait(&fstrans_state_cv, &fstrans_lock);
    391 	fli->fli_trans_cnt = 1;
    392 	fli->fli_lock_type = lock_type;
    393 	mutex_exit(&fstrans_lock);
    394 
    395 	return 0;
    396 }
    397 
    398 void
    399 fstrans_start(struct mount *mp)
    400 {
    401 	int error __diagused;
    402 
    403 	error = _fstrans_start(mp, FSTRANS_SHARED, 1);
    404 	KASSERT(error == 0);
    405 }
    406 
    407 int
    408 fstrans_start_nowait(struct mount *mp)
    409 {
    410 
    411 	return _fstrans_start(mp, FSTRANS_SHARED, 0);
    412 }
    413 
    414 /*
    415  * Finish a transaction.
    416  */
    417 void
    418 fstrans_done(struct mount *mp)
    419 {
    420 	int s;
    421 	struct fstrans_lwp_info *fli;
    422 	struct fstrans_mount_info *fmi;
    423 
    424 	if ((mp = fstrans_normalize_mount(mp)) == NULL)
    425 		return;
    426 	if ((fli = fstrans_get_lwp_info(mp, false)) == NULL)
    427 		return;
    428 	KASSERT(fli->fli_trans_cnt > 0);
    429 
    430 	if (fli->fli_trans_cnt > 1) {
    431 		fli->fli_trans_cnt -= 1;
    432 
    433 		return;
    434 	}
    435 
    436 	fstrans_clear_lwp_info();
    437 
    438 	s = pserialize_read_enter();
    439 	fmi = mp->mnt_transinfo;
    440 	if (__predict_true(fmi->fmi_state == FSTRANS_NORMAL)) {
    441 		fli->fli_trans_cnt = 0;
    442 		pserialize_read_exit(s);
    443 
    444 		return;
    445 	}
    446 	pserialize_read_exit(s);
    447 
    448 	mutex_enter(&fstrans_lock);
    449 	fli->fli_trans_cnt = 0;
    450 	cv_signal(&fstrans_count_cv);
    451 	mutex_exit(&fstrans_lock);
    452 }
    453 
    454 /*
    455  * Check if this thread has an exclusive lock.
    456  */
    457 int
    458 fstrans_is_owner(struct mount *mp)
    459 {
    460 	struct fstrans_lwp_info *fli;
    461 
    462 	if ((mp = fstrans_normalize_mount(mp)) == NULL)
    463 		return 0;
    464 	if ((fli = fstrans_get_lwp_info(mp, false)) == NULL)
    465 		return 0;
    466 
    467 	if (fli->fli_trans_cnt == 0)
    468 		return 0;
    469 
    470 	KASSERT(fli->fli_mount == mp);
    471 	KASSERT(fli->fli_trans_cnt > 0);
    472 
    473 	return (fli->fli_lock_type == FSTRANS_EXCL);
    474 }
    475 
    476 /*
    477  * True, if no thread is in a transaction not granted at the current state.
    478  */
    479 static bool
    480 state_change_done(const struct mount *mp)
    481 {
    482 	struct fstrans_lwp_info *fli;
    483 	struct fstrans_mount_info *fmi;
    484 
    485 	KASSERT(mutex_owned(&fstrans_lock));
    486 
    487 	fmi = mp->mnt_transinfo;
    488 	LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
    489 		if (fli->fli_mount != mp)
    490 			continue;
    491 		if (fli->fli_trans_cnt == 0)
    492 			continue;
    493 		if (grant_lock(fmi->fmi_state, fli->fli_lock_type))
    494 			continue;
    495 
    496 		return false;
    497 	}
    498 
    499 	return true;
    500 }
    501 
    502 /*
    503  * Set new file system state.
    504  */
    505 int
    506 fstrans_setstate(struct mount *mp, enum fstrans_state new_state)
    507 {
    508 	int error;
    509 	enum fstrans_state old_state;
    510 	struct fstrans_mount_info *fmi;
    511 
    512 	fmi = mp->mnt_transinfo;
    513 	old_state = fmi->fmi_state;
    514 	if (old_state == new_state)
    515 		return 0;
    516 
    517 	mutex_enter(&fstrans_lock);
    518 	fmi->fmi_state = new_state;
    519 	pserialize_perform(fstrans_psz);
    520 
    521 	/*
    522 	 * All threads see the new state now.
    523 	 * Wait for transactions invalid at this state to leave.
    524 	 */
    525 	error = 0;
    526 	while (! state_change_done(mp)) {
    527 		error = cv_wait_sig(&fstrans_count_cv, &fstrans_lock);
    528 		if (error) {
    529 			new_state = fmi->fmi_state = FSTRANS_NORMAL;
    530 			break;
    531 		}
    532 	}
    533 	cv_broadcast(&fstrans_state_cv);
    534 	mutex_exit(&fstrans_lock);
    535 
    536 	if (old_state != new_state) {
    537 		if (old_state == FSTRANS_NORMAL)
    538 			_fstrans_start(mp, FSTRANS_EXCL, 1);
    539 		if (new_state == FSTRANS_NORMAL)
    540 			fstrans_done(mp);
    541 	}
    542 
    543 	return error;
    544 }
    545 
    546 /*
    547  * Get current file system state.
    548  */
    549 enum fstrans_state
    550 fstrans_getstate(struct mount *mp)
    551 {
    552 	struct fstrans_mount_info *fmi;
    553 
    554 	fmi = mp->mnt_transinfo;
    555 	KASSERT(fmi != NULL);
    556 
    557 	return fmi->fmi_state;
    558 }
    559 
    560 /*
    561  * Request a filesystem to suspend all operations.
    562  */
    563 int
    564 vfs_suspend(struct mount *mp, int nowait)
    565 {
    566 	int error;
    567 
    568 	if ((mp = fstrans_normalize_mount(mp)) == NULL)
    569 		return EOPNOTSUPP;
    570 	if (nowait) {
    571 		if (!mutex_tryenter(&vfs_suspend_lock))
    572 			return EWOULDBLOCK;
    573 	} else
    574 		mutex_enter(&vfs_suspend_lock);
    575 
    576 	if ((error = VFS_SUSPENDCTL(mp, SUSPEND_SUSPEND)) != 0)
    577 		mutex_exit(&vfs_suspend_lock);
    578 
    579 	return error;
    580 }
    581 
    582 /*
    583  * Request a filesystem to resume all operations.
    584  */
    585 void
    586 vfs_resume(struct mount *mp)
    587 {
    588 
    589 	mp = fstrans_normalize_mount(mp);
    590 	KASSERT(mp != NULL);
    591 
    592 	VFS_SUSPENDCTL(mp, SUSPEND_RESUME);
    593 	mutex_exit(&vfs_suspend_lock);
    594 }
    595 
    596 
    597 /*
    598  * True, if no thread is running a cow handler.
    599  */
    600 static bool
    601 cow_state_change_done(const struct mount *mp)
    602 {
    603 	struct fstrans_lwp_info *fli;
    604 	struct fstrans_mount_info *fmi __diagused;
    605 
    606 	fmi = mp->mnt_transinfo;
    607 
    608 	KASSERT(mutex_owned(&fstrans_lock));
    609 	KASSERT(fmi->fmi_cow_change);
    610 
    611 	LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
    612 		if (fli->fli_mount != mp)
    613 			continue;
    614 		if (fli->fli_cow_cnt == 0)
    615 			continue;
    616 
    617 		return false;
    618 	}
    619 
    620 	return true;
    621 }
    622 
    623 /*
    624  * Prepare for changing this mounts cow list.
    625  * Returns with fstrans_lock locked.
    626  */
    627 static void
    628 cow_change_enter(const struct mount *mp)
    629 {
    630 	struct fstrans_mount_info *fmi;
    631 
    632 	fmi = mp->mnt_transinfo;
    633 
    634 	mutex_enter(&fstrans_lock);
    635 
    636 	/*
    637 	 * Wait for other threads changing the list.
    638 	 */
    639 	while (fmi->fmi_cow_change)
    640 		cv_wait(&fstrans_state_cv, &fstrans_lock);
    641 
    642 	/*
    643 	 * Wait until all threads are aware of a state change.
    644 	 */
    645 	fmi->fmi_cow_change = true;
    646 	pserialize_perform(fstrans_psz);
    647 
    648 	while (! cow_state_change_done(mp))
    649 		cv_wait(&fstrans_count_cv, &fstrans_lock);
    650 }
    651 
    652 /*
    653  * Done changing this mounts cow list.
    654  */
    655 static void
    656 cow_change_done(const struct mount *mp)
    657 {
    658 	struct fstrans_mount_info *fmi;
    659 
    660 	KASSERT(mutex_owned(&fstrans_lock));
    661 
    662 	fmi = mp->mnt_transinfo;
    663 
    664 	fmi->fmi_cow_change = false;
    665 	pserialize_perform(fstrans_psz);
    666 
    667 	cv_broadcast(&fstrans_state_cv);
    668 
    669 	mutex_exit(&fstrans_lock);
    670 }
    671 
    672 /*
    673  * Add a handler to this mount.
    674  */
    675 int
    676 fscow_establish(struct mount *mp, int (*func)(void *, struct buf *, bool),
    677     void *arg)
    678 {
    679 	struct fstrans_mount_info *fmi;
    680 	struct fscow_handler *newch;
    681 
    682 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
    683 		return EINVAL;
    684 
    685 	fmi = mp->mnt_transinfo;
    686 	KASSERT(fmi != NULL);
    687 
    688 	newch = kmem_alloc(sizeof(*newch), KM_SLEEP);
    689 	newch->ch_func = func;
    690 	newch->ch_arg = arg;
    691 
    692 	cow_change_enter(mp);
    693 	LIST_INSERT_HEAD(&fmi->fmi_cow_handler, newch, ch_list);
    694 	cow_change_done(mp);
    695 
    696 	return 0;
    697 }
    698 
    699 /*
    700  * Remove a handler from this mount.
    701  */
    702 int
    703 fscow_disestablish(struct mount *mp, int (*func)(void *, struct buf *, bool),
    704     void *arg)
    705 {
    706 	struct fstrans_mount_info *fmi;
    707 	struct fscow_handler *hp = NULL;
    708 
    709 	if ((mp->mnt_iflag & IMNT_HAS_TRANS) == 0)
    710 		return EINVAL;
    711 
    712 	fmi = mp->mnt_transinfo;
    713 	KASSERT(fmi != NULL);
    714 
    715 	cow_change_enter(mp);
    716 	LIST_FOREACH(hp, &fmi->fmi_cow_handler, ch_list)
    717 		if (hp->ch_func == func && hp->ch_arg == arg)
    718 			break;
    719 	if (hp != NULL) {
    720 		LIST_REMOVE(hp, ch_list);
    721 		kmem_free(hp, sizeof(*hp));
    722 	}
    723 	cow_change_done(mp);
    724 
    725 	return hp ? 0 : EINVAL;
    726 }
    727 
    728 /*
    729  * Check for need to copy block that is about to be written.
    730  */
    731 int
    732 fscow_run(struct buf *bp, bool data_valid)
    733 {
    734 	int error, s;
    735 	struct mount *mp;
    736 	struct fstrans_lwp_info *fli;
    737 	struct fstrans_mount_info *fmi;
    738 	struct fscow_handler *hp;
    739 
    740 	/*
    741 	 * First check if we need run the copy-on-write handler.
    742 	 */
    743 	if ((bp->b_flags & B_COWDONE))
    744 		return 0;
    745 	if (bp->b_vp == NULL) {
    746 		bp->b_flags |= B_COWDONE;
    747 		return 0;
    748 	}
    749 	if (bp->b_vp->v_type == VBLK)
    750 		mp = spec_node_getmountedfs(bp->b_vp);
    751 	else
    752 		mp = bp->b_vp->v_mount;
    753 	if (mp == NULL || (mp->mnt_iflag & IMNT_HAS_TRANS) == 0) {
    754 		bp->b_flags |= B_COWDONE;
    755 		return 0;
    756 	}
    757 
    758 	fli = fstrans_get_lwp_info(mp, true);
    759 	fmi = mp->mnt_transinfo;
    760 
    761 	/*
    762 	 * On non-recursed run check if other threads
    763 	 * want to change the list.
    764 	 */
    765 	if (fli->fli_cow_cnt == 0) {
    766 		s = pserialize_read_enter();
    767 		if (__predict_false(fmi->fmi_cow_change)) {
    768 			pserialize_read_exit(s);
    769 			mutex_enter(&fstrans_lock);
    770 			while (fmi->fmi_cow_change)
    771 				cv_wait(&fstrans_state_cv, &fstrans_lock);
    772 			fli->fli_cow_cnt = 1;
    773 			mutex_exit(&fstrans_lock);
    774 		} else {
    775 			fli->fli_cow_cnt = 1;
    776 			pserialize_read_exit(s);
    777 		}
    778 	} else
    779 		fli->fli_cow_cnt += 1;
    780 
    781 	/*
    782 	 * Run all copy-on-write handlers, stop on error.
    783 	 */
    784 	error = 0;
    785 	LIST_FOREACH(hp, &fmi->fmi_cow_handler, ch_list)
    786 		if ((error = (*hp->ch_func)(hp->ch_arg, bp, data_valid)) != 0)
    787 			break;
    788  	if (error == 0)
    789  		bp->b_flags |= B_COWDONE;
    790 
    791 	/*
    792 	 * Check if other threads want to change the list.
    793 	 */
    794 	if (fli->fli_cow_cnt > 1) {
    795 		fli->fli_cow_cnt -= 1;
    796 	} else {
    797 		s = pserialize_read_enter();
    798 		if (__predict_false(fmi->fmi_cow_change)) {
    799 			pserialize_read_exit(s);
    800 			mutex_enter(&fstrans_lock);
    801 			fli->fli_cow_cnt = 0;
    802 			cv_signal(&fstrans_count_cv);
    803 			mutex_exit(&fstrans_lock);
    804 		} else {
    805 			fli->fli_cow_cnt = 0;
    806 			pserialize_read_exit(s);
    807 		}
    808 	}
    809 
    810 	return error;
    811 }
    812 
    813 #if defined(DDB)
    814 void fstrans_dump(int);
    815 
    816 static void
    817 fstrans_print_lwp(struct proc *p, struct lwp *l, int verbose)
    818 {
    819 	char prefix[9];
    820 	struct fstrans_lwp_info *fli;
    821 
    822 	snprintf(prefix, sizeof(prefix), "%d.%d", p->p_pid, l->l_lid);
    823 	LIST_FOREACH(fli, &fstrans_fli_head, fli_list) {
    824 		if (fli->fli_self != l)
    825 			continue;
    826 		if (fli->fli_trans_cnt == 0 && fli->fli_cow_cnt == 0) {
    827 			if (! verbose)
    828 				continue;
    829 		}
    830 		printf("%-8s", prefix);
    831 		if (verbose)
    832 			printf(" @%p", fli);
    833 		if (fli->fli_mount != NULL)
    834 			printf(" (%s)", fli->fli_mount->mnt_stat.f_mntonname);
    835 		else
    836 			printf(" NULL");
    837 		if (fli->fli_trans_cnt == 0) {
    838 			printf(" -");
    839 		} else {
    840 			switch (fli->fli_lock_type) {
    841 			case FSTRANS_SHARED:
    842 				printf(" shared");
    843 				break;
    844 			case FSTRANS_EXCL:
    845 				printf(" excl");
    846 				break;
    847 			default:
    848 				printf(" %#x", fli->fli_lock_type);
    849 				break;
    850 			}
    851 		}
    852 		printf(" %d cow %d\n", fli->fli_trans_cnt, fli->fli_cow_cnt);
    853 		prefix[0] = '\0';
    854 	}
    855 }
    856 
    857 static void
    858 fstrans_print_mount(struct mount *mp, int verbose)
    859 {
    860 	struct fstrans_mount_info *fmi;
    861 
    862 	fmi = mp->mnt_transinfo;
    863 	if (!verbose && (fmi == NULL || fmi->fmi_state == FSTRANS_NORMAL))
    864 		return;
    865 
    866 	printf("%-16s ", mp->mnt_stat.f_mntonname);
    867 	if (fmi == NULL) {
    868 		printf("(null)\n");
    869 		return;
    870 	}
    871 	switch (fmi->fmi_state) {
    872 	case FSTRANS_NORMAL:
    873 		printf("state normal\n");
    874 		break;
    875 	case FSTRANS_SUSPENDED:
    876 		printf("state suspended\n");
    877 		break;
    878 	default:
    879 		printf("state %#x\n", fmi->fmi_state);
    880 		break;
    881 	}
    882 }
    883 
    884 void
    885 fstrans_dump(int full)
    886 {
    887 	const struct proclist_desc *pd;
    888 	struct proc *p;
    889 	struct lwp *l;
    890 	struct mount *mp;
    891 
    892 	printf("Fstrans locks by lwp:\n");
    893 	for (pd = proclists; pd->pd_list != NULL; pd++)
    894 		PROCLIST_FOREACH(p, pd->pd_list)
    895 			LIST_FOREACH(l, &p->p_lwps, l_sibling)
    896 				fstrans_print_lwp(p, l, full == 1);
    897 
    898 	printf("Fstrans state by mount:\n");
    899 	for (mp = _mountlist_next(NULL); mp; mp = _mountlist_next(mp))
    900 		fstrans_print_mount(mp, full == 1);
    901 }
    902 #endif /* defined(DDB) */
    903