xref: /netbsd-src/sys/kern/vfs_wapbl.c (revision a5847cc334d9a7029f6352b847e9e8d71a0f9e0c)
1 /*	$NetBSD: vfs_wapbl.c,v 1.47 2011/09/01 09:03:43 christos Exp $	*/
2 
3 /*-
4  * Copyright (c) 2003, 2008, 2009 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Wasabi Systems, Inc.
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 /*
33  * This implements file system independent write ahead filesystem logging.
34  */
35 
36 #define WAPBL_INTERNAL
37 
38 #include <sys/cdefs.h>
39 __KERNEL_RCSID(0, "$NetBSD: vfs_wapbl.c,v 1.47 2011/09/01 09:03:43 christos Exp $");
40 
41 #include <sys/param.h>
42 #include <sys/bitops.h>
43 
44 #ifdef _KERNEL
45 #include <sys/param.h>
46 #include <sys/namei.h>
47 #include <sys/proc.h>
48 #include <sys/sysctl.h>
49 #include <sys/uio.h>
50 #include <sys/vnode.h>
51 #include <sys/file.h>
52 #include <sys/malloc.h>
53 #include <sys/module.h>
54 #include <sys/resourcevar.h>
55 #include <sys/conf.h>
56 #include <sys/mount.h>
57 #include <sys/kernel.h>
58 #include <sys/kauth.h>
59 #include <sys/mutex.h>
60 #include <sys/atomic.h>
61 #include <sys/wapbl.h>
62 #include <sys/wapbl_replay.h>
63 
64 #include <miscfs/specfs/specdev.h>
65 
66 #if 0 /* notyet */
67 #define	wapbl_malloc(s) kmem_alloc((s), KM_SLEEP)
68 #define	wapbl_free(a, s) kmem_free((a), (s))
69 #define	wapbl_calloc(n, s) kmem_zalloc((n)*(s), KM_SLEEP)
70 #else
71 MALLOC_JUSTDEFINE(M_WAPBL, "wapbl", "write-ahead physical block logging");
72 #define	wapbl_malloc(s) malloc((s), M_WAPBL, M_WAITOK)
73 #define	wapbl_free(a, s) free((a), M_WAPBL)
74 #define	wapbl_calloc(n, s) malloc((n)*(s), M_WAPBL, M_WAITOK | M_ZERO)
75 #endif
76 
77 static struct sysctllog *wapbl_sysctl;
78 static int wapbl_flush_disk_cache = 1;
79 static int wapbl_verbose_commit = 0;
80 
81 #else /* !_KERNEL */
82 #include <assert.h>
83 #include <errno.h>
84 #include <stdio.h>
85 #include <stdbool.h>
86 #include <stdlib.h>
87 #include <string.h>
88 
89 #include <sys/time.h>
90 #include <sys/wapbl.h>
91 #include <sys/wapbl_replay.h>
92 
93 #define	KDASSERT(x) assert(x)
94 #define	KASSERT(x) assert(x)
95 #define	wapbl_malloc(s) malloc(s)
96 #define	wapbl_free(a, s) free(a)
97 #define	wapbl_calloc(n, s) calloc((n), (s))
98 
99 #endif /* !_KERNEL */
100 
101 /*
102  * INTERNAL DATA STRUCTURES
103  */
104 
105 /*
106  * This structure holds per-mount log information.
107  *
108  * Legend:	a = atomic access only
109  *		r = read-only after init
110  *		l = rwlock held
111  *		m = mutex held
112  *		lm = rwlock held writing or mutex held
113  *		u = unlocked access ok
114  *		b = bufcache_lock held
115  */
116 struct wapbl {
117 	struct vnode *wl_logvp;	/* r:	log here */
118 	struct vnode *wl_devvp;	/* r:	log on this device */
119 	struct mount *wl_mount;	/* r:	mountpoint wl is associated with */
120 	daddr_t wl_logpbn;	/* r:	Physical block number of start of log */
121 	int wl_log_dev_bshift;	/* r:	logarithm of device block size of log
122 					device */
123 	int wl_fs_dev_bshift;	/* r:	logarithm of device block size of
124 					filesystem device */
125 
126 	unsigned wl_lock_count;	/* m:	Count of transactions in progress */
127 
128 	size_t wl_circ_size; 	/* r:	Number of bytes in buffer of log */
129 	size_t wl_circ_off;	/* r:	Number of bytes reserved at start */
130 
131 	size_t wl_bufcount_max;	/* r:	Number of buffers reserved for log */
132 	size_t wl_bufbytes_max;	/* r:	Number of buf bytes reserved for log */
133 
134 	off_t wl_head;		/* l:	Byte offset of log head */
135 	off_t wl_tail;		/* l:	Byte offset of log tail */
136 	/*
137 	 * head == tail == 0 means log is empty
138 	 * head == tail != 0 means log is full
139 	 * see assertions in wapbl_advance() for other boundary conditions.
140 	 * only truncate moves the tail, except when flush sets it to
141 	 * wl_header_size only flush moves the head, except when truncate
142 	 * sets it to 0.
143 	 */
144 
145 	struct wapbl_wc_header *wl_wc_header;	/* l	*/
146 	void *wl_wc_scratch;	/* l:	scratch space (XXX: por que?!?) */
147 
148 	kmutex_t wl_mtx;	/* u:	short-term lock */
149 	krwlock_t wl_rwlock;	/* u:	File system transaction lock */
150 
151 	/*
152 	 * Must be held while accessing
153 	 * wl_count or wl_bufs or head or tail
154 	 */
155 
156 	/*
157 	 * Callback called from within the flush routine to flush any extra
158 	 * bits.  Note that flush may be skipped without calling this if
159 	 * there are no outstanding buffers in the transaction.
160 	 */
161 #if _KERNEL
162 	wapbl_flush_fn_t wl_flush;	/* r	*/
163 	wapbl_flush_fn_t wl_flush_abort;/* r	*/
164 #endif
165 
166 	size_t wl_bufbytes;	/* m:	Byte count of pages in wl_bufs */
167 	size_t wl_bufcount;	/* m:	Count of buffers in wl_bufs */
168 	size_t wl_bcount;	/* m:	Total bcount of wl_bufs */
169 
170 	LIST_HEAD(, buf) wl_bufs; /* m:	Buffers in current transaction */
171 
172 	kcondvar_t wl_reclaimable_cv;	/* m (obviously) */
173 	size_t wl_reclaimable_bytes; /* m:	Amount of space available for
174 						reclamation by truncate */
175 	int wl_error_count;	/* m:	# of wl_entries with errors */
176 	size_t wl_reserved_bytes; /* never truncate log smaller than this */
177 
178 #ifdef WAPBL_DEBUG_BUFBYTES
179 	size_t wl_unsynced_bufbytes; /* Byte count of unsynced buffers */
180 #endif
181 
182 	daddr_t *wl_deallocblks;/* lm:	address of block */
183 	int *wl_dealloclens;	/* lm:	size of block */
184 	int wl_dealloccnt;	/* lm:	total count */
185 	int wl_dealloclim;	/* l:	max count */
186 
187 	/* hashtable of inode numbers for allocated but unlinked inodes */
188 	/* synch ??? */
189 	LIST_HEAD(wapbl_ino_head, wapbl_ino) *wl_inohash;
190 	u_long wl_inohashmask;
191 	int wl_inohashcnt;
192 
193 	SIMPLEQ_HEAD(, wapbl_entry) wl_entries; /* On disk transaction
194 						   accounting */
195 };
196 
197 #ifdef WAPBL_DEBUG_PRINT
198 int wapbl_debug_print = WAPBL_DEBUG_PRINT;
199 #endif
200 
201 /****************************************************************/
202 #ifdef _KERNEL
203 
204 #ifdef WAPBL_DEBUG
205 struct wapbl *wapbl_debug_wl;
206 #endif
207 
208 static int wapbl_write_commit(struct wapbl *wl, off_t head, off_t tail);
209 static int wapbl_write_blocks(struct wapbl *wl, off_t *offp);
210 static int wapbl_write_revocations(struct wapbl *wl, off_t *offp);
211 static int wapbl_write_inodes(struct wapbl *wl, off_t *offp);
212 #endif /* _KERNEL */
213 
214 static int wapbl_replay_process(struct wapbl_replay *wr, off_t, off_t);
215 
216 static inline size_t wapbl_space_free(size_t avail, off_t head,
217 	off_t tail);
218 static inline size_t wapbl_space_used(size_t avail, off_t head,
219 	off_t tail);
220 
221 #ifdef _KERNEL
222 
223 #define	WAPBL_INODETRK_SIZE 83
224 static int wapbl_ino_pool_refcount;
225 static struct pool wapbl_ino_pool;
226 struct wapbl_ino {
227 	LIST_ENTRY(wapbl_ino) wi_hash;
228 	ino_t wi_ino;
229 	mode_t wi_mode;
230 };
231 
232 static void wapbl_inodetrk_init(struct wapbl *wl, u_int size);
233 static void wapbl_inodetrk_free(struct wapbl *wl);
234 static struct wapbl_ino *wapbl_inodetrk_get(struct wapbl *wl, ino_t ino);
235 
236 static size_t wapbl_transaction_len(struct wapbl *wl);
237 static inline size_t wapbl_transaction_inodes_len(struct wapbl *wl);
238 
239 #if 0
240 int wapbl_replay_verify(struct wapbl_replay *, struct vnode *);
241 #endif
242 
243 static int wapbl_replay_isopen1(struct wapbl_replay *);
244 
245 /*
246  * This is useful for debugging.  If set, the log will
247  * only be truncated when necessary.
248  */
249 int wapbl_lazy_truncate = 0;
250 
251 struct wapbl_ops wapbl_ops = {
252 	.wo_wapbl_discard	= wapbl_discard,
253 	.wo_wapbl_replay_isopen	= wapbl_replay_isopen1,
254 	.wo_wapbl_replay_can_read = wapbl_replay_can_read,
255 	.wo_wapbl_replay_read	= wapbl_replay_read,
256 	.wo_wapbl_add_buf	= wapbl_add_buf,
257 	.wo_wapbl_remove_buf	= wapbl_remove_buf,
258 	.wo_wapbl_resize_buf	= wapbl_resize_buf,
259 	.wo_wapbl_begin		= wapbl_begin,
260 	.wo_wapbl_end		= wapbl_end,
261 	.wo_wapbl_junlock_assert= wapbl_junlock_assert,
262 
263 	/* XXX: the following is only used to say "this is a wapbl buf" */
264 	.wo_wapbl_biodone	= wapbl_biodone,
265 };
266 
267 static int
268 wapbl_sysctl_init(void)
269 {
270 	int rv;
271 	const struct sysctlnode *rnode, *cnode;
272 
273 	wapbl_sysctl = NULL;
274 
275 	rv = sysctl_createv(&wapbl_sysctl, 0, NULL, &rnode,
276 		       CTLFLAG_PERMANENT,
277 		       CTLTYPE_NODE, "vfs", NULL,
278 		       NULL, 0, NULL, 0,
279 		       CTL_VFS, CTL_EOL);
280 	if (rv)
281 		return rv;
282 
283 	rv = sysctl_createv(&wapbl_sysctl, 0, &rnode, &rnode,
284 		       CTLFLAG_PERMANENT,
285 		       CTLTYPE_NODE, "wapbl",
286 		       SYSCTL_DESCR("WAPBL journaling options"),
287 		       NULL, 0, NULL, 0,
288 		       CTL_CREATE, CTL_EOL);
289 	if (rv)
290 		return rv;
291 
292 	rv = sysctl_createv(&wapbl_sysctl, 0, &rnode, &cnode,
293 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
294 		       CTLTYPE_INT, "flush_disk_cache",
295 		       SYSCTL_DESCR("flush disk cache"),
296 		       NULL, 0, &wapbl_flush_disk_cache, 0,
297 		       CTL_CREATE, CTL_EOL);
298 	if (rv)
299 		return rv;
300 
301 	rv = sysctl_createv(&wapbl_sysctl, 0, &rnode, &cnode,
302 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
303 		       CTLTYPE_INT, "verbose_commit",
304 		       SYSCTL_DESCR("show time and size of wapbl log commits"),
305 		       NULL, 0, &wapbl_verbose_commit, 0,
306 		       CTL_CREATE, CTL_EOL);
307 	return rv;
308 }
309 
310 static void
311 wapbl_init(void)
312 {
313 	malloc_type_attach(M_WAPBL);
314 	wapbl_sysctl_init();
315 }
316 
317 #ifdef notyet
318 static int
319 wapbl_fini(bool interface)
320 {
321 	if (aio_sysctl != NULL)
322 		 sysctl_teardown(&aio_sysctl);
323 	return 0;
324 }
325 #endif
326 
327 static int
328 wapbl_start_flush_inodes(struct wapbl *wl, struct wapbl_replay *wr)
329 {
330 	int error, i;
331 
332 	WAPBL_PRINTF(WAPBL_PRINT_REPLAY,
333 	    ("wapbl_start: reusing log with %d inodes\n", wr->wr_inodescnt));
334 
335 	/*
336 	 * Its only valid to reuse the replay log if its
337 	 * the same as the new log we just opened.
338 	 */
339 	KDASSERT(!wapbl_replay_isopen(wr));
340 	KASSERT(wl->wl_devvp->v_type == VBLK);
341 	KASSERT(wr->wr_devvp->v_type == VBLK);
342 	KASSERT(wl->wl_devvp->v_rdev == wr->wr_devvp->v_rdev);
343 	KASSERT(wl->wl_logpbn == wr->wr_logpbn);
344 	KASSERT(wl->wl_circ_size == wr->wr_circ_size);
345 	KASSERT(wl->wl_circ_off == wr->wr_circ_off);
346 	KASSERT(wl->wl_log_dev_bshift == wr->wr_log_dev_bshift);
347 	KASSERT(wl->wl_fs_dev_bshift == wr->wr_fs_dev_bshift);
348 
349 	wl->wl_wc_header->wc_generation = wr->wr_generation + 1;
350 
351 	for (i = 0; i < wr->wr_inodescnt; i++)
352 		wapbl_register_inode(wl, wr->wr_inodes[i].wr_inumber,
353 		    wr->wr_inodes[i].wr_imode);
354 
355 	/* Make sure new transaction won't overwrite old inodes list */
356 	KDASSERT(wapbl_transaction_len(wl) <=
357 	    wapbl_space_free(wl->wl_circ_size, wr->wr_inodeshead,
358 	    wr->wr_inodestail));
359 
360 	wl->wl_head = wl->wl_tail = wr->wr_inodeshead;
361 	wl->wl_reclaimable_bytes = wl->wl_reserved_bytes =
362 	    wapbl_transaction_len(wl);
363 
364 	error = wapbl_write_inodes(wl, &wl->wl_head);
365 	if (error)
366 		return error;
367 
368 	KASSERT(wl->wl_head != wl->wl_tail);
369 	KASSERT(wl->wl_head != 0);
370 
371 	return 0;
372 }
373 
374 int
375 wapbl_start(struct wapbl ** wlp, struct mount *mp, struct vnode *vp,
376 	daddr_t off, size_t count, size_t blksize, struct wapbl_replay *wr,
377 	wapbl_flush_fn_t flushfn, wapbl_flush_fn_t flushabortfn)
378 {
379 	struct wapbl *wl;
380 	struct vnode *devvp;
381 	daddr_t logpbn;
382 	int error;
383 	int log_dev_bshift = ilog2(blksize);
384 	int fs_dev_bshift = log_dev_bshift;
385 	int run;
386 
387 	WAPBL_PRINTF(WAPBL_PRINT_OPEN, ("wapbl_start: vp=%p off=%" PRId64
388 	    " count=%zu blksize=%zu\n", vp, off, count, blksize));
389 
390 	if (log_dev_bshift > fs_dev_bshift) {
391 		WAPBL_PRINTF(WAPBL_PRINT_OPEN,
392 			("wapbl: log device's block size cannot be larger "
393 			 "than filesystem's\n"));
394 		/*
395 		 * Not currently implemented, although it could be if
396 		 * needed someday.
397 		 */
398 		return ENOSYS;
399 	}
400 
401 	if (off < 0)
402 		return EINVAL;
403 
404 	if (blksize < DEV_BSIZE)
405 		return EINVAL;
406 	if (blksize % DEV_BSIZE)
407 		return EINVAL;
408 
409 	/* XXXTODO: verify that the full load is writable */
410 
411 	/*
412 	 * XXX check for minimum log size
413 	 * minimum is governed by minimum amount of space
414 	 * to complete a transaction. (probably truncate)
415 	 */
416 	/* XXX for now pick something minimal */
417 	if ((count * blksize) < MAXPHYS) {
418 		return ENOSPC;
419 	}
420 
421 	if ((error = VOP_BMAP(vp, off, &devvp, &logpbn, &run)) != 0) {
422 		return error;
423 	}
424 
425 	wl = wapbl_calloc(1, sizeof(*wl));
426 	rw_init(&wl->wl_rwlock);
427 	mutex_init(&wl->wl_mtx, MUTEX_DEFAULT, IPL_NONE);
428 	cv_init(&wl->wl_reclaimable_cv, "wapblrec");
429 	LIST_INIT(&wl->wl_bufs);
430 	SIMPLEQ_INIT(&wl->wl_entries);
431 
432 	wl->wl_logvp = vp;
433 	wl->wl_devvp = devvp;
434 	wl->wl_mount = mp;
435 	wl->wl_logpbn = logpbn;
436 	wl->wl_log_dev_bshift = log_dev_bshift;
437 	wl->wl_fs_dev_bshift = fs_dev_bshift;
438 
439 	wl->wl_flush = flushfn;
440 	wl->wl_flush_abort = flushabortfn;
441 
442 	/* Reserve two log device blocks for the commit headers */
443 	wl->wl_circ_off = 2<<wl->wl_log_dev_bshift;
444 	wl->wl_circ_size = ((count * blksize) - wl->wl_circ_off);
445 	/* truncate the log usage to a multiple of log_dev_bshift */
446 	wl->wl_circ_size >>= wl->wl_log_dev_bshift;
447 	wl->wl_circ_size <<= wl->wl_log_dev_bshift;
448 
449 	/*
450 	 * wl_bufbytes_max limits the size of the in memory transaction space.
451 	 * - Since buffers are allocated and accounted for in units of
452 	 *   PAGE_SIZE it is required to be a multiple of PAGE_SIZE
453 	 *   (i.e. 1<<PAGE_SHIFT)
454 	 * - Since the log device has to be written in units of
455 	 *   1<<wl_log_dev_bshift it is required to be a mulitple of
456 	 *   1<<wl_log_dev_bshift.
457 	 * - Since filesystem will provide data in units of 1<<wl_fs_dev_bshift,
458 	 *   it is convenient to be a multiple of 1<<wl_fs_dev_bshift.
459 	 * Therefore it must be multiple of the least common multiple of those
460 	 * three quantities.  Fortunately, all of those quantities are
461 	 * guaranteed to be a power of two, and the least common multiple of
462 	 * a set of numbers which are all powers of two is simply the maximum
463 	 * of those numbers.  Finally, the maximum logarithm of a power of two
464 	 * is the same as the log of the maximum power of two.  So we can do
465 	 * the following operations to size wl_bufbytes_max:
466 	 */
467 
468 	/* XXX fix actual number of pages reserved per filesystem. */
469 	wl->wl_bufbytes_max = MIN(wl->wl_circ_size, buf_memcalc() / 2);
470 
471 	/* Round wl_bufbytes_max to the largest power of two constraint */
472 	wl->wl_bufbytes_max >>= PAGE_SHIFT;
473 	wl->wl_bufbytes_max <<= PAGE_SHIFT;
474 	wl->wl_bufbytes_max >>= wl->wl_log_dev_bshift;
475 	wl->wl_bufbytes_max <<= wl->wl_log_dev_bshift;
476 	wl->wl_bufbytes_max >>= wl->wl_fs_dev_bshift;
477 	wl->wl_bufbytes_max <<= wl->wl_fs_dev_bshift;
478 
479 	/* XXX maybe use filesystem fragment size instead of 1024 */
480 	/* XXX fix actual number of buffers reserved per filesystem. */
481 	wl->wl_bufcount_max = (nbuf / 2) * 1024;
482 
483 	/* XXX tie this into resource estimation */
484 	wl->wl_dealloclim = wl->wl_bufbytes_max / mp->mnt_stat.f_bsize / 2;
485 
486 	wl->wl_deallocblks = wapbl_malloc(sizeof(*wl->wl_deallocblks) *
487 	    wl->wl_dealloclim);
488 	wl->wl_dealloclens = wapbl_malloc(sizeof(*wl->wl_dealloclens) *
489 	    wl->wl_dealloclim);
490 
491 	wapbl_inodetrk_init(wl, WAPBL_INODETRK_SIZE);
492 
493 	/* Initialize the commit header */
494 	{
495 		struct wapbl_wc_header *wc;
496 		size_t len = 1 << wl->wl_log_dev_bshift;
497 		wc = wapbl_calloc(1, len);
498 		wc->wc_type = WAPBL_WC_HEADER;
499 		wc->wc_len = len;
500 		wc->wc_circ_off = wl->wl_circ_off;
501 		wc->wc_circ_size = wl->wl_circ_size;
502 		/* XXX wc->wc_fsid */
503 		wc->wc_log_dev_bshift = wl->wl_log_dev_bshift;
504 		wc->wc_fs_dev_bshift = wl->wl_fs_dev_bshift;
505 		wl->wl_wc_header = wc;
506 		wl->wl_wc_scratch = wapbl_malloc(len);
507 	}
508 
509 	/*
510 	 * if there was an existing set of unlinked but
511 	 * allocated inodes, preserve it in the new
512 	 * log.
513 	 */
514 	if (wr && wr->wr_inodescnt) {
515 		error = wapbl_start_flush_inodes(wl, wr);
516 		if (error)
517 			goto errout;
518 	}
519 
520 	error = wapbl_write_commit(wl, wl->wl_head, wl->wl_tail);
521 	if (error) {
522 		goto errout;
523 	}
524 
525 	*wlp = wl;
526 #if defined(WAPBL_DEBUG)
527 	wapbl_debug_wl = wl;
528 #endif
529 
530 	return 0;
531  errout:
532 	wapbl_discard(wl);
533 	wapbl_free(wl->wl_wc_scratch, wl->wl_wc_header->wc_len);
534 	wapbl_free(wl->wl_wc_header, wl->wl_wc_header->wc_len);
535 	wapbl_free(wl->wl_deallocblks,
536 	    sizeof(*wl->wl_deallocblks) * wl->wl_dealloclim);
537 	wapbl_free(wl->wl_dealloclens,
538 	    sizeof(*wl->wl_dealloclens) * wl->wl_dealloclim);
539 	wapbl_inodetrk_free(wl);
540 	wapbl_free(wl, sizeof(*wl));
541 
542 	return error;
543 }
544 
545 /*
546  * Like wapbl_flush, only discards the transaction
547  * completely
548  */
549 
550 void
551 wapbl_discard(struct wapbl *wl)
552 {
553 	struct wapbl_entry *we;
554 	struct buf *bp;
555 	int i;
556 
557 	/*
558 	 * XXX we may consider using upgrade here
559 	 * if we want to call flush from inside a transaction
560 	 */
561 	rw_enter(&wl->wl_rwlock, RW_WRITER);
562 	wl->wl_flush(wl->wl_mount, wl->wl_deallocblks, wl->wl_dealloclens,
563 	    wl->wl_dealloccnt);
564 
565 #ifdef WAPBL_DEBUG_PRINT
566 	{
567 		pid_t pid = -1;
568 		lwpid_t lid = -1;
569 		if (curproc)
570 			pid = curproc->p_pid;
571 		if (curlwp)
572 			lid = curlwp->l_lid;
573 #ifdef WAPBL_DEBUG_BUFBYTES
574 		WAPBL_PRINTF(WAPBL_PRINT_DISCARD,
575 		    ("wapbl_discard: thread %d.%d discarding "
576 		    "transaction\n"
577 		    "\tbufcount=%zu bufbytes=%zu bcount=%zu "
578 		    "deallocs=%d inodes=%d\n"
579 		    "\terrcnt = %u, reclaimable=%zu reserved=%zu "
580 		    "unsynced=%zu\n",
581 		    pid, lid, wl->wl_bufcount, wl->wl_bufbytes,
582 		    wl->wl_bcount, wl->wl_dealloccnt,
583 		    wl->wl_inohashcnt, wl->wl_error_count,
584 		    wl->wl_reclaimable_bytes, wl->wl_reserved_bytes,
585 		    wl->wl_unsynced_bufbytes));
586 		SIMPLEQ_FOREACH(we, &wl->wl_entries, we_entries) {
587 			WAPBL_PRINTF(WAPBL_PRINT_DISCARD,
588 			    ("\tentry: bufcount = %zu, reclaimable = %zu, "
589 			     "error = %d, unsynced = %zu\n",
590 			     we->we_bufcount, we->we_reclaimable_bytes,
591 			     we->we_error, we->we_unsynced_bufbytes));
592 		}
593 #else /* !WAPBL_DEBUG_BUFBYTES */
594 		WAPBL_PRINTF(WAPBL_PRINT_DISCARD,
595 		    ("wapbl_discard: thread %d.%d discarding transaction\n"
596 		    "\tbufcount=%zu bufbytes=%zu bcount=%zu "
597 		    "deallocs=%d inodes=%d\n"
598 		    "\terrcnt = %u, reclaimable=%zu reserved=%zu\n",
599 		    pid, lid, wl->wl_bufcount, wl->wl_bufbytes,
600 		    wl->wl_bcount, wl->wl_dealloccnt,
601 		    wl->wl_inohashcnt, wl->wl_error_count,
602 		    wl->wl_reclaimable_bytes, wl->wl_reserved_bytes));
603 		SIMPLEQ_FOREACH(we, &wl->wl_entries, we_entries) {
604 			WAPBL_PRINTF(WAPBL_PRINT_DISCARD,
605 			    ("\tentry: bufcount = %zu, reclaimable = %zu, "
606 			     "error = %d\n",
607 			     we->we_bufcount, we->we_reclaimable_bytes,
608 			     we->we_error));
609 		}
610 #endif /* !WAPBL_DEBUG_BUFBYTES */
611 	}
612 #endif /* WAPBL_DEBUG_PRINT */
613 
614 	for (i = 0; i <= wl->wl_inohashmask; i++) {
615 		struct wapbl_ino_head *wih;
616 		struct wapbl_ino *wi;
617 
618 		wih = &wl->wl_inohash[i];
619 		while ((wi = LIST_FIRST(wih)) != NULL) {
620 			LIST_REMOVE(wi, wi_hash);
621 			pool_put(&wapbl_ino_pool, wi);
622 			KASSERT(wl->wl_inohashcnt > 0);
623 			wl->wl_inohashcnt--;
624 		}
625 	}
626 
627 	/*
628 	 * clean buffer list
629 	 */
630 	mutex_enter(&bufcache_lock);
631 	mutex_enter(&wl->wl_mtx);
632 	while ((bp = LIST_FIRST(&wl->wl_bufs)) != NULL) {
633 		if (bbusy(bp, 0, 0, &wl->wl_mtx) == 0) {
634 			/*
635 			 * The buffer will be unlocked and
636 			 * removed from the transaction in brelse
637 			 */
638 			mutex_exit(&wl->wl_mtx);
639 			brelsel(bp, 0);
640 			mutex_enter(&wl->wl_mtx);
641 		}
642 	}
643 	mutex_exit(&wl->wl_mtx);
644 	mutex_exit(&bufcache_lock);
645 
646 	/*
647 	 * Remove references to this wl from wl_entries, free any which
648 	 * no longer have buffers, others will be freed in wapbl_biodone
649 	 * when they no longer have any buffers.
650 	 */
651 	while ((we = SIMPLEQ_FIRST(&wl->wl_entries)) != NULL) {
652 		SIMPLEQ_REMOVE_HEAD(&wl->wl_entries, we_entries);
653 		/* XXX should we be accumulating wl_error_count
654 		 * and increasing reclaimable bytes ? */
655 		we->we_wapbl = NULL;
656 		if (we->we_bufcount == 0) {
657 #ifdef WAPBL_DEBUG_BUFBYTES
658 			KASSERT(we->we_unsynced_bufbytes == 0);
659 #endif
660 			wapbl_free(we, sizeof(*we));
661 		}
662 	}
663 
664 	/* Discard list of deallocs */
665 	wl->wl_dealloccnt = 0;
666 	/* XXX should we clear wl_reserved_bytes? */
667 
668 	KASSERT(wl->wl_bufbytes == 0);
669 	KASSERT(wl->wl_bcount == 0);
670 	KASSERT(wl->wl_bufcount == 0);
671 	KASSERT(LIST_EMPTY(&wl->wl_bufs));
672 	KASSERT(SIMPLEQ_EMPTY(&wl->wl_entries));
673 	KASSERT(wl->wl_inohashcnt == 0);
674 
675 	rw_exit(&wl->wl_rwlock);
676 }
677 
678 int
679 wapbl_stop(struct wapbl *wl, int force)
680 {
681 	struct vnode *vp;
682 	int error;
683 
684 	WAPBL_PRINTF(WAPBL_PRINT_OPEN, ("wapbl_stop called\n"));
685 	error = wapbl_flush(wl, 1);
686 	if (error) {
687 		if (force)
688 			wapbl_discard(wl);
689 		else
690 			return error;
691 	}
692 
693 	/* Unlinked inodes persist after a flush */
694 	if (wl->wl_inohashcnt) {
695 		if (force) {
696 			wapbl_discard(wl);
697 		} else {
698 			return EBUSY;
699 		}
700 	}
701 
702 	KASSERT(wl->wl_bufbytes == 0);
703 	KASSERT(wl->wl_bcount == 0);
704 	KASSERT(wl->wl_bufcount == 0);
705 	KASSERT(LIST_EMPTY(&wl->wl_bufs));
706 	KASSERT(wl->wl_dealloccnt == 0);
707 	KASSERT(SIMPLEQ_EMPTY(&wl->wl_entries));
708 	KASSERT(wl->wl_inohashcnt == 0);
709 
710 	vp = wl->wl_logvp;
711 
712 	wapbl_free(wl->wl_wc_scratch, wl->wl_wc_header->wc_len);
713 	wapbl_free(wl->wl_wc_header, wl->wl_wc_header->wc_len);
714 	wapbl_free(wl->wl_deallocblks,
715 	    sizeof(*wl->wl_deallocblks) * wl->wl_dealloclim);
716 	wapbl_free(wl->wl_dealloclens,
717 	    sizeof(*wl->wl_dealloclens) * wl->wl_dealloclim);
718 	wapbl_inodetrk_free(wl);
719 
720 	cv_destroy(&wl->wl_reclaimable_cv);
721 	mutex_destroy(&wl->wl_mtx);
722 	rw_destroy(&wl->wl_rwlock);
723 	wapbl_free(wl, sizeof(*wl));
724 
725 	return 0;
726 }
727 
728 static int
729 wapbl_doio(void *data, size_t len, struct vnode *devvp, daddr_t pbn, int flags)
730 {
731 	struct pstats *pstats = curlwp->l_proc->p_stats;
732 	struct buf *bp;
733 	int error;
734 
735 	KASSERT((flags & ~(B_WRITE | B_READ)) == 0);
736 	KASSERT(devvp->v_type == VBLK);
737 
738 	if ((flags & (B_WRITE | B_READ)) == B_WRITE) {
739 		mutex_enter(devvp->v_interlock);
740 		devvp->v_numoutput++;
741 		mutex_exit(devvp->v_interlock);
742 		pstats->p_ru.ru_oublock++;
743 	} else {
744 		pstats->p_ru.ru_inblock++;
745 	}
746 
747 	bp = getiobuf(devvp, true);
748 	bp->b_flags = flags;
749 	bp->b_cflags = BC_BUSY; /* silly & dubious */
750 	bp->b_dev = devvp->v_rdev;
751 	bp->b_data = data;
752 	bp->b_bufsize = bp->b_resid = bp->b_bcount = len;
753 	bp->b_blkno = pbn;
754 
755 	WAPBL_PRINTF(WAPBL_PRINT_IO,
756 	    ("wapbl_doio: %s %d bytes at block %"PRId64" on dev 0x%"PRIx64"\n",
757 	    BUF_ISWRITE(bp) ? "write" : "read", bp->b_bcount,
758 	    bp->b_blkno, bp->b_dev));
759 
760 	VOP_STRATEGY(devvp, bp);
761 
762 	error = biowait(bp);
763 	putiobuf(bp);
764 
765 	if (error) {
766 		WAPBL_PRINTF(WAPBL_PRINT_ERROR,
767 		    ("wapbl_doio: %s %zu bytes at block %" PRId64
768 		    " on dev 0x%"PRIx64" failed with error %d\n",
769 		    (((flags & (B_WRITE | B_READ)) == B_WRITE) ?
770 		     "write" : "read"),
771 		    len, pbn, devvp->v_rdev, error));
772 	}
773 
774 	return error;
775 }
776 
777 int
778 wapbl_write(void *data, size_t len, struct vnode *devvp, daddr_t pbn)
779 {
780 
781 	return wapbl_doio(data, len, devvp, pbn, B_WRITE);
782 }
783 
784 int
785 wapbl_read(void *data, size_t len, struct vnode *devvp, daddr_t pbn)
786 {
787 
788 	return wapbl_doio(data, len, devvp, pbn, B_READ);
789 }
790 
791 /*
792  * Off is byte offset returns new offset for next write
793  * handles log wraparound
794  */
795 static int
796 wapbl_circ_write(struct wapbl *wl, void *data, size_t len, off_t *offp)
797 {
798 	size_t slen;
799 	off_t off = *offp;
800 	int error;
801 	daddr_t pbn;
802 
803 	KDASSERT(((len >> wl->wl_log_dev_bshift) <<
804 	    wl->wl_log_dev_bshift) == len);
805 
806 	if (off < wl->wl_circ_off)
807 		off = wl->wl_circ_off;
808 	slen = wl->wl_circ_off + wl->wl_circ_size - off;
809 	if (slen < len) {
810 		pbn = wl->wl_logpbn + (off >> wl->wl_log_dev_bshift);
811 #ifdef _KERNEL
812 		pbn = btodb(pbn << wl->wl_log_dev_bshift);
813 #endif
814 		error = wapbl_write(data, slen, wl->wl_devvp, pbn);
815 		if (error)
816 			return error;
817 		data = (uint8_t *)data + slen;
818 		len -= slen;
819 		off = wl->wl_circ_off;
820 	}
821 	pbn = wl->wl_logpbn + (off >> wl->wl_log_dev_bshift);
822 #ifdef _KERNEL
823 	pbn = btodb(pbn << wl->wl_log_dev_bshift);
824 #endif
825 	error = wapbl_write(data, len, wl->wl_devvp, pbn);
826 	if (error)
827 		return error;
828 	off += len;
829 	if (off >= wl->wl_circ_off + wl->wl_circ_size)
830 		off = wl->wl_circ_off;
831 	*offp = off;
832 	return 0;
833 }
834 
835 /****************************************************************/
836 
837 int
838 wapbl_begin(struct wapbl *wl, const char *file, int line)
839 {
840 	int doflush;
841 	unsigned lockcount;
842 
843 	KDASSERT(wl);
844 
845 	/*
846 	 * XXX this needs to be made much more sophisticated.
847 	 * perhaps each wapbl_begin could reserve a specified
848 	 * number of buffers and bytes.
849 	 */
850 	mutex_enter(&wl->wl_mtx);
851 	lockcount = wl->wl_lock_count;
852 	doflush = ((wl->wl_bufbytes + (lockcount * MAXPHYS)) >
853 		   wl->wl_bufbytes_max / 2) ||
854 		  ((wl->wl_bufcount + (lockcount * 10)) >
855 		   wl->wl_bufcount_max / 2) ||
856 		  (wapbl_transaction_len(wl) > wl->wl_circ_size / 2) ||
857 		  (wl->wl_dealloccnt >= (wl->wl_dealloclim / 2));
858 	mutex_exit(&wl->wl_mtx);
859 
860 	if (doflush) {
861 		WAPBL_PRINTF(WAPBL_PRINT_FLUSH,
862 		    ("force flush lockcnt=%d bufbytes=%zu "
863 		    "(max=%zu) bufcount=%zu (max=%zu) "
864 		    "dealloccnt %d (lim=%d)\n",
865 		    lockcount, wl->wl_bufbytes,
866 		    wl->wl_bufbytes_max, wl->wl_bufcount,
867 		    wl->wl_bufcount_max,
868 		    wl->wl_dealloccnt, wl->wl_dealloclim));
869 	}
870 
871 	if (doflush) {
872 		int error = wapbl_flush(wl, 0);
873 		if (error)
874 			return error;
875 	}
876 
877 	rw_enter(&wl->wl_rwlock, RW_READER);
878 	mutex_enter(&wl->wl_mtx);
879 	wl->wl_lock_count++;
880 	mutex_exit(&wl->wl_mtx);
881 
882 #if defined(WAPBL_DEBUG_PRINT)
883 	WAPBL_PRINTF(WAPBL_PRINT_TRANSACTION,
884 	    ("wapbl_begin thread %d.%d with bufcount=%zu "
885 	    "bufbytes=%zu bcount=%zu at %s:%d\n",
886 	    curproc->p_pid, curlwp->l_lid, wl->wl_bufcount,
887 	    wl->wl_bufbytes, wl->wl_bcount, file, line));
888 #endif
889 
890 	return 0;
891 }
892 
893 void
894 wapbl_end(struct wapbl *wl)
895 {
896 
897 #if defined(WAPBL_DEBUG_PRINT)
898 	WAPBL_PRINTF(WAPBL_PRINT_TRANSACTION,
899 	     ("wapbl_end thread %d.%d with bufcount=%zu "
900 	      "bufbytes=%zu bcount=%zu\n",
901 	      curproc->p_pid, curlwp->l_lid, wl->wl_bufcount,
902 	      wl->wl_bufbytes, wl->wl_bcount));
903 #endif
904 
905 #ifdef DIAGNOSTIC
906 	size_t flushsize = wapbl_transaction_len(wl);
907 	if (flushsize > (wl->wl_circ_size - wl->wl_reserved_bytes)) {
908 		/*
909 		 * XXX this could be handled more gracefully, perhaps place
910 		 * only a partial transaction in the log and allow the
911 		 * remaining to flush without the protection of the journal.
912 		 */
913 		panic("wapbl_end: current transaction too big to flush\n");
914 	}
915 #endif
916 
917 	mutex_enter(&wl->wl_mtx);
918 	KASSERT(wl->wl_lock_count > 0);
919 	wl->wl_lock_count--;
920 	mutex_exit(&wl->wl_mtx);
921 
922 	rw_exit(&wl->wl_rwlock);
923 }
924 
925 void
926 wapbl_add_buf(struct wapbl *wl, struct buf * bp)
927 {
928 
929 	KASSERT(bp->b_cflags & BC_BUSY);
930 	KASSERT(bp->b_vp);
931 
932 	wapbl_jlock_assert(wl);
933 
934 #if 0
935 	/*
936 	 * XXX this might be an issue for swapfiles.
937 	 * see uvm_swap.c:1702
938 	 *
939 	 * XXX2 why require it then?  leap of semantics?
940 	 */
941 	KASSERT((bp->b_cflags & BC_NOCACHE) == 0);
942 #endif
943 
944 	mutex_enter(&wl->wl_mtx);
945 	if (bp->b_flags & B_LOCKED) {
946 		LIST_REMOVE(bp, b_wapbllist);
947 		WAPBL_PRINTF(WAPBL_PRINT_BUFFER2,
948 		   ("wapbl_add_buf thread %d.%d re-adding buf %p "
949 		    "with %d bytes %d bcount\n",
950 		    curproc->p_pid, curlwp->l_lid, bp, bp->b_bufsize,
951 		    bp->b_bcount));
952 	} else {
953 		/* unlocked by dirty buffers shouldn't exist */
954 		KASSERT(!(bp->b_oflags & BO_DELWRI));
955 		wl->wl_bufbytes += bp->b_bufsize;
956 		wl->wl_bcount += bp->b_bcount;
957 		wl->wl_bufcount++;
958 		WAPBL_PRINTF(WAPBL_PRINT_BUFFER,
959 		   ("wapbl_add_buf thread %d.%d adding buf %p "
960 		    "with %d bytes %d bcount\n",
961 		    curproc->p_pid, curlwp->l_lid, bp, bp->b_bufsize,
962 		    bp->b_bcount));
963 	}
964 	LIST_INSERT_HEAD(&wl->wl_bufs, bp, b_wapbllist);
965 	mutex_exit(&wl->wl_mtx);
966 
967 	bp->b_flags |= B_LOCKED;
968 }
969 
970 static void
971 wapbl_remove_buf_locked(struct wapbl * wl, struct buf *bp)
972 {
973 
974 	KASSERT(mutex_owned(&wl->wl_mtx));
975 	KASSERT(bp->b_cflags & BC_BUSY);
976 	wapbl_jlock_assert(wl);
977 
978 #if 0
979 	/*
980 	 * XXX this might be an issue for swapfiles.
981 	 * see uvm_swap.c:1725
982 	 *
983 	 * XXXdeux: see above
984 	 */
985 	KASSERT((bp->b_flags & BC_NOCACHE) == 0);
986 #endif
987 	KASSERT(bp->b_flags & B_LOCKED);
988 
989 	WAPBL_PRINTF(WAPBL_PRINT_BUFFER,
990 	   ("wapbl_remove_buf thread %d.%d removing buf %p with "
991 	    "%d bytes %d bcount\n",
992 	    curproc->p_pid, curlwp->l_lid, bp, bp->b_bufsize, bp->b_bcount));
993 
994 	KASSERT(wl->wl_bufbytes >= bp->b_bufsize);
995 	wl->wl_bufbytes -= bp->b_bufsize;
996 	KASSERT(wl->wl_bcount >= bp->b_bcount);
997 	wl->wl_bcount -= bp->b_bcount;
998 	KASSERT(wl->wl_bufcount > 0);
999 	wl->wl_bufcount--;
1000 	KASSERT((wl->wl_bufcount == 0) == (wl->wl_bufbytes == 0));
1001 	KASSERT((wl->wl_bufcount == 0) == (wl->wl_bcount == 0));
1002 	LIST_REMOVE(bp, b_wapbllist);
1003 
1004 	bp->b_flags &= ~B_LOCKED;
1005 }
1006 
1007 /* called from brelsel() in vfs_bio among other places */
1008 void
1009 wapbl_remove_buf(struct wapbl * wl, struct buf *bp)
1010 {
1011 
1012 	mutex_enter(&wl->wl_mtx);
1013 	wapbl_remove_buf_locked(wl, bp);
1014 	mutex_exit(&wl->wl_mtx);
1015 }
1016 
1017 void
1018 wapbl_resize_buf(struct wapbl *wl, struct buf *bp, long oldsz, long oldcnt)
1019 {
1020 
1021 	KASSERT(bp->b_cflags & BC_BUSY);
1022 
1023 	/*
1024 	 * XXX: why does this depend on B_LOCKED?  otherwise the buf
1025 	 * is not for a transaction?  if so, why is this called in the
1026 	 * first place?
1027 	 */
1028 	if (bp->b_flags & B_LOCKED) {
1029 		mutex_enter(&wl->wl_mtx);
1030 		wl->wl_bufbytes += bp->b_bufsize - oldsz;
1031 		wl->wl_bcount += bp->b_bcount - oldcnt;
1032 		mutex_exit(&wl->wl_mtx);
1033 	}
1034 }
1035 
1036 #endif /* _KERNEL */
1037 
1038 /****************************************************************/
1039 /* Some utility inlines */
1040 
1041 /* This is used to advance the pointer at old to new value at old+delta */
1042 static inline off_t
1043 wapbl_advance(size_t size, size_t off, off_t old, size_t delta)
1044 {
1045 	off_t new;
1046 
1047 	/* Define acceptable ranges for inputs. */
1048 	KASSERT(delta <= (size_t)size);
1049 	KASSERT((old == 0) || ((size_t)old >= off));
1050 	KASSERT(old < (off_t)(size + off));
1051 
1052 	if ((old == 0) && (delta != 0))
1053 		new = off + delta;
1054 	else if ((old + delta) < (size + off))
1055 		new = old + delta;
1056 	else
1057 		new = (old + delta) - size;
1058 
1059 	/* Note some interesting axioms */
1060 	KASSERT((delta != 0) || (new == old));
1061 	KASSERT((delta == 0) || (new != 0));
1062 	KASSERT((delta != (size)) || (new == old));
1063 
1064 	/* Define acceptable ranges for output. */
1065 	KASSERT((new == 0) || ((size_t)new >= off));
1066 	KASSERT((size_t)new < (size + off));
1067 	return new;
1068 }
1069 
1070 static inline size_t
1071 wapbl_space_used(size_t avail, off_t head, off_t tail)
1072 {
1073 
1074 	if (tail == 0) {
1075 		KASSERT(head == 0);
1076 		return 0;
1077 	}
1078 	return ((head + (avail - 1) - tail) % avail) + 1;
1079 }
1080 
1081 static inline size_t
1082 wapbl_space_free(size_t avail, off_t head, off_t tail)
1083 {
1084 
1085 	return avail - wapbl_space_used(avail, head, tail);
1086 }
1087 
1088 static inline void
1089 wapbl_advance_head(size_t size, size_t off, size_t delta, off_t *headp,
1090 		   off_t *tailp)
1091 {
1092 	off_t head = *headp;
1093 	off_t tail = *tailp;
1094 
1095 	KASSERT(delta <= wapbl_space_free(size, head, tail));
1096 	head = wapbl_advance(size, off, head, delta);
1097 	if ((tail == 0) && (head != 0))
1098 		tail = off;
1099 	*headp = head;
1100 	*tailp = tail;
1101 }
1102 
1103 static inline void
1104 wapbl_advance_tail(size_t size, size_t off, size_t delta, off_t *headp,
1105 		   off_t *tailp)
1106 {
1107 	off_t head = *headp;
1108 	off_t tail = *tailp;
1109 
1110 	KASSERT(delta <= wapbl_space_used(size, head, tail));
1111 	tail = wapbl_advance(size, off, tail, delta);
1112 	if (head == tail) {
1113 		head = tail = 0;
1114 	}
1115 	*headp = head;
1116 	*tailp = tail;
1117 }
1118 
1119 #ifdef _KERNEL
1120 
1121 /****************************************************************/
1122 
1123 /*
1124  * Remove transactions whose buffers are completely flushed to disk.
1125  * Will block until at least minfree space is available.
1126  * only intended to be called from inside wapbl_flush and therefore
1127  * does not protect against commit races with itself or with flush.
1128  */
1129 static int
1130 wapbl_truncate(struct wapbl *wl, size_t minfree, int waitonly)
1131 {
1132 	size_t delta;
1133 	size_t avail;
1134 	off_t head;
1135 	off_t tail;
1136 	int error = 0;
1137 
1138 	KASSERT(minfree <= (wl->wl_circ_size - wl->wl_reserved_bytes));
1139 	KASSERT(rw_write_held(&wl->wl_rwlock));
1140 
1141 	mutex_enter(&wl->wl_mtx);
1142 
1143 	/*
1144 	 * First check to see if we have to do a commit
1145 	 * at all.
1146 	 */
1147 	avail = wapbl_space_free(wl->wl_circ_size, wl->wl_head, wl->wl_tail);
1148 	if (minfree < avail) {
1149 		mutex_exit(&wl->wl_mtx);
1150 		return 0;
1151 	}
1152 	minfree -= avail;
1153 	while ((wl->wl_error_count == 0) &&
1154 	    (wl->wl_reclaimable_bytes < minfree)) {
1155         	WAPBL_PRINTF(WAPBL_PRINT_TRUNCATE,
1156                    ("wapbl_truncate: sleeping on %p wl=%p bytes=%zd "
1157 		    "minfree=%zd\n",
1158                     &wl->wl_reclaimable_bytes, wl, wl->wl_reclaimable_bytes,
1159 		    minfree));
1160 
1161 		cv_wait(&wl->wl_reclaimable_cv, &wl->wl_mtx);
1162 	}
1163 	if (wl->wl_reclaimable_bytes < minfree) {
1164 		KASSERT(wl->wl_error_count);
1165 		/* XXX maybe get actual error from buffer instead someday? */
1166 		error = EIO;
1167 	}
1168 	head = wl->wl_head;
1169 	tail = wl->wl_tail;
1170 	delta = wl->wl_reclaimable_bytes;
1171 
1172 	/* If all of of the entries are flushed, then be sure to keep
1173 	 * the reserved bytes reserved.  Watch out for discarded transactions,
1174 	 * which could leave more bytes reserved than are reclaimable.
1175 	 */
1176 	if (SIMPLEQ_EMPTY(&wl->wl_entries) &&
1177 	    (delta >= wl->wl_reserved_bytes)) {
1178 		delta -= wl->wl_reserved_bytes;
1179 	}
1180 	wapbl_advance_tail(wl->wl_circ_size, wl->wl_circ_off, delta, &head,
1181 			   &tail);
1182 	KDASSERT(wl->wl_reserved_bytes <=
1183 		wapbl_space_used(wl->wl_circ_size, head, tail));
1184 	mutex_exit(&wl->wl_mtx);
1185 
1186 	if (error)
1187 		return error;
1188 
1189 	if (waitonly)
1190 		return 0;
1191 
1192 	/*
1193 	 * This is where head, tail and delta are unprotected
1194 	 * from races against itself or flush.  This is ok since
1195 	 * we only call this routine from inside flush itself.
1196 	 *
1197 	 * XXX: how can it race against itself when accessed only
1198 	 * from behind the write-locked rwlock?
1199 	 */
1200 	error = wapbl_write_commit(wl, head, tail);
1201 	if (error)
1202 		return error;
1203 
1204 	wl->wl_head = head;
1205 	wl->wl_tail = tail;
1206 
1207 	mutex_enter(&wl->wl_mtx);
1208 	KASSERT(wl->wl_reclaimable_bytes >= delta);
1209 	wl->wl_reclaimable_bytes -= delta;
1210 	mutex_exit(&wl->wl_mtx);
1211 	WAPBL_PRINTF(WAPBL_PRINT_TRUNCATE,
1212 	    ("wapbl_truncate thread %d.%d truncating %zu bytes\n",
1213 	    curproc->p_pid, curlwp->l_lid, delta));
1214 
1215 	return 0;
1216 }
1217 
1218 /****************************************************************/
1219 
1220 void
1221 wapbl_biodone(struct buf *bp)
1222 {
1223 	struct wapbl_entry *we = bp->b_private;
1224 	struct wapbl *wl = we->we_wapbl;
1225 
1226 	/*
1227 	 * Handle possible flushing of buffers after log has been
1228 	 * decomissioned.
1229 	 */
1230 	if (!wl) {
1231 		KASSERT(we->we_bufcount > 0);
1232 		we->we_bufcount--;
1233 #ifdef WAPBL_DEBUG_BUFBYTES
1234 		KASSERT(we->we_unsynced_bufbytes >= bp->b_bufsize);
1235 		we->we_unsynced_bufbytes -= bp->b_bufsize;
1236 #endif
1237 
1238 		if (we->we_bufcount == 0) {
1239 #ifdef WAPBL_DEBUG_BUFBYTES
1240 			KASSERT(we->we_unsynced_bufbytes == 0);
1241 #endif
1242 			wapbl_free(we, sizeof(*we));
1243 		}
1244 
1245 		brelse(bp, 0);
1246 		return;
1247 	}
1248 
1249 #ifdef ohbother
1250 	KDASSERT(bp->b_oflags & BO_DONE);
1251 	KDASSERT(!(bp->b_oflags & BO_DELWRI));
1252 	KDASSERT(bp->b_flags & B_ASYNC);
1253 	KDASSERT(bp->b_cflags & BC_BUSY);
1254 	KDASSERT(!(bp->b_flags & B_LOCKED));
1255 	KDASSERT(!(bp->b_flags & B_READ));
1256 	KDASSERT(!(bp->b_cflags & BC_INVAL));
1257 	KDASSERT(!(bp->b_cflags & BC_NOCACHE));
1258 #endif
1259 
1260 	if (bp->b_error) {
1261 #ifdef notyet /* Can't currently handle possible dirty buffer reuse */
1262 		/*
1263 		 * XXXpooka: interfaces not fully updated
1264 		 * Note: this was not enabled in the original patch
1265 		 * against netbsd4 either.  I don't know if comment
1266 		 * above is true or not.
1267 		 */
1268 
1269 		/*
1270 		 * If an error occurs, report the error and leave the
1271 		 * buffer as a delayed write on the LRU queue.
1272 		 * restarting the write would likely result in
1273 		 * an error spinloop, so let it be done harmlessly
1274 		 * by the syncer.
1275 		 */
1276 		bp->b_flags &= ~(B_DONE);
1277 		simple_unlock(&bp->b_interlock);
1278 
1279 		if (we->we_error == 0) {
1280 			mutex_enter(&wl->wl_mtx);
1281 			wl->wl_error_count++;
1282 			mutex_exit(&wl->wl_mtx);
1283 			cv_broadcast(&wl->wl_reclaimable_cv);
1284 		}
1285 		we->we_error = bp->b_error;
1286 		bp->b_error = 0;
1287 		brelse(bp);
1288 		return;
1289 #else
1290 		/* For now, just mark the log permanently errored out */
1291 
1292 		mutex_enter(&wl->wl_mtx);
1293 		if (wl->wl_error_count == 0) {
1294 			wl->wl_error_count++;
1295 			cv_broadcast(&wl->wl_reclaimable_cv);
1296 		}
1297 		mutex_exit(&wl->wl_mtx);
1298 #endif
1299 	}
1300 
1301 	mutex_enter(&wl->wl_mtx);
1302 
1303 	KASSERT(we->we_bufcount > 0);
1304 	we->we_bufcount--;
1305 #ifdef WAPBL_DEBUG_BUFBYTES
1306 	KASSERT(we->we_unsynced_bufbytes >= bp->b_bufsize);
1307 	we->we_unsynced_bufbytes -= bp->b_bufsize;
1308 	KASSERT(wl->wl_unsynced_bufbytes >= bp->b_bufsize);
1309 	wl->wl_unsynced_bufbytes -= bp->b_bufsize;
1310 #endif
1311 
1312 	/*
1313 	 * If the current transaction can be reclaimed, start
1314 	 * at the beginning and reclaim any consecutive reclaimable
1315 	 * transactions.  If we successfully reclaim anything,
1316 	 * then wakeup anyone waiting for the reclaim.
1317 	 */
1318 	if (we->we_bufcount == 0) {
1319 		size_t delta = 0;
1320 		int errcnt = 0;
1321 #ifdef WAPBL_DEBUG_BUFBYTES
1322 		KDASSERT(we->we_unsynced_bufbytes == 0);
1323 #endif
1324 		/*
1325 		 * clear any posted error, since the buffer it came from
1326 		 * has successfully flushed by now
1327 		 */
1328 		while ((we = SIMPLEQ_FIRST(&wl->wl_entries)) &&
1329 		       (we->we_bufcount == 0)) {
1330 			delta += we->we_reclaimable_bytes;
1331 			if (we->we_error)
1332 				errcnt++;
1333 			SIMPLEQ_REMOVE_HEAD(&wl->wl_entries, we_entries);
1334 			wapbl_free(we, sizeof(*we));
1335 		}
1336 
1337 		if (delta) {
1338 			wl->wl_reclaimable_bytes += delta;
1339 			KASSERT(wl->wl_error_count >= errcnt);
1340 			wl->wl_error_count -= errcnt;
1341 			cv_broadcast(&wl->wl_reclaimable_cv);
1342 		}
1343 	}
1344 
1345 	mutex_exit(&wl->wl_mtx);
1346 	brelse(bp, 0);
1347 }
1348 
1349 /*
1350  * Write transactions to disk + start I/O for contents
1351  */
1352 int
1353 wapbl_flush(struct wapbl *wl, int waitfor)
1354 {
1355 	struct buf *bp;
1356 	struct wapbl_entry *we;
1357 	off_t off;
1358 	off_t head;
1359 	off_t tail;
1360 	size_t delta = 0;
1361 	size_t flushsize;
1362 	size_t reserved;
1363 	int error = 0;
1364 
1365 	/*
1366 	 * Do a quick check to see if a full flush can be skipped
1367 	 * This assumes that the flush callback does not need to be called
1368 	 * unless there are other outstanding bufs.
1369 	 */
1370 	if (!waitfor) {
1371 		size_t nbufs;
1372 		mutex_enter(&wl->wl_mtx);	/* XXX need mutex here to
1373 						   protect the KASSERTS */
1374 		nbufs = wl->wl_bufcount;
1375 		KASSERT((wl->wl_bufcount == 0) == (wl->wl_bufbytes == 0));
1376 		KASSERT((wl->wl_bufcount == 0) == (wl->wl_bcount == 0));
1377 		mutex_exit(&wl->wl_mtx);
1378 		if (nbufs == 0)
1379 			return 0;
1380 	}
1381 
1382 	/*
1383 	 * XXX we may consider using LK_UPGRADE here
1384 	 * if we want to call flush from inside a transaction
1385 	 */
1386 	rw_enter(&wl->wl_rwlock, RW_WRITER);
1387 	wl->wl_flush(wl->wl_mount, wl->wl_deallocblks, wl->wl_dealloclens,
1388 	    wl->wl_dealloccnt);
1389 
1390 	/*
1391 	 * Now that we are fully locked and flushed,
1392 	 * do another check for nothing to do.
1393 	 */
1394 	if (wl->wl_bufcount == 0) {
1395 		goto out;
1396 	}
1397 
1398 #if 0
1399 	WAPBL_PRINTF(WAPBL_PRINT_FLUSH,
1400 		     ("wapbl_flush thread %d.%d flushing entries with "
1401 		      "bufcount=%zu bufbytes=%zu\n",
1402 		      curproc->p_pid, curlwp->l_lid, wl->wl_bufcount,
1403 		      wl->wl_bufbytes));
1404 #endif
1405 
1406 	/* Calculate amount of space needed to flush */
1407 	flushsize = wapbl_transaction_len(wl);
1408 	if (wapbl_verbose_commit) {
1409 		struct timespec ts;
1410 		getnanotime(&ts);
1411 		printf("%s: %lld.%09ld this transaction = %zu bytes\n",
1412 		    __func__, (long long)ts.tv_sec,
1413 		    (long)ts.tv_nsec, flushsize);
1414 	}
1415 
1416 	if (flushsize > (wl->wl_circ_size - wl->wl_reserved_bytes)) {
1417 		/*
1418 		 * XXX this could be handled more gracefully, perhaps place
1419 		 * only a partial transaction in the log and allow the
1420 		 * remaining to flush without the protection of the journal.
1421 		 */
1422 		panic("wapbl_flush: current transaction too big to flush\n");
1423 	}
1424 
1425 	error = wapbl_truncate(wl, flushsize, 0);
1426 	if (error)
1427 		goto out2;
1428 
1429 	off = wl->wl_head;
1430 	KASSERT((off == 0) || ((off >= wl->wl_circ_off) &&
1431 	                      (off < wl->wl_circ_off + wl->wl_circ_size)));
1432 	error = wapbl_write_blocks(wl, &off);
1433 	if (error)
1434 		goto out2;
1435 	error = wapbl_write_revocations(wl, &off);
1436 	if (error)
1437 		goto out2;
1438 	error = wapbl_write_inodes(wl, &off);
1439 	if (error)
1440 		goto out2;
1441 
1442 	reserved = 0;
1443 	if (wl->wl_inohashcnt)
1444 		reserved = wapbl_transaction_inodes_len(wl);
1445 
1446 	head = wl->wl_head;
1447 	tail = wl->wl_tail;
1448 
1449 	wapbl_advance_head(wl->wl_circ_size, wl->wl_circ_off, flushsize,
1450 	    &head, &tail);
1451 #ifdef WAPBL_DEBUG
1452 	if (head != off) {
1453 		panic("lost head! head=%"PRIdMAX" tail=%" PRIdMAX
1454 		      " off=%"PRIdMAX" flush=%zu\n",
1455 		      (intmax_t)head, (intmax_t)tail, (intmax_t)off,
1456 		      flushsize);
1457 	}
1458 #else
1459 	KASSERT(head == off);
1460 #endif
1461 
1462 	/* Opportunistically move the tail forward if we can */
1463 	if (!wapbl_lazy_truncate) {
1464 		mutex_enter(&wl->wl_mtx);
1465 		delta = wl->wl_reclaimable_bytes;
1466 		mutex_exit(&wl->wl_mtx);
1467 		wapbl_advance_tail(wl->wl_circ_size, wl->wl_circ_off, delta,
1468 		    &head, &tail);
1469 	}
1470 
1471 	error = wapbl_write_commit(wl, head, tail);
1472 	if (error)
1473 		goto out2;
1474 
1475 	we = wapbl_calloc(1, sizeof(*we));
1476 
1477 #ifdef WAPBL_DEBUG_BUFBYTES
1478 	WAPBL_PRINTF(WAPBL_PRINT_FLUSH,
1479 		("wapbl_flush: thread %d.%d head+=%zu tail+=%zu used=%zu"
1480 		 " unsynced=%zu"
1481 		 "\n\tbufcount=%zu bufbytes=%zu bcount=%zu deallocs=%d "
1482 		 "inodes=%d\n",
1483 		 curproc->p_pid, curlwp->l_lid, flushsize, delta,
1484 		 wapbl_space_used(wl->wl_circ_size, head, tail),
1485 		 wl->wl_unsynced_bufbytes, wl->wl_bufcount,
1486 		 wl->wl_bufbytes, wl->wl_bcount, wl->wl_dealloccnt,
1487 		 wl->wl_inohashcnt));
1488 #else
1489 	WAPBL_PRINTF(WAPBL_PRINT_FLUSH,
1490 		("wapbl_flush: thread %d.%d head+=%zu tail+=%zu used=%zu"
1491 		 "\n\tbufcount=%zu bufbytes=%zu bcount=%zu deallocs=%d "
1492 		 "inodes=%d\n",
1493 		 curproc->p_pid, curlwp->l_lid, flushsize, delta,
1494 		 wapbl_space_used(wl->wl_circ_size, head, tail),
1495 		 wl->wl_bufcount, wl->wl_bufbytes, wl->wl_bcount,
1496 		 wl->wl_dealloccnt, wl->wl_inohashcnt));
1497 #endif
1498 
1499 
1500 	mutex_enter(&bufcache_lock);
1501 	mutex_enter(&wl->wl_mtx);
1502 
1503 	wl->wl_reserved_bytes = reserved;
1504 	wl->wl_head = head;
1505 	wl->wl_tail = tail;
1506 	KASSERT(wl->wl_reclaimable_bytes >= delta);
1507 	wl->wl_reclaimable_bytes -= delta;
1508 	wl->wl_dealloccnt = 0;
1509 #ifdef WAPBL_DEBUG_BUFBYTES
1510 	wl->wl_unsynced_bufbytes += wl->wl_bufbytes;
1511 #endif
1512 
1513 	we->we_wapbl = wl;
1514 	we->we_bufcount = wl->wl_bufcount;
1515 #ifdef WAPBL_DEBUG_BUFBYTES
1516 	we->we_unsynced_bufbytes = wl->wl_bufbytes;
1517 #endif
1518 	we->we_reclaimable_bytes = flushsize;
1519 	we->we_error = 0;
1520 	SIMPLEQ_INSERT_TAIL(&wl->wl_entries, we, we_entries);
1521 
1522 	/*
1523 	 * this flushes bufs in reverse order than they were queued
1524 	 * it shouldn't matter, but if we care we could use TAILQ instead.
1525 	 * XXX Note they will get put on the lru queue when they flush
1526 	 * so we might actually want to change this to preserve order.
1527 	 */
1528 	while ((bp = LIST_FIRST(&wl->wl_bufs)) != NULL) {
1529 		if (bbusy(bp, 0, 0, &wl->wl_mtx)) {
1530 			continue;
1531 		}
1532 		bp->b_iodone = wapbl_biodone;
1533 		bp->b_private = we;
1534 		bremfree(bp);
1535 		wapbl_remove_buf_locked(wl, bp);
1536 		mutex_exit(&wl->wl_mtx);
1537 		mutex_exit(&bufcache_lock);
1538 		bawrite(bp);
1539 		mutex_enter(&bufcache_lock);
1540 		mutex_enter(&wl->wl_mtx);
1541 	}
1542 	mutex_exit(&wl->wl_mtx);
1543 	mutex_exit(&bufcache_lock);
1544 
1545 #if 0
1546 	WAPBL_PRINTF(WAPBL_PRINT_FLUSH,
1547 		     ("wapbl_flush thread %d.%d done flushing entries...\n",
1548 		     curproc->p_pid, curlwp->l_lid));
1549 #endif
1550 
1551  out:
1552 
1553 	/*
1554 	 * If the waitfor flag is set, don't return until everything is
1555 	 * fully flushed and the on disk log is empty.
1556 	 */
1557 	if (waitfor) {
1558 		error = wapbl_truncate(wl, wl->wl_circ_size -
1559 			wl->wl_reserved_bytes, wapbl_lazy_truncate);
1560 	}
1561 
1562  out2:
1563 	if (error) {
1564 		wl->wl_flush_abort(wl->wl_mount, wl->wl_deallocblks,
1565 		    wl->wl_dealloclens, wl->wl_dealloccnt);
1566 	}
1567 
1568 #ifdef WAPBL_DEBUG_PRINT
1569 	if (error) {
1570 		pid_t pid = -1;
1571 		lwpid_t lid = -1;
1572 		if (curproc)
1573 			pid = curproc->p_pid;
1574 		if (curlwp)
1575 			lid = curlwp->l_lid;
1576 		mutex_enter(&wl->wl_mtx);
1577 #ifdef WAPBL_DEBUG_BUFBYTES
1578 		WAPBL_PRINTF(WAPBL_PRINT_ERROR,
1579 		    ("wapbl_flush: thread %d.%d aborted flush: "
1580 		    "error = %d\n"
1581 		    "\tbufcount=%zu bufbytes=%zu bcount=%zu "
1582 		    "deallocs=%d inodes=%d\n"
1583 		    "\terrcnt = %d, reclaimable=%zu reserved=%zu "
1584 		    "unsynced=%zu\n",
1585 		    pid, lid, error, wl->wl_bufcount,
1586 		    wl->wl_bufbytes, wl->wl_bcount,
1587 		    wl->wl_dealloccnt, wl->wl_inohashcnt,
1588 		    wl->wl_error_count, wl->wl_reclaimable_bytes,
1589 		    wl->wl_reserved_bytes, wl->wl_unsynced_bufbytes));
1590 		SIMPLEQ_FOREACH(we, &wl->wl_entries, we_entries) {
1591 			WAPBL_PRINTF(WAPBL_PRINT_ERROR,
1592 			    ("\tentry: bufcount = %zu, reclaimable = %zu, "
1593 			     "error = %d, unsynced = %zu\n",
1594 			     we->we_bufcount, we->we_reclaimable_bytes,
1595 			     we->we_error, we->we_unsynced_bufbytes));
1596 		}
1597 #else
1598 		WAPBL_PRINTF(WAPBL_PRINT_ERROR,
1599 		    ("wapbl_flush: thread %d.%d aborted flush: "
1600 		     "error = %d\n"
1601 		     "\tbufcount=%zu bufbytes=%zu bcount=%zu "
1602 		     "deallocs=%d inodes=%d\n"
1603 		     "\terrcnt = %d, reclaimable=%zu reserved=%zu\n",
1604 		     pid, lid, error, wl->wl_bufcount,
1605 		     wl->wl_bufbytes, wl->wl_bcount,
1606 		     wl->wl_dealloccnt, wl->wl_inohashcnt,
1607 		     wl->wl_error_count, wl->wl_reclaimable_bytes,
1608 		     wl->wl_reserved_bytes));
1609 		SIMPLEQ_FOREACH(we, &wl->wl_entries, we_entries) {
1610 			WAPBL_PRINTF(WAPBL_PRINT_ERROR,
1611 			    ("\tentry: bufcount = %zu, reclaimable = %zu, "
1612 			     "error = %d\n", we->we_bufcount,
1613 			     we->we_reclaimable_bytes, we->we_error));
1614 		}
1615 #endif
1616 		mutex_exit(&wl->wl_mtx);
1617 	}
1618 #endif
1619 
1620 	rw_exit(&wl->wl_rwlock);
1621 	return error;
1622 }
1623 
1624 /****************************************************************/
1625 
1626 void
1627 wapbl_jlock_assert(struct wapbl *wl)
1628 {
1629 
1630 	KASSERT(rw_lock_held(&wl->wl_rwlock));
1631 }
1632 
1633 void
1634 wapbl_junlock_assert(struct wapbl *wl)
1635 {
1636 
1637 	KASSERT(!rw_write_held(&wl->wl_rwlock));
1638 }
1639 
1640 /****************************************************************/
1641 
1642 /* locks missing */
1643 void
1644 wapbl_print(struct wapbl *wl,
1645 		int full,
1646 		void (*pr)(const char *, ...))
1647 {
1648 	struct buf *bp;
1649 	struct wapbl_entry *we;
1650 	(*pr)("wapbl %p", wl);
1651 	(*pr)("\nlogvp = %p, devvp = %p, logpbn = %"PRId64"\n",
1652 	      wl->wl_logvp, wl->wl_devvp, wl->wl_logpbn);
1653 	(*pr)("circ = %zu, header = %zu, head = %"PRIdMAX" tail = %"PRIdMAX"\n",
1654 	      wl->wl_circ_size, wl->wl_circ_off,
1655 	      (intmax_t)wl->wl_head, (intmax_t)wl->wl_tail);
1656 	(*pr)("fs_dev_bshift = %d, log_dev_bshift = %d\n",
1657 	      wl->wl_log_dev_bshift, wl->wl_fs_dev_bshift);
1658 #ifdef WAPBL_DEBUG_BUFBYTES
1659 	(*pr)("bufcount = %zu, bufbytes = %zu bcount = %zu reclaimable = %zu "
1660 	      "reserved = %zu errcnt = %d unsynced = %zu\n",
1661 	      wl->wl_bufcount, wl->wl_bufbytes, wl->wl_bcount,
1662 	      wl->wl_reclaimable_bytes, wl->wl_reserved_bytes,
1663 				wl->wl_error_count, wl->wl_unsynced_bufbytes);
1664 #else
1665 	(*pr)("bufcount = %zu, bufbytes = %zu bcount = %zu reclaimable = %zu "
1666 	      "reserved = %zu errcnt = %d\n", wl->wl_bufcount, wl->wl_bufbytes,
1667 	      wl->wl_bcount, wl->wl_reclaimable_bytes, wl->wl_reserved_bytes,
1668 				wl->wl_error_count);
1669 #endif
1670 	(*pr)("\tdealloccnt = %d, dealloclim = %d\n",
1671 	      wl->wl_dealloccnt, wl->wl_dealloclim);
1672 	(*pr)("\tinohashcnt = %d, inohashmask = 0x%08x\n",
1673 	      wl->wl_inohashcnt, wl->wl_inohashmask);
1674 	(*pr)("entries:\n");
1675 	SIMPLEQ_FOREACH(we, &wl->wl_entries, we_entries) {
1676 #ifdef WAPBL_DEBUG_BUFBYTES
1677 		(*pr)("\tbufcount = %zu, reclaimable = %zu, error = %d, "
1678 		      "unsynced = %zu\n",
1679 		      we->we_bufcount, we->we_reclaimable_bytes,
1680 		      we->we_error, we->we_unsynced_bufbytes);
1681 #else
1682 		(*pr)("\tbufcount = %zu, reclaimable = %zu, error = %d\n",
1683 		      we->we_bufcount, we->we_reclaimable_bytes, we->we_error);
1684 #endif
1685 	}
1686 	if (full) {
1687 		int cnt = 0;
1688 		(*pr)("bufs =");
1689 		LIST_FOREACH(bp, &wl->wl_bufs, b_wapbllist) {
1690 			if (!LIST_NEXT(bp, b_wapbllist)) {
1691 				(*pr)(" %p", bp);
1692 			} else if ((++cnt % 6) == 0) {
1693 				(*pr)(" %p,\n\t", bp);
1694 			} else {
1695 				(*pr)(" %p,", bp);
1696 			}
1697 		}
1698 		(*pr)("\n");
1699 
1700 		(*pr)("dealloced blks = ");
1701 		{
1702 			int i;
1703 			cnt = 0;
1704 			for (i = 0; i < wl->wl_dealloccnt; i++) {
1705 				(*pr)(" %"PRId64":%d,",
1706 				      wl->wl_deallocblks[i],
1707 				      wl->wl_dealloclens[i]);
1708 				if ((++cnt % 4) == 0) {
1709 					(*pr)("\n\t");
1710 				}
1711 			}
1712 		}
1713 		(*pr)("\n");
1714 
1715 		(*pr)("registered inodes = ");
1716 		{
1717 			int i;
1718 			cnt = 0;
1719 			for (i = 0; i <= wl->wl_inohashmask; i++) {
1720 				struct wapbl_ino_head *wih;
1721 				struct wapbl_ino *wi;
1722 
1723 				wih = &wl->wl_inohash[i];
1724 				LIST_FOREACH(wi, wih, wi_hash) {
1725 					if (wi->wi_ino == 0)
1726 						continue;
1727 					(*pr)(" %"PRId32"/0%06"PRIo32",",
1728 					    wi->wi_ino, wi->wi_mode);
1729 					if ((++cnt % 4) == 0) {
1730 						(*pr)("\n\t");
1731 					}
1732 				}
1733 			}
1734 			(*pr)("\n");
1735 		}
1736 	}
1737 }
1738 
1739 #if defined(WAPBL_DEBUG) || defined(DDB)
1740 void
1741 wapbl_dump(struct wapbl *wl)
1742 {
1743 #if defined(WAPBL_DEBUG)
1744 	if (!wl)
1745 		wl = wapbl_debug_wl;
1746 #endif
1747 	if (!wl)
1748 		return;
1749 	wapbl_print(wl, 1, printf);
1750 }
1751 #endif
1752 
1753 /****************************************************************/
1754 
1755 void
1756 wapbl_register_deallocation(struct wapbl *wl, daddr_t blk, int len)
1757 {
1758 
1759 	wapbl_jlock_assert(wl);
1760 
1761 	mutex_enter(&wl->wl_mtx);
1762 	/* XXX should eventually instead tie this into resource estimation */
1763 	/*
1764 	 * XXX this panic needs locking/mutex analysis and the
1765 	 * ability to cope with the failure.
1766 	 */
1767 	/* XXX this XXX doesn't have enough XXX */
1768 	if (__predict_false(wl->wl_dealloccnt >= wl->wl_dealloclim))
1769 		panic("wapbl_register_deallocation: out of resources");
1770 
1771 	wl->wl_deallocblks[wl->wl_dealloccnt] = blk;
1772 	wl->wl_dealloclens[wl->wl_dealloccnt] = len;
1773 	wl->wl_dealloccnt++;
1774 	WAPBL_PRINTF(WAPBL_PRINT_ALLOC,
1775 	    ("wapbl_register_deallocation: blk=%"PRId64" len=%d\n", blk, len));
1776 	mutex_exit(&wl->wl_mtx);
1777 }
1778 
1779 /****************************************************************/
1780 
1781 static void
1782 wapbl_inodetrk_init(struct wapbl *wl, u_int size)
1783 {
1784 
1785 	wl->wl_inohash = hashinit(size, HASH_LIST, true, &wl->wl_inohashmask);
1786 	if (atomic_inc_uint_nv(&wapbl_ino_pool_refcount) == 1) {
1787 		pool_init(&wapbl_ino_pool, sizeof(struct wapbl_ino), 0, 0, 0,
1788 		    "wapblinopl", &pool_allocator_nointr, IPL_NONE);
1789 	}
1790 }
1791 
1792 static void
1793 wapbl_inodetrk_free(struct wapbl *wl)
1794 {
1795 
1796 	/* XXX this KASSERT needs locking/mutex analysis */
1797 	KASSERT(wl->wl_inohashcnt == 0);
1798 	hashdone(wl->wl_inohash, HASH_LIST, wl->wl_inohashmask);
1799 	if (atomic_dec_uint_nv(&wapbl_ino_pool_refcount) == 0) {
1800 		pool_destroy(&wapbl_ino_pool);
1801 	}
1802 }
1803 
1804 static struct wapbl_ino *
1805 wapbl_inodetrk_get(struct wapbl *wl, ino_t ino)
1806 {
1807 	struct wapbl_ino_head *wih;
1808 	struct wapbl_ino *wi;
1809 
1810 	KASSERT(mutex_owned(&wl->wl_mtx));
1811 
1812 	wih = &wl->wl_inohash[ino & wl->wl_inohashmask];
1813 	LIST_FOREACH(wi, wih, wi_hash) {
1814 		if (ino == wi->wi_ino)
1815 			return wi;
1816 	}
1817 	return 0;
1818 }
1819 
1820 void
1821 wapbl_register_inode(struct wapbl *wl, ino_t ino, mode_t mode)
1822 {
1823 	struct wapbl_ino_head *wih;
1824 	struct wapbl_ino *wi;
1825 
1826 	wi = pool_get(&wapbl_ino_pool, PR_WAITOK);
1827 
1828 	mutex_enter(&wl->wl_mtx);
1829 	if (wapbl_inodetrk_get(wl, ino) == NULL) {
1830 		wi->wi_ino = ino;
1831 		wi->wi_mode = mode;
1832 		wih = &wl->wl_inohash[ino & wl->wl_inohashmask];
1833 		LIST_INSERT_HEAD(wih, wi, wi_hash);
1834 		wl->wl_inohashcnt++;
1835 		WAPBL_PRINTF(WAPBL_PRINT_INODE,
1836 		    ("wapbl_register_inode: ino=%"PRId64"\n", ino));
1837 		mutex_exit(&wl->wl_mtx);
1838 	} else {
1839 		mutex_exit(&wl->wl_mtx);
1840 		pool_put(&wapbl_ino_pool, wi);
1841 	}
1842 }
1843 
1844 void
1845 wapbl_unregister_inode(struct wapbl *wl, ino_t ino, mode_t mode)
1846 {
1847 	struct wapbl_ino *wi;
1848 
1849 	mutex_enter(&wl->wl_mtx);
1850 	wi = wapbl_inodetrk_get(wl, ino);
1851 	if (wi) {
1852 		WAPBL_PRINTF(WAPBL_PRINT_INODE,
1853 		    ("wapbl_unregister_inode: ino=%"PRId64"\n", ino));
1854 		KASSERT(wl->wl_inohashcnt > 0);
1855 		wl->wl_inohashcnt--;
1856 		LIST_REMOVE(wi, wi_hash);
1857 		mutex_exit(&wl->wl_mtx);
1858 
1859 		pool_put(&wapbl_ino_pool, wi);
1860 	} else {
1861 		mutex_exit(&wl->wl_mtx);
1862 	}
1863 }
1864 
1865 /****************************************************************/
1866 
1867 static inline size_t
1868 wapbl_transaction_inodes_len(struct wapbl *wl)
1869 {
1870 	int blocklen = 1<<wl->wl_log_dev_bshift;
1871 	int iph;
1872 
1873 	/* Calculate number of inodes described in a inodelist header */
1874 	iph = (blocklen - offsetof(struct wapbl_wc_inodelist, wc_inodes)) /
1875 	    sizeof(((struct wapbl_wc_inodelist *)0)->wc_inodes[0]);
1876 
1877 	KASSERT(iph > 0);
1878 
1879 	return MAX(1, howmany(wl->wl_inohashcnt, iph)) * blocklen;
1880 }
1881 
1882 
1883 /* Calculate amount of space a transaction will take on disk */
1884 static size_t
1885 wapbl_transaction_len(struct wapbl *wl)
1886 {
1887 	int blocklen = 1<<wl->wl_log_dev_bshift;
1888 	size_t len;
1889 	int bph;
1890 
1891 	/* Calculate number of blocks described in a blocklist header */
1892 	bph = (blocklen - offsetof(struct wapbl_wc_blocklist, wc_blocks)) /
1893 	    sizeof(((struct wapbl_wc_blocklist *)0)->wc_blocks[0]);
1894 
1895 	KASSERT(bph > 0);
1896 
1897 	len = wl->wl_bcount;
1898 	len += howmany(wl->wl_bufcount, bph) * blocklen;
1899 	len += howmany(wl->wl_dealloccnt, bph) * blocklen;
1900 	len += wapbl_transaction_inodes_len(wl);
1901 
1902 	return len;
1903 }
1904 
1905 /*
1906  * Perform commit operation
1907  *
1908  * Note that generation number incrementation needs to
1909  * be protected against racing with other invocations
1910  * of wapbl_commit.  This is ok since this routine
1911  * is only invoked from wapbl_flush
1912  */
1913 static int
1914 wapbl_write_commit(struct wapbl *wl, off_t head, off_t tail)
1915 {
1916 	struct wapbl_wc_header *wc = wl->wl_wc_header;
1917 	struct timespec ts;
1918 	int error;
1919 	int force = 1;
1920 	daddr_t pbn;
1921 
1922 	if (wapbl_flush_disk_cache) {
1923 		/* XXX Calc checksum here, instead we do this for now */
1924 		error = VOP_IOCTL(wl->wl_devvp, DIOCCACHESYNC, &force,
1925 		    FWRITE, FSCRED);
1926 		if (error) {
1927 			WAPBL_PRINTF(WAPBL_PRINT_ERROR,
1928 			    ("wapbl_write_commit: DIOCCACHESYNC on dev 0x%x "
1929 			    "returned %d\n", wl->wl_devvp->v_rdev, error));
1930 		}
1931 	}
1932 
1933 	wc->wc_head = head;
1934 	wc->wc_tail = tail;
1935 	wc->wc_checksum = 0;
1936 	wc->wc_version = 1;
1937 	getnanotime(&ts);
1938 	wc->wc_time = ts.tv_sec;
1939 	wc->wc_timensec = ts.tv_nsec;
1940 
1941 	WAPBL_PRINTF(WAPBL_PRINT_WRITE,
1942 	    ("wapbl_write_commit: head = %"PRIdMAX "tail = %"PRIdMAX"\n",
1943 	    (intmax_t)head, (intmax_t)tail));
1944 
1945 	/*
1946 	 * XXX if generation will rollover, then first zero
1947 	 * over second commit header before trying to write both headers.
1948 	 */
1949 
1950 	pbn = wl->wl_logpbn + (wc->wc_generation % 2);
1951 #ifdef _KERNEL
1952 	pbn = btodb(pbn << wc->wc_log_dev_bshift);
1953 #endif
1954 	error = wapbl_write(wc, wc->wc_len, wl->wl_devvp, pbn);
1955 	if (error)
1956 		return error;
1957 
1958 	if (wapbl_flush_disk_cache) {
1959 		error = VOP_IOCTL(wl->wl_devvp, DIOCCACHESYNC, &force,
1960 		    FWRITE, FSCRED);
1961 		if (error) {
1962 			WAPBL_PRINTF(WAPBL_PRINT_ERROR,
1963 			    ("wapbl_write_commit: DIOCCACHESYNC on dev 0x%x "
1964 			    "returned %d\n", wl->wl_devvp->v_rdev, error));
1965 		}
1966 	}
1967 
1968 	/*
1969 	 * If the generation number was zero, write it out a second time.
1970 	 * This handles initialization and generation number rollover
1971 	 */
1972 	if (wc->wc_generation++ == 0) {
1973 		error = wapbl_write_commit(wl, head, tail);
1974 		/*
1975 		 * This panic should be able to be removed if we do the
1976 		 * zero'ing mentioned above, and we are certain to roll
1977 		 * back generation number on failure.
1978 		 */
1979 		if (error)
1980 			panic("wapbl_write_commit: error writing duplicate "
1981 			      "log header: %d\n", error);
1982 	}
1983 	return 0;
1984 }
1985 
1986 /* Returns new offset value */
1987 static int
1988 wapbl_write_blocks(struct wapbl *wl, off_t *offp)
1989 {
1990 	struct wapbl_wc_blocklist *wc =
1991 	    (struct wapbl_wc_blocklist *)wl->wl_wc_scratch;
1992 	int blocklen = 1<<wl->wl_log_dev_bshift;
1993 	int bph;
1994 	struct buf *bp;
1995 	off_t off = *offp;
1996 	int error;
1997 	size_t padding;
1998 
1999 	KASSERT(rw_write_held(&wl->wl_rwlock));
2000 
2001 	bph = (blocklen - offsetof(struct wapbl_wc_blocklist, wc_blocks)) /
2002 	    sizeof(((struct wapbl_wc_blocklist *)0)->wc_blocks[0]);
2003 
2004 	bp = LIST_FIRST(&wl->wl_bufs);
2005 
2006 	while (bp) {
2007 		int cnt;
2008 		struct buf *obp = bp;
2009 
2010 		KASSERT(bp->b_flags & B_LOCKED);
2011 
2012 		wc->wc_type = WAPBL_WC_BLOCKS;
2013 		wc->wc_len = blocklen;
2014 		wc->wc_blkcount = 0;
2015 		while (bp && (wc->wc_blkcount < bph)) {
2016 			/*
2017 			 * Make sure all the physical block numbers are up to
2018 			 * date.  If this is not always true on a given
2019 			 * filesystem, then VOP_BMAP must be called.  We
2020 			 * could call VOP_BMAP here, or else in the filesystem
2021 			 * specific flush callback, although neither of those
2022 			 * solutions allow us to take the vnode lock.  If a
2023 			 * filesystem requires that we must take the vnode lock
2024 			 * to call VOP_BMAP, then we can probably do it in
2025 			 * bwrite when the vnode lock should already be held
2026 			 * by the invoking code.
2027 			 */
2028 			KASSERT((bp->b_vp->v_type == VBLK) ||
2029 				 (bp->b_blkno != bp->b_lblkno));
2030 			KASSERT(bp->b_blkno > 0);
2031 
2032 			wc->wc_blocks[wc->wc_blkcount].wc_daddr = bp->b_blkno;
2033 			wc->wc_blocks[wc->wc_blkcount].wc_dlen = bp->b_bcount;
2034 			wc->wc_len += bp->b_bcount;
2035 			wc->wc_blkcount++;
2036 			bp = LIST_NEXT(bp, b_wapbllist);
2037 		}
2038 		if (wc->wc_len % blocklen != 0) {
2039 			padding = blocklen - wc->wc_len % blocklen;
2040 			wc->wc_len += padding;
2041 		} else {
2042 			padding = 0;
2043 		}
2044 
2045 		WAPBL_PRINTF(WAPBL_PRINT_WRITE,
2046 		    ("wapbl_write_blocks: len = %u (padding %zu) off = %"PRIdMAX"\n",
2047 		    wc->wc_len, padding, (intmax_t)off));
2048 
2049 		error = wapbl_circ_write(wl, wc, blocklen, &off);
2050 		if (error)
2051 			return error;
2052 		bp = obp;
2053 		cnt = 0;
2054 		while (bp && (cnt++ < bph)) {
2055 			error = wapbl_circ_write(wl, bp->b_data,
2056 			    bp->b_bcount, &off);
2057 			if (error)
2058 				return error;
2059 			bp = LIST_NEXT(bp, b_wapbllist);
2060 		}
2061 		if (padding) {
2062 			void *zero;
2063 
2064 			zero = wapbl_malloc(padding);
2065 			memset(zero, 0, padding);
2066 			error = wapbl_circ_write(wl, zero, padding, &off);
2067 			wapbl_free(zero, padding);
2068 			if (error)
2069 				return error;
2070 		}
2071 	}
2072 	*offp = off;
2073 	return 0;
2074 }
2075 
2076 static int
2077 wapbl_write_revocations(struct wapbl *wl, off_t *offp)
2078 {
2079 	struct wapbl_wc_blocklist *wc =
2080 	    (struct wapbl_wc_blocklist *)wl->wl_wc_scratch;
2081 	int i;
2082 	int blocklen = 1<<wl->wl_log_dev_bshift;
2083 	int bph;
2084 	off_t off = *offp;
2085 	int error;
2086 
2087 	if (wl->wl_dealloccnt == 0)
2088 		return 0;
2089 
2090 	bph = (blocklen - offsetof(struct wapbl_wc_blocklist, wc_blocks)) /
2091 	    sizeof(((struct wapbl_wc_blocklist *)0)->wc_blocks[0]);
2092 
2093 	i = 0;
2094 	while (i < wl->wl_dealloccnt) {
2095 		wc->wc_type = WAPBL_WC_REVOCATIONS;
2096 		wc->wc_len = blocklen;
2097 		wc->wc_blkcount = 0;
2098 		while ((i < wl->wl_dealloccnt) && (wc->wc_blkcount < bph)) {
2099 			wc->wc_blocks[wc->wc_blkcount].wc_daddr =
2100 			    wl->wl_deallocblks[i];
2101 			wc->wc_blocks[wc->wc_blkcount].wc_dlen =
2102 			    wl->wl_dealloclens[i];
2103 			wc->wc_blkcount++;
2104 			i++;
2105 		}
2106 		WAPBL_PRINTF(WAPBL_PRINT_WRITE,
2107 		    ("wapbl_write_revocations: len = %u off = %"PRIdMAX"\n",
2108 		    wc->wc_len, (intmax_t)off));
2109 		error = wapbl_circ_write(wl, wc, blocklen, &off);
2110 		if (error)
2111 			return error;
2112 	}
2113 	*offp = off;
2114 	return 0;
2115 }
2116 
2117 static int
2118 wapbl_write_inodes(struct wapbl *wl, off_t *offp)
2119 {
2120 	struct wapbl_wc_inodelist *wc =
2121 	    (struct wapbl_wc_inodelist *)wl->wl_wc_scratch;
2122 	int i;
2123 	int blocklen = 1 << wl->wl_log_dev_bshift;
2124 	off_t off = *offp;
2125 	int error;
2126 
2127 	struct wapbl_ino_head *wih;
2128 	struct wapbl_ino *wi;
2129 	int iph;
2130 
2131 	iph = (blocklen - offsetof(struct wapbl_wc_inodelist, wc_inodes)) /
2132 	    sizeof(((struct wapbl_wc_inodelist *)0)->wc_inodes[0]);
2133 
2134 	i = 0;
2135 	wih = &wl->wl_inohash[0];
2136 	wi = 0;
2137 	do {
2138 		wc->wc_type = WAPBL_WC_INODES;
2139 		wc->wc_len = blocklen;
2140 		wc->wc_inocnt = 0;
2141 		wc->wc_clear = (i == 0);
2142 		while ((i < wl->wl_inohashcnt) && (wc->wc_inocnt < iph)) {
2143 			while (!wi) {
2144 				KASSERT((wih - &wl->wl_inohash[0])
2145 				    <= wl->wl_inohashmask);
2146 				wi = LIST_FIRST(wih++);
2147 			}
2148 			wc->wc_inodes[wc->wc_inocnt].wc_inumber = wi->wi_ino;
2149 			wc->wc_inodes[wc->wc_inocnt].wc_imode = wi->wi_mode;
2150 			wc->wc_inocnt++;
2151 			i++;
2152 			wi = LIST_NEXT(wi, wi_hash);
2153 		}
2154 		WAPBL_PRINTF(WAPBL_PRINT_WRITE,
2155 		    ("wapbl_write_inodes: len = %u off = %"PRIdMAX"\n",
2156 		    wc->wc_len, (intmax_t)off));
2157 		error = wapbl_circ_write(wl, wc, blocklen, &off);
2158 		if (error)
2159 			return error;
2160 	} while (i < wl->wl_inohashcnt);
2161 
2162 	*offp = off;
2163 	return 0;
2164 }
2165 
2166 #endif /* _KERNEL */
2167 
2168 /****************************************************************/
2169 
2170 struct wapbl_blk {
2171 	LIST_ENTRY(wapbl_blk) wb_hash;
2172 	daddr_t wb_blk;
2173 	off_t wb_off; /* Offset of this block in the log */
2174 };
2175 #define	WAPBL_BLKPOOL_MIN 83
2176 
2177 static void
2178 wapbl_blkhash_init(struct wapbl_replay *wr, u_int size)
2179 {
2180 	if (size < WAPBL_BLKPOOL_MIN)
2181 		size = WAPBL_BLKPOOL_MIN;
2182 	KASSERT(wr->wr_blkhash == 0);
2183 #ifdef _KERNEL
2184 	wr->wr_blkhash = hashinit(size, HASH_LIST, true, &wr->wr_blkhashmask);
2185 #else /* ! _KERNEL */
2186 	/* Manually implement hashinit */
2187 	{
2188 		unsigned long i, hashsize;
2189 		for (hashsize = 1; hashsize < size; hashsize <<= 1)
2190 			continue;
2191 		wr->wr_blkhash = wapbl_malloc(hashsize * sizeof(*wr->wr_blkhash));
2192 		for (i = 0; i < hashsize; i++)
2193 			LIST_INIT(&wr->wr_blkhash[i]);
2194 		wr->wr_blkhashmask = hashsize - 1;
2195 	}
2196 #endif /* ! _KERNEL */
2197 }
2198 
2199 static void
2200 wapbl_blkhash_free(struct wapbl_replay *wr)
2201 {
2202 	KASSERT(wr->wr_blkhashcnt == 0);
2203 #ifdef _KERNEL
2204 	hashdone(wr->wr_blkhash, HASH_LIST, wr->wr_blkhashmask);
2205 #else /* ! _KERNEL */
2206 	wapbl_free(wr->wr_blkhash,
2207 	    (wr->wr_blkhashmask + 1) * sizeof(*wr->wr_blkhash));
2208 #endif /* ! _KERNEL */
2209 }
2210 
2211 static struct wapbl_blk *
2212 wapbl_blkhash_get(struct wapbl_replay *wr, daddr_t blk)
2213 {
2214 	struct wapbl_blk_head *wbh;
2215 	struct wapbl_blk *wb;
2216 	wbh = &wr->wr_blkhash[blk & wr->wr_blkhashmask];
2217 	LIST_FOREACH(wb, wbh, wb_hash) {
2218 		if (blk == wb->wb_blk)
2219 			return wb;
2220 	}
2221 	return 0;
2222 }
2223 
2224 static void
2225 wapbl_blkhash_ins(struct wapbl_replay *wr, daddr_t blk, off_t off)
2226 {
2227 	struct wapbl_blk_head *wbh;
2228 	struct wapbl_blk *wb;
2229 	wb = wapbl_blkhash_get(wr, blk);
2230 	if (wb) {
2231 		KASSERT(wb->wb_blk == blk);
2232 		wb->wb_off = off;
2233 	} else {
2234 		wb = wapbl_malloc(sizeof(*wb));
2235 		wb->wb_blk = blk;
2236 		wb->wb_off = off;
2237 		wbh = &wr->wr_blkhash[blk & wr->wr_blkhashmask];
2238 		LIST_INSERT_HEAD(wbh, wb, wb_hash);
2239 		wr->wr_blkhashcnt++;
2240 	}
2241 }
2242 
2243 static void
2244 wapbl_blkhash_rem(struct wapbl_replay *wr, daddr_t blk)
2245 {
2246 	struct wapbl_blk *wb = wapbl_blkhash_get(wr, blk);
2247 	if (wb) {
2248 		KASSERT(wr->wr_blkhashcnt > 0);
2249 		wr->wr_blkhashcnt--;
2250 		LIST_REMOVE(wb, wb_hash);
2251 		wapbl_free(wb, sizeof(*wb));
2252 	}
2253 }
2254 
2255 static void
2256 wapbl_blkhash_clear(struct wapbl_replay *wr)
2257 {
2258 	unsigned long i;
2259 	for (i = 0; i <= wr->wr_blkhashmask; i++) {
2260 		struct wapbl_blk *wb;
2261 
2262 		while ((wb = LIST_FIRST(&wr->wr_blkhash[i]))) {
2263 			KASSERT(wr->wr_blkhashcnt > 0);
2264 			wr->wr_blkhashcnt--;
2265 			LIST_REMOVE(wb, wb_hash);
2266 			wapbl_free(wb, sizeof(*wb));
2267 		}
2268 	}
2269 	KASSERT(wr->wr_blkhashcnt == 0);
2270 }
2271 
2272 /****************************************************************/
2273 
2274 static int
2275 wapbl_circ_read(struct wapbl_replay *wr, void *data, size_t len, off_t *offp)
2276 {
2277 	size_t slen;
2278 	off_t off = *offp;
2279 	int error;
2280 	daddr_t pbn;
2281 
2282 	KASSERT(((len >> wr->wr_log_dev_bshift) <<
2283 	    wr->wr_log_dev_bshift) == len);
2284 
2285 	if (off < wr->wr_circ_off)
2286 		off = wr->wr_circ_off;
2287 	slen = wr->wr_circ_off + wr->wr_circ_size - off;
2288 	if (slen < len) {
2289 		pbn = wr->wr_logpbn + (off >> wr->wr_log_dev_bshift);
2290 #ifdef _KERNEL
2291 		pbn = btodb(pbn << wr->wr_log_dev_bshift);
2292 #endif
2293 		error = wapbl_read(data, slen, wr->wr_devvp, pbn);
2294 		if (error)
2295 			return error;
2296 		data = (uint8_t *)data + slen;
2297 		len -= slen;
2298 		off = wr->wr_circ_off;
2299 	}
2300 	pbn = wr->wr_logpbn + (off >> wr->wr_log_dev_bshift);
2301 #ifdef _KERNEL
2302 	pbn = btodb(pbn << wr->wr_log_dev_bshift);
2303 #endif
2304 	error = wapbl_read(data, len, wr->wr_devvp, pbn);
2305 	if (error)
2306 		return error;
2307 	off += len;
2308 	if (off >= wr->wr_circ_off + wr->wr_circ_size)
2309 		off = wr->wr_circ_off;
2310 	*offp = off;
2311 	return 0;
2312 }
2313 
2314 static void
2315 wapbl_circ_advance(struct wapbl_replay *wr, size_t len, off_t *offp)
2316 {
2317 	size_t slen;
2318 	off_t off = *offp;
2319 
2320 	KASSERT(((len >> wr->wr_log_dev_bshift) <<
2321 	    wr->wr_log_dev_bshift) == len);
2322 
2323 	if (off < wr->wr_circ_off)
2324 		off = wr->wr_circ_off;
2325 	slen = wr->wr_circ_off + wr->wr_circ_size - off;
2326 	if (slen < len) {
2327 		len -= slen;
2328 		off = wr->wr_circ_off;
2329 	}
2330 	off += len;
2331 	if (off >= wr->wr_circ_off + wr->wr_circ_size)
2332 		off = wr->wr_circ_off;
2333 	*offp = off;
2334 }
2335 
2336 /****************************************************************/
2337 
2338 int
2339 wapbl_replay_start(struct wapbl_replay **wrp, struct vnode *vp,
2340 	daddr_t off, size_t count, size_t blksize)
2341 {
2342 	struct wapbl_replay *wr;
2343 	int error;
2344 	struct vnode *devvp;
2345 	daddr_t logpbn;
2346 	uint8_t *scratch;
2347 	struct wapbl_wc_header *wch;
2348 	struct wapbl_wc_header *wch2;
2349 	/* Use this until we read the actual log header */
2350 	int log_dev_bshift = ilog2(blksize);
2351 	size_t used;
2352 	daddr_t pbn;
2353 
2354 	WAPBL_PRINTF(WAPBL_PRINT_REPLAY,
2355 	    ("wapbl_replay_start: vp=%p off=%"PRId64 " count=%zu blksize=%zu\n",
2356 	    vp, off, count, blksize));
2357 
2358 	if (off < 0)
2359 		return EINVAL;
2360 
2361 	if (blksize < DEV_BSIZE)
2362 		return EINVAL;
2363 	if (blksize % DEV_BSIZE)
2364 		return EINVAL;
2365 
2366 #ifdef _KERNEL
2367 #if 0
2368 	/* XXX vp->v_size isn't reliably set for VBLK devices,
2369 	 * especially root.  However, we might still want to verify
2370 	 * that the full load is readable */
2371 	if ((off + count) * blksize > vp->v_size)
2372 		return EINVAL;
2373 #endif
2374 	if ((error = VOP_BMAP(vp, off, &devvp, &logpbn, 0)) != 0) {
2375 		return error;
2376 	}
2377 #else /* ! _KERNEL */
2378 	devvp = vp;
2379 	logpbn = off;
2380 #endif /* ! _KERNEL */
2381 
2382 	scratch = wapbl_malloc(MAXBSIZE);
2383 
2384 	pbn = logpbn;
2385 #ifdef _KERNEL
2386 	pbn = btodb(pbn << log_dev_bshift);
2387 #endif
2388 	error = wapbl_read(scratch, 2<<log_dev_bshift, devvp, pbn);
2389 	if (error)
2390 		goto errout;
2391 
2392 	wch = (struct wapbl_wc_header *)scratch;
2393 	wch2 =
2394 	    (struct wapbl_wc_header *)(scratch + (1<<log_dev_bshift));
2395 	/* XXX verify checksums and magic numbers */
2396 	if (wch->wc_type != WAPBL_WC_HEADER) {
2397 		printf("Unrecognized wapbl magic: 0x%08x\n", wch->wc_type);
2398 		error = EFTYPE;
2399 		goto errout;
2400 	}
2401 
2402 	if (wch2->wc_generation > wch->wc_generation)
2403 		wch = wch2;
2404 
2405 	wr = wapbl_calloc(1, sizeof(*wr));
2406 
2407 	wr->wr_logvp = vp;
2408 	wr->wr_devvp = devvp;
2409 	wr->wr_logpbn = logpbn;
2410 
2411 	wr->wr_scratch = scratch;
2412 
2413 	wr->wr_log_dev_bshift = wch->wc_log_dev_bshift;
2414 	wr->wr_fs_dev_bshift = wch->wc_fs_dev_bshift;
2415 	wr->wr_circ_off = wch->wc_circ_off;
2416 	wr->wr_circ_size = wch->wc_circ_size;
2417 	wr->wr_generation = wch->wc_generation;
2418 
2419 	used = wapbl_space_used(wch->wc_circ_size, wch->wc_head, wch->wc_tail);
2420 
2421 	WAPBL_PRINTF(WAPBL_PRINT_REPLAY,
2422 	    ("wapbl_replay: head=%"PRId64" tail=%"PRId64" off=%"PRId64
2423 	    " len=%"PRId64" used=%zu\n",
2424 	    wch->wc_head, wch->wc_tail, wch->wc_circ_off,
2425 	    wch->wc_circ_size, used));
2426 
2427 	wapbl_blkhash_init(wr, (used >> wch->wc_fs_dev_bshift));
2428 
2429 	error = wapbl_replay_process(wr, wch->wc_head, wch->wc_tail);
2430 	if (error) {
2431 		wapbl_replay_stop(wr);
2432 		wapbl_replay_free(wr);
2433 		return error;
2434 	}
2435 
2436 	*wrp = wr;
2437 	return 0;
2438 
2439  errout:
2440 	wapbl_free(scratch, MAXBSIZE);
2441 	return error;
2442 }
2443 
2444 void
2445 wapbl_replay_stop(struct wapbl_replay *wr)
2446 {
2447 
2448 	if (!wapbl_replay_isopen(wr))
2449 		return;
2450 
2451 	WAPBL_PRINTF(WAPBL_PRINT_REPLAY, ("wapbl_replay_stop called\n"));
2452 
2453 	wapbl_free(wr->wr_scratch, MAXBSIZE);
2454 	wr->wr_scratch = NULL;
2455 
2456 	wr->wr_logvp = NULL;
2457 
2458 	wapbl_blkhash_clear(wr);
2459 	wapbl_blkhash_free(wr);
2460 }
2461 
2462 void
2463 wapbl_replay_free(struct wapbl_replay *wr)
2464 {
2465 
2466 	KDASSERT(!wapbl_replay_isopen(wr));
2467 
2468 	if (wr->wr_inodes)
2469 		wapbl_free(wr->wr_inodes,
2470 		    wr->wr_inodescnt * sizeof(wr->wr_inodes[0]));
2471 	wapbl_free(wr, sizeof(*wr));
2472 }
2473 
2474 #ifdef _KERNEL
2475 int
2476 wapbl_replay_isopen1(struct wapbl_replay *wr)
2477 {
2478 
2479 	return wapbl_replay_isopen(wr);
2480 }
2481 #endif
2482 
2483 static void
2484 wapbl_replay_process_blocks(struct wapbl_replay *wr, off_t *offp)
2485 {
2486 	struct wapbl_wc_blocklist *wc =
2487 	    (struct wapbl_wc_blocklist *)wr->wr_scratch;
2488 	int fsblklen = 1 << wr->wr_fs_dev_bshift;
2489 	int i, j, n;
2490 
2491 	for (i = 0; i < wc->wc_blkcount; i++) {
2492 		/*
2493 		 * Enter each physical block into the hashtable independently.
2494 		 */
2495 		n = wc->wc_blocks[i].wc_dlen >> wr->wr_fs_dev_bshift;
2496 		for (j = 0; j < n; j++) {
2497 			wapbl_blkhash_ins(wr, wc->wc_blocks[i].wc_daddr + btodb(j * fsblklen),
2498 			    *offp);
2499 			wapbl_circ_advance(wr, fsblklen, offp);
2500 		}
2501 	}
2502 }
2503 
2504 static void
2505 wapbl_replay_process_revocations(struct wapbl_replay *wr)
2506 {
2507 	struct wapbl_wc_blocklist *wc =
2508 	    (struct wapbl_wc_blocklist *)wr->wr_scratch;
2509 	int fsblklen = 1 << wr->wr_fs_dev_bshift;
2510 	int i, j, n;
2511 
2512 	for (i = 0; i < wc->wc_blkcount; i++) {
2513 		/*
2514 		 * Remove any blocks found from the hashtable.
2515 		 */
2516 		n = wc->wc_blocks[i].wc_dlen >> wr->wr_fs_dev_bshift;
2517 		for (j = 0; j < n; j++)
2518 			wapbl_blkhash_rem(wr, wc->wc_blocks[i].wc_daddr + btodb(j * fsblklen));
2519 	}
2520 }
2521 
2522 static void
2523 wapbl_replay_process_inodes(struct wapbl_replay *wr, off_t oldoff, off_t newoff)
2524 {
2525 	struct wapbl_wc_inodelist *wc =
2526 	    (struct wapbl_wc_inodelist *)wr->wr_scratch;
2527 	void *new_inodes;
2528 	const size_t oldsize = wr->wr_inodescnt * sizeof(wr->wr_inodes[0]);
2529 
2530 	KASSERT(sizeof(wr->wr_inodes[0]) == sizeof(wc->wc_inodes[0]));
2531 
2532 	/*
2533 	 * Keep track of where we found this so location won't be
2534 	 * overwritten.
2535 	 */
2536 	if (wc->wc_clear) {
2537 		wr->wr_inodestail = oldoff;
2538 		wr->wr_inodescnt = 0;
2539 		if (wr->wr_inodes != NULL) {
2540 			wapbl_free(wr->wr_inodes, oldsize);
2541 			wr->wr_inodes = NULL;
2542 		}
2543 	}
2544 	wr->wr_inodeshead = newoff;
2545 	if (wc->wc_inocnt == 0)
2546 		return;
2547 
2548 	new_inodes = wapbl_malloc((wr->wr_inodescnt + wc->wc_inocnt) *
2549 	    sizeof(wr->wr_inodes[0]));
2550 	if (wr->wr_inodes != NULL) {
2551 		memcpy(new_inodes, wr->wr_inodes, oldsize);
2552 		wapbl_free(wr->wr_inodes, oldsize);
2553 	}
2554 	wr->wr_inodes = new_inodes;
2555 	memcpy(&wr->wr_inodes[wr->wr_inodescnt], wc->wc_inodes,
2556 	    wc->wc_inocnt * sizeof(wr->wr_inodes[0]));
2557 	wr->wr_inodescnt += wc->wc_inocnt;
2558 }
2559 
2560 static int
2561 wapbl_replay_process(struct wapbl_replay *wr, off_t head, off_t tail)
2562 {
2563 	off_t off;
2564 	int error;
2565 
2566 	int logblklen = 1 << wr->wr_log_dev_bshift;
2567 
2568 	wapbl_blkhash_clear(wr);
2569 
2570 	off = tail;
2571 	while (off != head) {
2572 		struct wapbl_wc_null *wcn;
2573 		off_t saveoff = off;
2574 		error = wapbl_circ_read(wr, wr->wr_scratch, logblklen, &off);
2575 		if (error)
2576 			goto errout;
2577 		wcn = (struct wapbl_wc_null *)wr->wr_scratch;
2578 		switch (wcn->wc_type) {
2579 		case WAPBL_WC_BLOCKS:
2580 			wapbl_replay_process_blocks(wr, &off);
2581 			break;
2582 
2583 		case WAPBL_WC_REVOCATIONS:
2584 			wapbl_replay_process_revocations(wr);
2585 			break;
2586 
2587 		case WAPBL_WC_INODES:
2588 			wapbl_replay_process_inodes(wr, saveoff, off);
2589 			break;
2590 
2591 		default:
2592 			printf("Unrecognized wapbl type: 0x%08x\n",
2593 			       wcn->wc_type);
2594  			error = EFTYPE;
2595 			goto errout;
2596 		}
2597 		wapbl_circ_advance(wr, wcn->wc_len, &saveoff);
2598 		if (off != saveoff) {
2599 			printf("wapbl_replay: corrupted records\n");
2600 			error = EFTYPE;
2601 			goto errout;
2602 		}
2603 	}
2604 	return 0;
2605 
2606  errout:
2607 	wapbl_blkhash_clear(wr);
2608 	return error;
2609 }
2610 
2611 #if 0
2612 int
2613 wapbl_replay_verify(struct wapbl_replay *wr, struct vnode *fsdevvp)
2614 {
2615 	off_t off;
2616 	int mismatchcnt = 0;
2617 	int logblklen = 1 << wr->wr_log_dev_bshift;
2618 	int fsblklen = 1 << wr->wr_fs_dev_bshift;
2619 	void *scratch1 = wapbl_malloc(MAXBSIZE);
2620 	void *scratch2 = wapbl_malloc(MAXBSIZE);
2621 	int error = 0;
2622 
2623 	KDASSERT(wapbl_replay_isopen(wr));
2624 
2625 	off = wch->wc_tail;
2626 	while (off != wch->wc_head) {
2627 		struct wapbl_wc_null *wcn;
2628 #ifdef DEBUG
2629 		off_t saveoff = off;
2630 #endif
2631 		error = wapbl_circ_read(wr, wr->wr_scratch, logblklen, &off);
2632 		if (error)
2633 			goto out;
2634 		wcn = (struct wapbl_wc_null *)wr->wr_scratch;
2635 		switch (wcn->wc_type) {
2636 		case WAPBL_WC_BLOCKS:
2637 			{
2638 				struct wapbl_wc_blocklist *wc =
2639 				    (struct wapbl_wc_blocklist *)wr->wr_scratch;
2640 				int i;
2641 				for (i = 0; i < wc->wc_blkcount; i++) {
2642 					int foundcnt = 0;
2643 					int dirtycnt = 0;
2644 					int j, n;
2645 					/*
2646 					 * Check each physical block into the
2647 					 * hashtable independently
2648 					 */
2649 					n = wc->wc_blocks[i].wc_dlen >>
2650 					    wch->wc_fs_dev_bshift;
2651 					for (j = 0; j < n; j++) {
2652 						struct wapbl_blk *wb =
2653 						   wapbl_blkhash_get(wr,
2654 						   wc->wc_blocks[i].wc_daddr + btodb(j * fsblklen));
2655 						if (wb && (wb->wb_off == off)) {
2656 							foundcnt++;
2657 							error =
2658 							    wapbl_circ_read(wr,
2659 							    scratch1, fsblklen,
2660 							    &off);
2661 							if (error)
2662 								goto out;
2663 							error =
2664 							    wapbl_read(scratch2,
2665 							    fsblklen, fsdevvp,
2666 							    wb->wb_blk);
2667 							if (error)
2668 								goto out;
2669 							if (memcmp(scratch1,
2670 								   scratch2,
2671 								   fsblklen)) {
2672 								printf(
2673 		"wapbl_verify: mismatch block %"PRId64" at off %"PRIdMAX"\n",
2674 		wb->wb_blk, (intmax_t)off);
2675 								dirtycnt++;
2676 								mismatchcnt++;
2677 							}
2678 						} else {
2679 							wapbl_circ_advance(wr,
2680 							    fsblklen, &off);
2681 						}
2682 					}
2683 #if 0
2684 					/*
2685 					 * If all of the blocks in an entry
2686 					 * are clean, then remove all of its
2687 					 * blocks from the hashtable since they
2688 					 * never will need replay.
2689 					 */
2690 					if ((foundcnt != 0) &&
2691 					    (dirtycnt == 0)) {
2692 						off = saveoff;
2693 						wapbl_circ_advance(wr,
2694 						    logblklen, &off);
2695 						for (j = 0; j < n; j++) {
2696 							struct wapbl_blk *wb =
2697 							   wapbl_blkhash_get(wr,
2698 							   wc->wc_blocks[i].wc_daddr + btodb(j * fsblklen));
2699 							if (wb &&
2700 							  (wb->wb_off == off)) {
2701 								wapbl_blkhash_rem(wr, wb->wb_blk);
2702 							}
2703 							wapbl_circ_advance(wr,
2704 							    fsblklen, &off);
2705 						}
2706 					}
2707 #endif
2708 				}
2709 			}
2710 			break;
2711 		case WAPBL_WC_REVOCATIONS:
2712 		case WAPBL_WC_INODES:
2713 			break;
2714 		default:
2715 			KASSERT(0);
2716 		}
2717 #ifdef DEBUG
2718 		wapbl_circ_advance(wr, wcn->wc_len, &saveoff);
2719 		KASSERT(off == saveoff);
2720 #endif
2721 	}
2722  out:
2723 	wapbl_free(scratch1, MAXBSIZE);
2724 	wapbl_free(scratch2, MAXBSIZE);
2725 	if (!error && mismatchcnt)
2726 		error = EFTYPE;
2727 	return error;
2728 }
2729 #endif
2730 
2731 int
2732 wapbl_replay_write(struct wapbl_replay *wr, struct vnode *fsdevvp)
2733 {
2734 	struct wapbl_blk *wb;
2735 	size_t i;
2736 	off_t off;
2737 	void *scratch;
2738 	int error = 0;
2739 	int fsblklen = 1 << wr->wr_fs_dev_bshift;
2740 
2741 	KDASSERT(wapbl_replay_isopen(wr));
2742 
2743 	scratch = wapbl_malloc(MAXBSIZE);
2744 
2745 	for (i = 0; i <= wr->wr_blkhashmask; ++i) {
2746 		LIST_FOREACH(wb, &wr->wr_blkhash[i], wb_hash) {
2747 			off = wb->wb_off;
2748 			error = wapbl_circ_read(wr, scratch, fsblklen, &off);
2749 			if (error)
2750 				break;
2751 			error = wapbl_write(scratch, fsblklen, fsdevvp,
2752 			    wb->wb_blk);
2753 			if (error)
2754 				break;
2755 		}
2756 	}
2757 
2758 	wapbl_free(scratch, MAXBSIZE);
2759 	return error;
2760 }
2761 
2762 int
2763 wapbl_replay_can_read(struct wapbl_replay *wr, daddr_t blk, long len)
2764 {
2765 	int fsblklen = 1 << wr->wr_fs_dev_bshift;
2766 
2767 	KDASSERT(wapbl_replay_isopen(wr));
2768 	KASSERT((len % fsblklen) == 0);
2769 
2770 	while (len != 0) {
2771 		struct wapbl_blk *wb = wapbl_blkhash_get(wr, blk);
2772 		if (wb)
2773 			return 1;
2774 		len -= fsblklen;
2775 	}
2776 	return 0;
2777 }
2778 
2779 int
2780 wapbl_replay_read(struct wapbl_replay *wr, void *data, daddr_t blk, long len)
2781 {
2782 	int fsblklen = 1 << wr->wr_fs_dev_bshift;
2783 
2784 	KDASSERT(wapbl_replay_isopen(wr));
2785 
2786 	KASSERT((len % fsblklen) == 0);
2787 
2788 	while (len != 0) {
2789 		struct wapbl_blk *wb = wapbl_blkhash_get(wr, blk);
2790 		if (wb) {
2791 			off_t off = wb->wb_off;
2792 			int error;
2793 			error = wapbl_circ_read(wr, data, fsblklen, &off);
2794 			if (error)
2795 				return error;
2796 		}
2797 		data = (uint8_t *)data + fsblklen;
2798 		len -= fsblklen;
2799 		blk++;
2800 	}
2801 	return 0;
2802 }
2803 
2804 #ifdef _KERNEL
2805 /*
2806  * This is not really a module now, but maybe on it's way to
2807  * being one some day.
2808  */
2809 MODULE(MODULE_CLASS_VFS, wapbl, NULL);
2810 
2811 static int
2812 wapbl_modcmd(modcmd_t cmd, void *arg)
2813 {
2814 
2815 	switch (cmd) {
2816 	case MODULE_CMD_INIT:
2817 		wapbl_init();
2818 		return 0;
2819 	case MODULE_CMD_FINI:
2820 #ifdef notyet
2821 		return wapbl_fini(true);
2822 #endif
2823 		return EOPNOTSUPP;
2824 	default:
2825 		return ENOTTY;
2826 	}
2827 }
2828 #endif /* _KERNEL */
2829