xref: /freebsd/sys/geom/journal/g_journal.c (revision 4c1a82cea504df7a79f5bd8f7d0a41cacccff16e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2005-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
35 #include <sys/module.h>
36 #include <sys/limits.h>
37 #include <sys/lock.h>
38 #include <sys/mutex.h>
39 #include <sys/bio.h>
40 #include <sys/sysctl.h>
41 #include <sys/malloc.h>
42 #include <sys/mount.h>
43 #include <sys/eventhandler.h>
44 #include <sys/proc.h>
45 #include <sys/kthread.h>
46 #include <sys/sched.h>
47 #include <sys/taskqueue.h>
48 #include <sys/vnode.h>
49 #include <sys/sbuf.h>
50 #ifdef GJ_MEMDEBUG
51 #include <sys/stack.h>
52 #include <sys/kdb.h>
53 #endif
54 #include <vm/vm.h>
55 #include <vm/vm_kern.h>
56 #include <geom/geom.h>
57 #include <geom/geom_dbg.h>
58 
59 #include <geom/journal/g_journal.h>
60 
61 FEATURE(geom_journal, "GEOM journaling support");
62 
63 /*
64  * On-disk journal format:
65  *
66  * JH - Journal header
67  * RH - Record header
68  *
69  * %%%%%% ****** +------+ +------+     ****** +------+     %%%%%%
70  * % JH % * RH * | Data | | Data | ... * RH * | Data | ... % JH % ...
71  * %%%%%% ****** +------+ +------+     ****** +------+     %%%%%%
72  *
73  */
74 
75 CTASSERT(sizeof(struct g_journal_header) <= 512);
76 CTASSERT(sizeof(struct g_journal_record_header) <= 512);
77 
78 static MALLOC_DEFINE(M_JOURNAL, "journal_data", "GEOM_JOURNAL Data");
79 static struct mtx g_journal_cache_mtx;
80 MTX_SYSINIT(g_journal_cache, &g_journal_cache_mtx, "cache usage", MTX_DEF);
81 
82 const struct g_journal_desc *g_journal_filesystems[] = {
83 	&g_journal_ufs,
84 	NULL
85 };
86 
87 SYSCTL_DECL(_kern_geom);
88 
89 int g_journal_debug = 0;
90 static u_int g_journal_switch_time = 10;
91 static u_int g_journal_force_switch = 70;
92 static u_int g_journal_parallel_flushes = 16;
93 static u_int g_journal_parallel_copies = 16;
94 static u_int g_journal_accept_immediately = 64;
95 static u_int g_journal_record_entries = GJ_RECORD_HEADER_NENTRIES;
96 static u_int g_journal_do_optimize = 1;
97 
98 static SYSCTL_NODE(_kern_geom, OID_AUTO, journal, CTLFLAG_RW, 0,
99     "GEOM_JOURNAL stuff");
100 SYSCTL_INT(_kern_geom_journal, OID_AUTO, debug, CTLFLAG_RWTUN, &g_journal_debug, 0,
101     "Debug level");
102 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, switch_time, CTLFLAG_RW,
103     &g_journal_switch_time, 0, "Switch journals every N seconds");
104 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, force_switch, CTLFLAG_RW,
105     &g_journal_force_switch, 0, "Force switch when journal is N% full");
106 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_flushes, CTLFLAG_RW,
107     &g_journal_parallel_flushes, 0,
108     "Number of flush I/O requests to send in parallel");
109 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, accept_immediately, CTLFLAG_RW,
110     &g_journal_accept_immediately, 0,
111     "Number of I/O requests accepted immediately");
112 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_copies, CTLFLAG_RW,
113     &g_journal_parallel_copies, 0,
114     "Number of copy I/O requests to send in parallel");
115 static int
116 g_journal_record_entries_sysctl(SYSCTL_HANDLER_ARGS)
117 {
118 	u_int entries;
119 	int error;
120 
121 	entries = g_journal_record_entries;
122 	error = sysctl_handle_int(oidp, &entries, 0, req);
123 	if (error != 0 || req->newptr == NULL)
124 		return (error);
125 	if (entries < 1 || entries > GJ_RECORD_HEADER_NENTRIES)
126 		return (EINVAL);
127 	g_journal_record_entries = entries;
128 	return (0);
129 }
130 SYSCTL_PROC(_kern_geom_journal, OID_AUTO, record_entries,
131     CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_record_entries_sysctl, "I",
132     "Maximum number of entires in one journal record");
133 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, optimize, CTLFLAG_RW,
134     &g_journal_do_optimize, 0, "Try to combine bios on flush and copy");
135 
136 static u_long g_journal_cache_used = 0;
137 static u_long g_journal_cache_limit = 64 * 1024 * 1024;
138 static u_int g_journal_cache_divisor = 2;
139 static u_int g_journal_cache_switch = 90;
140 static u_int g_journal_cache_misses = 0;
141 static u_int g_journal_cache_alloc_failures = 0;
142 static u_long g_journal_cache_low = 0;
143 
144 static SYSCTL_NODE(_kern_geom_journal, OID_AUTO, cache, CTLFLAG_RW, 0,
145     "GEOM_JOURNAL cache");
146 SYSCTL_ULONG(_kern_geom_journal_cache, OID_AUTO, used, CTLFLAG_RD,
147     &g_journal_cache_used, 0, "Number of allocated bytes");
148 static int
149 g_journal_cache_limit_sysctl(SYSCTL_HANDLER_ARGS)
150 {
151 	u_long limit;
152 	int error;
153 
154 	limit = g_journal_cache_limit;
155 	error = sysctl_handle_long(oidp, &limit, 0, req);
156 	if (error != 0 || req->newptr == NULL)
157 		return (error);
158 	g_journal_cache_limit = limit;
159 	g_journal_cache_low = (limit / 100) * g_journal_cache_switch;
160 	return (0);
161 }
162 SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, limit,
163     CTLTYPE_ULONG | CTLFLAG_RWTUN, NULL, 0, g_journal_cache_limit_sysctl, "I",
164     "Maximum number of allocated bytes");
165 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, divisor, CTLFLAG_RDTUN,
166     &g_journal_cache_divisor, 0,
167     "(kmem_size / kern.geom.journal.cache.divisor) == cache size");
168 static int
169 g_journal_cache_switch_sysctl(SYSCTL_HANDLER_ARGS)
170 {
171 	u_int cswitch;
172 	int error;
173 
174 	cswitch = g_journal_cache_switch;
175 	error = sysctl_handle_int(oidp, &cswitch, 0, req);
176 	if (error != 0 || req->newptr == NULL)
177 		return (error);
178 	if (cswitch > 100)
179 		return (EINVAL);
180 	g_journal_cache_switch = cswitch;
181 	g_journal_cache_low = (g_journal_cache_limit / 100) * cswitch;
182 	return (0);
183 }
184 SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, switch,
185     CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_cache_switch_sysctl, "I",
186     "Force switch when we hit this percent of cache use");
187 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, misses, CTLFLAG_RW,
188     &g_journal_cache_misses, 0, "Number of cache misses");
189 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, alloc_failures, CTLFLAG_RW,
190     &g_journal_cache_alloc_failures, 0, "Memory allocation failures");
191 
192 static u_long g_journal_stats_bytes_skipped = 0;
193 static u_long g_journal_stats_combined_ios = 0;
194 static u_long g_journal_stats_switches = 0;
195 static u_long g_journal_stats_wait_for_copy = 0;
196 static u_long g_journal_stats_journal_full = 0;
197 static u_long g_journal_stats_low_mem = 0;
198 
199 static SYSCTL_NODE(_kern_geom_journal, OID_AUTO, stats, CTLFLAG_RW, 0,
200     "GEOM_JOURNAL statistics");
201 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, skipped_bytes, CTLFLAG_RW,
202     &g_journal_stats_bytes_skipped, 0, "Number of skipped bytes");
203 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, combined_ios, CTLFLAG_RW,
204     &g_journal_stats_combined_ios, 0, "Number of combined I/O requests");
205 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, switches, CTLFLAG_RW,
206     &g_journal_stats_switches, 0, "Number of journal switches");
207 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, wait_for_copy, CTLFLAG_RW,
208     &g_journal_stats_wait_for_copy, 0, "Wait for journal copy on switch");
209 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, journal_full, CTLFLAG_RW,
210     &g_journal_stats_journal_full, 0,
211     "Number of times journal was almost full.");
212 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, low_mem, CTLFLAG_RW,
213     &g_journal_stats_low_mem, 0, "Number of times low_mem hook was called.");
214 
215 static g_taste_t g_journal_taste;
216 static g_ctl_req_t g_journal_config;
217 static g_dumpconf_t g_journal_dumpconf;
218 static g_init_t g_journal_init;
219 static g_fini_t g_journal_fini;
220 
221 struct g_class g_journal_class = {
222 	.name = G_JOURNAL_CLASS_NAME,
223 	.version = G_VERSION,
224 	.taste = g_journal_taste,
225 	.ctlreq = g_journal_config,
226 	.dumpconf = g_journal_dumpconf,
227 	.init = g_journal_init,
228 	.fini = g_journal_fini
229 };
230 
231 static int g_journal_destroy(struct g_journal_softc *sc);
232 static void g_journal_metadata_update(struct g_journal_softc *sc);
233 static void g_journal_start_switcher(struct g_class *mp);
234 static void g_journal_stop_switcher(void);
235 static void g_journal_switch_wait(struct g_journal_softc *sc);
236 
237 #define	GJ_SWITCHER_WORKING	0
238 #define	GJ_SWITCHER_DIE		1
239 #define	GJ_SWITCHER_DIED	2
240 static struct proc *g_journal_switcher_proc = NULL;
241 static int g_journal_switcher_state = GJ_SWITCHER_WORKING;
242 static int g_journal_switcher_wokenup = 0;
243 static int g_journal_sync_requested = 0;
244 
245 #ifdef GJ_MEMDEBUG
246 struct meminfo {
247 	size_t		mi_size;
248 	struct stack	mi_stack;
249 };
250 #endif
251 
252 /*
253  * We use our own malloc/realloc/free funtions, so we can collect statistics
254  * and force journal switch when we're running out of cache.
255  */
256 static void *
257 gj_malloc(size_t size, int flags)
258 {
259 	void *p;
260 #ifdef GJ_MEMDEBUG
261 	struct meminfo *mi;
262 #endif
263 
264 	mtx_lock(&g_journal_cache_mtx);
265 	if (g_journal_cache_limit > 0 && !g_journal_switcher_wokenup &&
266 	    g_journal_cache_used + size > g_journal_cache_low) {
267 		GJ_DEBUG(1, "No cache, waking up the switcher.");
268 		g_journal_switcher_wokenup = 1;
269 		wakeup(&g_journal_switcher_state);
270 	}
271 	if ((flags & M_NOWAIT) && g_journal_cache_limit > 0 &&
272 	    g_journal_cache_used + size > g_journal_cache_limit) {
273 		mtx_unlock(&g_journal_cache_mtx);
274 		g_journal_cache_alloc_failures++;
275 		return (NULL);
276 	}
277 	g_journal_cache_used += size;
278 	mtx_unlock(&g_journal_cache_mtx);
279 	flags &= ~M_NOWAIT;
280 #ifndef GJ_MEMDEBUG
281 	p = malloc(size, M_JOURNAL, flags | M_WAITOK);
282 #else
283 	mi = malloc(sizeof(*mi) + size, M_JOURNAL, flags | M_WAITOK);
284 	p = (u_char *)mi + sizeof(*mi);
285 	mi->mi_size = size;
286 	stack_save(&mi->mi_stack);
287 #endif
288 	return (p);
289 }
290 
291 static void
292 gj_free(void *p, size_t size)
293 {
294 #ifdef GJ_MEMDEBUG
295 	struct meminfo *mi;
296 #endif
297 
298 	KASSERT(p != NULL, ("p=NULL"));
299 	KASSERT(size > 0, ("size=0"));
300 	mtx_lock(&g_journal_cache_mtx);
301 	KASSERT(g_journal_cache_used >= size, ("Freeing too much?"));
302 	g_journal_cache_used -= size;
303 	mtx_unlock(&g_journal_cache_mtx);
304 #ifdef GJ_MEMDEBUG
305 	mi = p = (void *)((u_char *)p - sizeof(*mi));
306 	if (mi->mi_size != size) {
307 		printf("GJOURNAL: Size mismatch! %zu != %zu\n", size,
308 		    mi->mi_size);
309 		printf("GJOURNAL: Alloc backtrace:\n");
310 		stack_print(&mi->mi_stack);
311 		printf("GJOURNAL: Free backtrace:\n");
312 		kdb_backtrace();
313 	}
314 #endif
315 	free(p, M_JOURNAL);
316 }
317 
318 static void *
319 gj_realloc(void *p, size_t size, size_t oldsize)
320 {
321 	void *np;
322 
323 #ifndef GJ_MEMDEBUG
324 	mtx_lock(&g_journal_cache_mtx);
325 	g_journal_cache_used -= oldsize;
326 	g_journal_cache_used += size;
327 	mtx_unlock(&g_journal_cache_mtx);
328 	np = realloc(p, size, M_JOURNAL, M_WAITOK);
329 #else
330 	np = gj_malloc(size, M_WAITOK);
331 	bcopy(p, np, MIN(oldsize, size));
332 	gj_free(p, oldsize);
333 #endif
334 	return (np);
335 }
336 
337 static void
338 g_journal_check_overflow(struct g_journal_softc *sc)
339 {
340 	off_t length, used;
341 
342 	if ((sc->sc_active.jj_offset < sc->sc_inactive.jj_offset &&
343 	     sc->sc_journal_offset >= sc->sc_inactive.jj_offset) ||
344 	    (sc->sc_active.jj_offset > sc->sc_inactive.jj_offset &&
345 	     sc->sc_journal_offset >= sc->sc_inactive.jj_offset &&
346 	     sc->sc_journal_offset < sc->sc_active.jj_offset)) {
347 		panic("Journal overflow "
348 		    "(id = %u joffset=%jd active=%jd inactive=%jd)",
349 		    (unsigned)sc->sc_id,
350 		    (intmax_t)sc->sc_journal_offset,
351 		    (intmax_t)sc->sc_active.jj_offset,
352 		    (intmax_t)sc->sc_inactive.jj_offset);
353 	}
354 	if (sc->sc_active.jj_offset < sc->sc_inactive.jj_offset) {
355 		length = sc->sc_inactive.jj_offset - sc->sc_active.jj_offset;
356 		used = sc->sc_journal_offset - sc->sc_active.jj_offset;
357 	} else {
358 		length = sc->sc_jend - sc->sc_active.jj_offset;
359 		length += sc->sc_inactive.jj_offset - sc->sc_jstart;
360 		if (sc->sc_journal_offset >= sc->sc_active.jj_offset)
361 			used = sc->sc_journal_offset - sc->sc_active.jj_offset;
362 		else {
363 			used = sc->sc_jend - sc->sc_active.jj_offset;
364 			used += sc->sc_journal_offset - sc->sc_jstart;
365 		}
366 	}
367 	/* Already woken up? */
368 	if (g_journal_switcher_wokenup)
369 		return;
370 	/*
371 	 * If the active journal takes more than g_journal_force_switch precent
372 	 * of free journal space, we force journal switch.
373 	 */
374 	KASSERT(length > 0,
375 	    ("length=%jd used=%jd active=%jd inactive=%jd joffset=%jd",
376 	    (intmax_t)length, (intmax_t)used,
377 	    (intmax_t)sc->sc_active.jj_offset,
378 	    (intmax_t)sc->sc_inactive.jj_offset,
379 	    (intmax_t)sc->sc_journal_offset));
380 	if ((used * 100) / length > g_journal_force_switch) {
381 		g_journal_stats_journal_full++;
382 		GJ_DEBUG(1, "Journal %s %jd%% full, forcing journal switch.",
383 		    sc->sc_name, (used * 100) / length);
384 		mtx_lock(&g_journal_cache_mtx);
385 		g_journal_switcher_wokenup = 1;
386 		wakeup(&g_journal_switcher_state);
387 		mtx_unlock(&g_journal_cache_mtx);
388 	}
389 }
390 
391 static void
392 g_journal_orphan(struct g_consumer *cp)
393 {
394 	struct g_journal_softc *sc;
395 	char name[256];
396 	int error;
397 
398 	g_topology_assert();
399 	sc = cp->geom->softc;
400 	strlcpy(name, cp->provider->name, sizeof(name));
401 	GJ_DEBUG(0, "Lost provider %s.", name);
402 	if (sc == NULL)
403 		return;
404 	error = g_journal_destroy(sc);
405 	if (error == 0)
406 		GJ_DEBUG(0, "Journal %s destroyed.", name);
407 	else {
408 		GJ_DEBUG(0, "Cannot destroy journal %s (error=%d). "
409 		    "Destroy it manually after last close.", sc->sc_name,
410 		    error);
411 	}
412 }
413 
414 static int
415 g_journal_access(struct g_provider *pp, int acr, int acw, int ace)
416 {
417 	struct g_journal_softc *sc;
418 	int dcr, dcw, dce;
419 
420 	g_topology_assert();
421 	GJ_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name,
422 	    acr, acw, ace);
423 
424 	dcr = pp->acr + acr;
425 	dcw = pp->acw + acw;
426 	dce = pp->ace + ace;
427 
428 	sc = pp->geom->softc;
429 	if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY)) {
430 		if (acr <= 0 && acw <= 0 && ace <= 0)
431 			return (0);
432 		else
433 			return (ENXIO);
434 	}
435 	if (pp->acw == 0 && dcw > 0) {
436 		GJ_DEBUG(1, "Marking %s as dirty.", sc->sc_name);
437 		sc->sc_flags &= ~GJF_DEVICE_CLEAN;
438 		g_topology_unlock();
439 		g_journal_metadata_update(sc);
440 		g_topology_lock();
441 	} /* else if (pp->acw == 0 && dcw > 0 && JEMPTY(sc)) {
442 		GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
443 		sc->sc_flags |= GJF_DEVICE_CLEAN;
444 		g_topology_unlock();
445 		g_journal_metadata_update(sc);
446 		g_topology_lock();
447 	} */
448 	return (0);
449 }
450 
451 static void
452 g_journal_header_encode(struct g_journal_header *hdr, u_char *data)
453 {
454 
455 	bcopy(GJ_HEADER_MAGIC, data, sizeof(GJ_HEADER_MAGIC));
456 	data += sizeof(GJ_HEADER_MAGIC);
457 	le32enc(data, hdr->jh_journal_id);
458 	data += 4;
459 	le32enc(data, hdr->jh_journal_next_id);
460 }
461 
462 static int
463 g_journal_header_decode(const u_char *data, struct g_journal_header *hdr)
464 {
465 
466 	bcopy(data, hdr->jh_magic, sizeof(hdr->jh_magic));
467 	data += sizeof(hdr->jh_magic);
468 	if (bcmp(hdr->jh_magic, GJ_HEADER_MAGIC, sizeof(GJ_HEADER_MAGIC)) != 0)
469 		return (EINVAL);
470 	hdr->jh_journal_id = le32dec(data);
471 	data += 4;
472 	hdr->jh_journal_next_id = le32dec(data);
473 	return (0);
474 }
475 
476 static void
477 g_journal_flush_cache(struct g_journal_softc *sc)
478 {
479 	struct bintime bt;
480 	int error;
481 
482 	if (sc->sc_bio_flush == 0)
483 		return;
484 	GJ_TIMER_START(1, &bt);
485 	if (sc->sc_bio_flush & GJ_FLUSH_JOURNAL) {
486 		error = g_io_flush(sc->sc_jconsumer);
487 		GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.",
488 		    sc->sc_jconsumer->provider->name, error);
489 	}
490 	if (sc->sc_bio_flush & GJ_FLUSH_DATA) {
491 		/*
492 		 * TODO: This could be called in parallel with the
493 		 *       previous call.
494 		 */
495 		error = g_io_flush(sc->sc_dconsumer);
496 		GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.",
497 		    sc->sc_dconsumer->provider->name, error);
498 	}
499 	GJ_TIMER_STOP(1, &bt, "Cache flush time");
500 }
501 
502 static int
503 g_journal_write_header(struct g_journal_softc *sc)
504 {
505 	struct g_journal_header hdr;
506 	struct g_consumer *cp;
507 	u_char *buf;
508 	int error;
509 
510 	cp = sc->sc_jconsumer;
511 	buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
512 
513 	strlcpy(hdr.jh_magic, GJ_HEADER_MAGIC, sizeof(hdr.jh_magic));
514 	hdr.jh_journal_id = sc->sc_journal_id;
515 	hdr.jh_journal_next_id = sc->sc_journal_next_id;
516 	g_journal_header_encode(&hdr, buf);
517 	error = g_write_data(cp, sc->sc_journal_offset, buf,
518 	    cp->provider->sectorsize);
519 	/* if (error == 0) */
520 	sc->sc_journal_offset += cp->provider->sectorsize;
521 
522 	gj_free(buf, cp->provider->sectorsize);
523 	return (error);
524 }
525 
526 /*
527  * Every journal record has a header and data following it.
528  * Functions below are used to decode the header before storing it to
529  * little endian and to encode it after reading to system endianness.
530  */
531 static void
532 g_journal_record_header_encode(struct g_journal_record_header *hdr,
533     u_char *data)
534 {
535 	struct g_journal_entry *ent;
536 	u_int i;
537 
538 	bcopy(GJ_RECORD_HEADER_MAGIC, data, sizeof(GJ_RECORD_HEADER_MAGIC));
539 	data += sizeof(GJ_RECORD_HEADER_MAGIC);
540 	le32enc(data, hdr->jrh_journal_id);
541 	data += 8;
542 	le16enc(data, hdr->jrh_nentries);
543 	data += 2;
544 	bcopy(hdr->jrh_sum, data, sizeof(hdr->jrh_sum));
545 	data += 8;
546 	for (i = 0; i < hdr->jrh_nentries; i++) {
547 		ent = &hdr->jrh_entries[i];
548 		le64enc(data, ent->je_joffset);
549 		data += 8;
550 		le64enc(data, ent->je_offset);
551 		data += 8;
552 		le64enc(data, ent->je_length);
553 		data += 8;
554 	}
555 }
556 
557 static int
558 g_journal_record_header_decode(const u_char *data,
559     struct g_journal_record_header *hdr)
560 {
561 	struct g_journal_entry *ent;
562 	u_int i;
563 
564 	bcopy(data, hdr->jrh_magic, sizeof(hdr->jrh_magic));
565 	data += sizeof(hdr->jrh_magic);
566 	if (strcmp(hdr->jrh_magic, GJ_RECORD_HEADER_MAGIC) != 0)
567 		return (EINVAL);
568 	hdr->jrh_journal_id = le32dec(data);
569 	data += 8;
570 	hdr->jrh_nentries = le16dec(data);
571 	data += 2;
572 	if (hdr->jrh_nentries > GJ_RECORD_HEADER_NENTRIES)
573 		return (EINVAL);
574 	bcopy(data, hdr->jrh_sum, sizeof(hdr->jrh_sum));
575 	data += 8;
576 	for (i = 0; i < hdr->jrh_nentries; i++) {
577 		ent = &hdr->jrh_entries[i];
578 		ent->je_joffset = le64dec(data);
579 		data += 8;
580 		ent->je_offset = le64dec(data);
581 		data += 8;
582 		ent->je_length = le64dec(data);
583 		data += 8;
584 	}
585 	return (0);
586 }
587 
588 /*
589  * Function reads metadata from a provider (via the given consumer), decodes
590  * it to system endianness and verifies its correctness.
591  */
592 static int
593 g_journal_metadata_read(struct g_consumer *cp, struct g_journal_metadata *md)
594 {
595 	struct g_provider *pp;
596 	u_char *buf;
597 	int error;
598 
599 	g_topology_assert();
600 
601 	error = g_access(cp, 1, 0, 0);
602 	if (error != 0)
603 		return (error);
604 	pp = cp->provider;
605 	g_topology_unlock();
606 	/* Metadata is stored in last sector. */
607 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
608 	    &error);
609 	g_topology_lock();
610 	g_access(cp, -1, 0, 0);
611 	if (buf == NULL) {
612 		GJ_DEBUG(1, "Cannot read metadata from %s (error=%d).",
613 		    cp->provider->name, error);
614 		return (error);
615 	}
616 
617 	/* Decode metadata. */
618 	error = journal_metadata_decode(buf, md);
619 	g_free(buf);
620 	/* Is this is gjournal provider at all? */
621 	if (strcmp(md->md_magic, G_JOURNAL_MAGIC) != 0)
622 		return (EINVAL);
623 	/*
624 	 * Are we able to handle this version of metadata?
625 	 * We only maintain backward compatibility.
626 	 */
627 	if (md->md_version > G_JOURNAL_VERSION) {
628 		GJ_DEBUG(0,
629 		    "Kernel module is too old to handle metadata from %s.",
630 		    cp->provider->name);
631 		return (EINVAL);
632 	}
633 	/* Is checksum correct? */
634 	if (error != 0) {
635 		GJ_DEBUG(0, "MD5 metadata hash mismatch for provider %s.",
636 		    cp->provider->name);
637 		return (error);
638 	}
639 	return (0);
640 }
641 
642 /*
643  * Two functions below are responsible for updating metadata.
644  * Only metadata on the data provider is updated (we need to update
645  * information about active journal in there).
646  */
647 static void
648 g_journal_metadata_done(struct bio *bp)
649 {
650 
651 	/*
652 	 * There is not much we can do on error except informing about it.
653 	 */
654 	if (bp->bio_error != 0) {
655 		GJ_LOGREQ(0, bp, "Cannot update metadata (error=%d).",
656 		    bp->bio_error);
657 	} else {
658 		GJ_LOGREQ(2, bp, "Metadata updated.");
659 	}
660 	gj_free(bp->bio_data, bp->bio_length);
661 	g_destroy_bio(bp);
662 }
663 
664 static void
665 g_journal_metadata_update(struct g_journal_softc *sc)
666 {
667 	struct g_journal_metadata md;
668 	struct g_consumer *cp;
669 	struct bio *bp;
670 	u_char *sector;
671 
672 	cp = sc->sc_dconsumer;
673 	sector = gj_malloc(cp->provider->sectorsize, M_WAITOK);
674 	strlcpy(md.md_magic, G_JOURNAL_MAGIC, sizeof(md.md_magic));
675 	md.md_version = G_JOURNAL_VERSION;
676 	md.md_id = sc->sc_id;
677 	md.md_type = sc->sc_orig_type;
678 	md.md_jstart = sc->sc_jstart;
679 	md.md_jend = sc->sc_jend;
680 	md.md_joffset = sc->sc_inactive.jj_offset;
681 	md.md_jid = sc->sc_journal_previous_id;
682 	md.md_flags = 0;
683 	if (sc->sc_flags & GJF_DEVICE_CLEAN)
684 		md.md_flags |= GJ_FLAG_CLEAN;
685 
686 	if (sc->sc_flags & GJF_DEVICE_HARDCODED)
687 		strlcpy(md.md_provider, sc->sc_name, sizeof(md.md_provider));
688 	else
689 		bzero(md.md_provider, sizeof(md.md_provider));
690 	md.md_provsize = cp->provider->mediasize;
691 	journal_metadata_encode(&md, sector);
692 
693 	/*
694 	 * Flush the cache, so we know all data are on disk.
695 	 * We write here informations like "journal is consistent", so we need
696 	 * to be sure it is. Without BIO_FLUSH here, we can end up in situation
697 	 * where metadata is stored on disk, but not all data.
698 	 */
699 	g_journal_flush_cache(sc);
700 
701 	bp = g_alloc_bio();
702 	bp->bio_offset = cp->provider->mediasize - cp->provider->sectorsize;
703 	bp->bio_length = cp->provider->sectorsize;
704 	bp->bio_data = sector;
705 	bp->bio_cmd = BIO_WRITE;
706 	if (!(sc->sc_flags & GJF_DEVICE_DESTROY)) {
707 		bp->bio_done = g_journal_metadata_done;
708 		g_io_request(bp, cp);
709 	} else {
710 		bp->bio_done = NULL;
711 		g_io_request(bp, cp);
712 		biowait(bp, "gjmdu");
713 		g_journal_metadata_done(bp);
714 	}
715 
716 	/*
717 	 * Be sure metadata reached the disk.
718 	 */
719 	g_journal_flush_cache(sc);
720 }
721 
722 /*
723  * This is where the I/O request comes from the GEOM.
724  */
725 static void
726 g_journal_start(struct bio *bp)
727 {
728 	struct g_journal_softc *sc;
729 
730 	sc = bp->bio_to->geom->softc;
731 	GJ_LOGREQ(3, bp, "Request received.");
732 
733 	switch (bp->bio_cmd) {
734 	case BIO_READ:
735 	case BIO_WRITE:
736 		mtx_lock(&sc->sc_mtx);
737 		bioq_insert_tail(&sc->sc_regular_queue, bp);
738 		wakeup(sc);
739 		mtx_unlock(&sc->sc_mtx);
740 		return;
741 	case BIO_GETATTR:
742 		if (strcmp(bp->bio_attribute, "GJOURNAL::provider") == 0) {
743 			strlcpy(bp->bio_data, bp->bio_to->name, bp->bio_length);
744 			bp->bio_completed = strlen(bp->bio_to->name) + 1;
745 			g_io_deliver(bp, 0);
746 			return;
747 		}
748 		/* FALLTHROUGH */
749 	case BIO_DELETE:
750 	default:
751 		g_io_deliver(bp, EOPNOTSUPP);
752 		return;
753 	}
754 }
755 
756 static void
757 g_journal_std_done(struct bio *bp)
758 {
759 	struct g_journal_softc *sc;
760 
761 	sc = bp->bio_from->geom->softc;
762 	mtx_lock(&sc->sc_mtx);
763 	bioq_insert_tail(&sc->sc_back_queue, bp);
764 	wakeup(sc);
765 	mtx_unlock(&sc->sc_mtx);
766 }
767 
768 static struct bio *
769 g_journal_new_bio(off_t start, off_t end, off_t joffset, u_char *data,
770     int flags)
771 {
772 	struct bio *bp;
773 
774 	bp = g_alloc_bio();
775 	bp->bio_offset = start;
776 	bp->bio_joffset = joffset;
777 	bp->bio_length = end - start;
778 	bp->bio_cmd = BIO_WRITE;
779 	bp->bio_done = g_journal_std_done;
780 	if (data == NULL)
781 		bp->bio_data = NULL;
782 	else {
783 		bp->bio_data = gj_malloc(bp->bio_length, flags);
784 		if (bp->bio_data != NULL)
785 			bcopy(data, bp->bio_data, bp->bio_length);
786 	}
787 	return (bp);
788 }
789 
790 #define	g_journal_insert_bio(head, bp, flags)				\
791 	g_journal_insert((head), (bp)->bio_offset,			\
792 		(bp)->bio_offset + (bp)->bio_length, (bp)->bio_joffset,	\
793 		(bp)->bio_data, flags)
794 /*
795  * The function below does a lot more than just inserting bio to the queue.
796  * It keeps the queue sorted by offset and ensures that there are no doubled
797  * data (it combines bios where ranges overlap).
798  *
799  * The function returns the number of bios inserted (as bio can be splitted).
800  */
801 static int
802 g_journal_insert(struct bio **head, off_t nstart, off_t nend, off_t joffset,
803     u_char *data, int flags)
804 {
805 	struct bio *nbp, *cbp, *pbp;
806 	off_t cstart, cend;
807 	u_char *tmpdata;
808 	int n;
809 
810 	GJ_DEBUG(3, "INSERT(%p): (%jd, %jd, %jd)", *head, nstart, nend,
811 	    joffset);
812 	n = 0;
813 	pbp = NULL;
814 	GJQ_FOREACH(*head, cbp) {
815 		cstart = cbp->bio_offset;
816 		cend = cbp->bio_offset + cbp->bio_length;
817 
818 		if (nstart >= cend) {
819 			/*
820 			 *  +-------------+
821 			 *  |             |
822 			 *  |   current   |  +-------------+
823 			 *  |     bio     |  |             |
824 			 *  |             |  |     new     |
825 			 *  +-------------+  |     bio     |
826 			 *                   |             |
827 			 *                   +-------------+
828 			 */
829 			GJ_DEBUG(3, "INSERT(%p): 1", *head);
830 		} else if (nend <= cstart) {
831 			/*
832 			 *                   +-------------+
833 			 *                   |             |
834 			 *  +-------------+  |   current   |
835 			 *  |             |  |     bio     |
836 			 *  |     new     |  |             |
837 			 *  |     bio     |  +-------------+
838 			 *  |             |
839 			 *  +-------------+
840 			 */
841 			nbp = g_journal_new_bio(nstart, nend, joffset, data,
842 			    flags);
843 			if (pbp == NULL)
844 				*head = nbp;
845 			else
846 				pbp->bio_next = nbp;
847 			nbp->bio_next = cbp;
848 			n++;
849 			GJ_DEBUG(3, "INSERT(%p): 2 (nbp=%p pbp=%p)", *head, nbp,
850 			    pbp);
851 			goto end;
852 		} else if (nstart <= cstart && nend >= cend) {
853 			/*
854 			 *      +-------------+      +-------------+
855 			 *      | current bio |      | current bio |
856 			 *  +---+-------------+---+  +-------------+---+
857 			 *  |   |             |   |  |             |   |
858 			 *  |   |             |   |  |             |   |
859 			 *  |   +-------------+   |  +-------------+   |
860 			 *  |       new bio       |  |     new bio     |
861 			 *  +---------------------+  +-----------------+
862 			 *
863 			 *      +-------------+  +-------------+
864 			 *      | current bio |  | current bio |
865 			 *  +---+-------------+  +-------------+
866 			 *  |   |             |  |             |
867 			 *  |   |             |  |             |
868 			 *  |   +-------------+  +-------------+
869 			 *  |     new bio     |  |   new bio   |
870 			 *  +-----------------+  +-------------+
871 			 */
872 			g_journal_stats_bytes_skipped += cbp->bio_length;
873 			cbp->bio_offset = nstart;
874 			cbp->bio_joffset = joffset;
875 			cbp->bio_length = cend - nstart;
876 			if (cbp->bio_data != NULL) {
877 				gj_free(cbp->bio_data, cend - cstart);
878 				cbp->bio_data = NULL;
879 			}
880 			if (data != NULL) {
881 				cbp->bio_data = gj_malloc(cbp->bio_length,
882 				    flags);
883 				if (cbp->bio_data != NULL) {
884 					bcopy(data, cbp->bio_data,
885 					    cbp->bio_length);
886 				}
887 				data += cend - nstart;
888 			}
889 			joffset += cend - nstart;
890 			nstart = cend;
891 			GJ_DEBUG(3, "INSERT(%p): 3 (cbp=%p)", *head, cbp);
892 		} else if (nstart > cstart && nend >= cend) {
893 			/*
894 			 *  +-----------------+  +-------------+
895 			 *  |   current bio   |  | current bio |
896 			 *  |   +-------------+  |   +---------+---+
897 			 *  |   |             |  |   |         |   |
898 			 *  |   |             |  |   |         |   |
899 			 *  +---+-------------+  +---+---------+   |
900 			 *      |   new bio   |      |   new bio   |
901 			 *      +-------------+      +-------------+
902 			 */
903 			g_journal_stats_bytes_skipped += cend - nstart;
904 			nbp = g_journal_new_bio(nstart, cend, joffset, data,
905 			    flags);
906 			nbp->bio_next = cbp->bio_next;
907 			cbp->bio_next = nbp;
908 			cbp->bio_length = nstart - cstart;
909 			if (cbp->bio_data != NULL) {
910 				cbp->bio_data = gj_realloc(cbp->bio_data,
911 				    cbp->bio_length, cend - cstart);
912 			}
913 			if (data != NULL)
914 				data += cend - nstart;
915 			joffset += cend - nstart;
916 			nstart = cend;
917 			n++;
918 			GJ_DEBUG(3, "INSERT(%p): 4 (cbp=%p)", *head, cbp);
919 		} else if (nstart > cstart && nend < cend) {
920 			/*
921 			 *  +---------------------+
922 			 *  |     current bio     |
923 			 *  |   +-------------+   |
924 			 *  |   |             |   |
925 			 *  |   |             |   |
926 			 *  +---+-------------+---+
927 			 *      |   new bio   |
928 			 *      +-------------+
929 			 */
930 			g_journal_stats_bytes_skipped += nend - nstart;
931 			nbp = g_journal_new_bio(nstart, nend, joffset, data,
932 			    flags);
933 			nbp->bio_next = cbp->bio_next;
934 			cbp->bio_next = nbp;
935 			if (cbp->bio_data == NULL)
936 				tmpdata = NULL;
937 			else
938 				tmpdata = cbp->bio_data + nend - cstart;
939 			nbp = g_journal_new_bio(nend, cend,
940 			    cbp->bio_joffset + nend - cstart, tmpdata, flags);
941 			nbp->bio_next = ((struct bio *)cbp->bio_next)->bio_next;
942 			((struct bio *)cbp->bio_next)->bio_next = nbp;
943 			cbp->bio_length = nstart - cstart;
944 			if (cbp->bio_data != NULL) {
945 				cbp->bio_data = gj_realloc(cbp->bio_data,
946 				    cbp->bio_length, cend - cstart);
947 			}
948 			n += 2;
949 			GJ_DEBUG(3, "INSERT(%p): 5 (cbp=%p)", *head, cbp);
950 			goto end;
951 		} else if (nstart <= cstart && nend < cend) {
952 			/*
953 			 *  +-----------------+      +-------------+
954 			 *  |   current bio   |      | current bio |
955 			 *  +-------------+   |  +---+---------+   |
956 			 *  |             |   |  |   |         |   |
957 			 *  |             |   |  |   |         |   |
958 			 *  +-------------+---+  |   +---------+---+
959 			 *  |   new bio   |      |   new bio   |
960 			 *  +-------------+      +-------------+
961 			 */
962 			g_journal_stats_bytes_skipped += nend - nstart;
963 			nbp = g_journal_new_bio(nstart, nend, joffset, data,
964 			    flags);
965 			if (pbp == NULL)
966 				*head = nbp;
967 			else
968 				pbp->bio_next = nbp;
969 			nbp->bio_next = cbp;
970 			cbp->bio_offset = nend;
971 			cbp->bio_length = cend - nend;
972 			cbp->bio_joffset += nend - cstart;
973 			tmpdata = cbp->bio_data;
974 			if (tmpdata != NULL) {
975 				cbp->bio_data = gj_malloc(cbp->bio_length,
976 				    flags);
977 				if (cbp->bio_data != NULL) {
978 					bcopy(tmpdata + nend - cstart,
979 					    cbp->bio_data, cbp->bio_length);
980 				}
981 				gj_free(tmpdata, cend - cstart);
982 			}
983 			n++;
984 			GJ_DEBUG(3, "INSERT(%p): 6 (cbp=%p)", *head, cbp);
985 			goto end;
986 		}
987 		if (nstart == nend)
988 			goto end;
989 		pbp = cbp;
990 	}
991 	nbp = g_journal_new_bio(nstart, nend, joffset, data, flags);
992 	if (pbp == NULL)
993 		*head = nbp;
994 	else
995 		pbp->bio_next = nbp;
996 	nbp->bio_next = NULL;
997 	n++;
998 	GJ_DEBUG(3, "INSERT(%p): 8 (nbp=%p pbp=%p)", *head, nbp, pbp);
999 end:
1000 	if (g_journal_debug >= 3) {
1001 		GJQ_FOREACH(*head, cbp) {
1002 			GJ_DEBUG(3, "ELEMENT: %p (%jd, %jd, %jd, %p)", cbp,
1003 			    (intmax_t)cbp->bio_offset,
1004 			    (intmax_t)cbp->bio_length,
1005 			    (intmax_t)cbp->bio_joffset, cbp->bio_data);
1006 		}
1007 		GJ_DEBUG(3, "INSERT(%p): DONE %d", *head, n);
1008 	}
1009 	return (n);
1010 }
1011 
1012 /*
1013  * The function combines neighbour bios trying to squeeze as much data as
1014  * possible into one bio.
1015  *
1016  * The function returns the number of bios combined (negative value).
1017  */
1018 static int
1019 g_journal_optimize(struct bio *head)
1020 {
1021 	struct bio *cbp, *pbp;
1022 	int n;
1023 
1024 	n = 0;
1025 	pbp = NULL;
1026 	GJQ_FOREACH(head, cbp) {
1027 		/* Skip bios which has to be read first. */
1028 		if (cbp->bio_data == NULL) {
1029 			pbp = NULL;
1030 			continue;
1031 		}
1032 		/* There is no previous bio yet. */
1033 		if (pbp == NULL) {
1034 			pbp = cbp;
1035 			continue;
1036 		}
1037 		/* Is this a neighbour bio? */
1038 		if (pbp->bio_offset + pbp->bio_length != cbp->bio_offset) {
1039 			/* Be sure that bios queue is sorted. */
1040 			KASSERT(pbp->bio_offset + pbp->bio_length < cbp->bio_offset,
1041 			    ("poffset=%jd plength=%jd coffset=%jd",
1042 			    (intmax_t)pbp->bio_offset,
1043 			    (intmax_t)pbp->bio_length,
1044 			    (intmax_t)cbp->bio_offset));
1045 			pbp = cbp;
1046 			continue;
1047 		}
1048 		/* Be sure we don't end up with too big bio. */
1049 		if (pbp->bio_length + cbp->bio_length > MAXPHYS) {
1050 			pbp = cbp;
1051 			continue;
1052 		}
1053 		/* Ok, we can join bios. */
1054 		GJ_LOGREQ(4, pbp, "Join: ");
1055 		GJ_LOGREQ(4, cbp, "and: ");
1056 		pbp->bio_data = gj_realloc(pbp->bio_data,
1057 		    pbp->bio_length + cbp->bio_length, pbp->bio_length);
1058 		bcopy(cbp->bio_data, pbp->bio_data + pbp->bio_length,
1059 		    cbp->bio_length);
1060 		gj_free(cbp->bio_data, cbp->bio_length);
1061 		pbp->bio_length += cbp->bio_length;
1062 		pbp->bio_next = cbp->bio_next;
1063 		g_destroy_bio(cbp);
1064 		cbp = pbp;
1065 		g_journal_stats_combined_ios++;
1066 		n--;
1067 		GJ_LOGREQ(4, pbp, "Got: ");
1068 	}
1069 	return (n);
1070 }
1071 
1072 /*
1073  * TODO: Update comment.
1074  * These are functions responsible for copying one portion of data from journal
1075  * to the destination provider.
1076  * The order goes like this:
1077  * 1. Read the header, which contains informations about data blocks
1078  *    following it.
1079  * 2. Read the data blocks from the journal.
1080  * 3. Write the data blocks on the data provider.
1081  *
1082  * g_journal_copy_start()
1083  * g_journal_copy_done() - got finished write request, logs potential errors.
1084  */
1085 
1086 /*
1087  * When there is no data in cache, this function is used to read it.
1088  */
1089 static void
1090 g_journal_read_first(struct g_journal_softc *sc, struct bio *bp)
1091 {
1092 	struct bio *cbp;
1093 
1094 	/*
1095 	 * We were short in memory, so data was freed.
1096 	 * In that case we need to read it back from journal.
1097 	 */
1098 	cbp = g_alloc_bio();
1099 	cbp->bio_cflags = bp->bio_cflags;
1100 	cbp->bio_parent = bp;
1101 	cbp->bio_offset = bp->bio_joffset;
1102 	cbp->bio_length = bp->bio_length;
1103 	cbp->bio_data = gj_malloc(bp->bio_length, M_WAITOK);
1104 	cbp->bio_cmd = BIO_READ;
1105 	cbp->bio_done = g_journal_std_done;
1106 	GJ_LOGREQ(4, cbp, "READ FIRST");
1107 	g_io_request(cbp, sc->sc_jconsumer);
1108 	g_journal_cache_misses++;
1109 }
1110 
1111 static void
1112 g_journal_copy_send(struct g_journal_softc *sc)
1113 {
1114 	struct bio *bioq, *bp, *lbp;
1115 
1116 	bioq = lbp = NULL;
1117 	mtx_lock(&sc->sc_mtx);
1118 	for (; sc->sc_copy_in_progress < g_journal_parallel_copies;) {
1119 		bp = GJQ_FIRST(sc->sc_inactive.jj_queue);
1120 		if (bp == NULL)
1121 			break;
1122 		GJQ_REMOVE(sc->sc_inactive.jj_queue, bp);
1123 		sc->sc_copy_in_progress++;
1124 		GJQ_INSERT_AFTER(bioq, bp, lbp);
1125 		lbp = bp;
1126 	}
1127 	mtx_unlock(&sc->sc_mtx);
1128 	if (g_journal_do_optimize)
1129 		sc->sc_copy_in_progress += g_journal_optimize(bioq);
1130 	while ((bp = GJQ_FIRST(bioq)) != NULL) {
1131 		GJQ_REMOVE(bioq, bp);
1132 		GJQ_INSERT_HEAD(sc->sc_copy_queue, bp);
1133 		bp->bio_cflags = GJ_BIO_COPY;
1134 		if (bp->bio_data == NULL)
1135 			g_journal_read_first(sc, bp);
1136 		else {
1137 			bp->bio_joffset = 0;
1138 			GJ_LOGREQ(4, bp, "SEND");
1139 			g_io_request(bp, sc->sc_dconsumer);
1140 		}
1141 	}
1142 }
1143 
1144 static void
1145 g_journal_copy_start(struct g_journal_softc *sc)
1146 {
1147 
1148 	/*
1149 	 * Remember in metadata that we're starting to copy journaled data
1150 	 * to the data provider.
1151 	 * In case of power failure, we will copy these data once again on boot.
1152 	 */
1153 	if (!sc->sc_journal_copying) {
1154 		sc->sc_journal_copying = 1;
1155 		GJ_DEBUG(1, "Starting copy of journal.");
1156 		g_journal_metadata_update(sc);
1157 	}
1158 	g_journal_copy_send(sc);
1159 }
1160 
1161 /*
1162  * Data block has been read from the journal provider.
1163  */
1164 static int
1165 g_journal_copy_read_done(struct bio *bp)
1166 {
1167 	struct g_journal_softc *sc;
1168 	struct g_consumer *cp;
1169 	struct bio *pbp;
1170 
1171 	KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1172 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1173 
1174 	sc = bp->bio_from->geom->softc;
1175 	pbp = bp->bio_parent;
1176 
1177 	if (bp->bio_error != 0) {
1178 		GJ_DEBUG(0, "Error while reading data from %s (error=%d).",
1179 		    bp->bio_to->name, bp->bio_error);
1180 		/*
1181 		 * We will not be able to deliver WRITE request as well.
1182 		 */
1183 		gj_free(bp->bio_data, bp->bio_length);
1184 		g_destroy_bio(pbp);
1185 		g_destroy_bio(bp);
1186 		sc->sc_copy_in_progress--;
1187 		return (1);
1188 	}
1189 	pbp->bio_data = bp->bio_data;
1190 	cp = sc->sc_dconsumer;
1191 	g_io_request(pbp, cp);
1192 	GJ_LOGREQ(4, bp, "READ DONE");
1193 	g_destroy_bio(bp);
1194 	return (0);
1195 }
1196 
1197 /*
1198  * Data block has been written to the data provider.
1199  */
1200 static void
1201 g_journal_copy_write_done(struct bio *bp)
1202 {
1203 	struct g_journal_softc *sc;
1204 
1205 	KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1206 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1207 
1208 	sc = bp->bio_from->geom->softc;
1209 	sc->sc_copy_in_progress--;
1210 
1211 	if (bp->bio_error != 0) {
1212 		GJ_LOGREQ(0, bp, "[copy] Error while writing data (error=%d)",
1213 		    bp->bio_error);
1214 	}
1215 	GJQ_REMOVE(sc->sc_copy_queue, bp);
1216 	gj_free(bp->bio_data, bp->bio_length);
1217 	GJ_LOGREQ(4, bp, "DONE");
1218 	g_destroy_bio(bp);
1219 
1220 	if (sc->sc_copy_in_progress == 0) {
1221 		/*
1222 		 * This was the last write request for this journal.
1223 		 */
1224 		GJ_DEBUG(1, "Data has been copied.");
1225 		sc->sc_journal_copying = 0;
1226 	}
1227 }
1228 
1229 static void g_journal_flush_done(struct bio *bp);
1230 
1231 /*
1232  * Flush one record onto active journal provider.
1233  */
1234 static void
1235 g_journal_flush(struct g_journal_softc *sc)
1236 {
1237 	struct g_journal_record_header hdr;
1238 	struct g_journal_entry *ent;
1239 	struct g_provider *pp;
1240 	struct bio **bioq;
1241 	struct bio *bp, *fbp, *pbp;
1242 	off_t joffset;
1243 	u_char *data, hash[16];
1244 	MD5_CTX ctx;
1245 	u_int i;
1246 
1247 	if (sc->sc_current_count == 0)
1248 		return;
1249 
1250 	pp = sc->sc_jprovider;
1251 	GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1252 	joffset = sc->sc_journal_offset;
1253 
1254 	GJ_DEBUG(2, "Storing %d journal entries on %s at %jd.",
1255 	    sc->sc_current_count, pp->name, (intmax_t)joffset);
1256 
1257 	/*
1258 	 * Store 'journal id', so we know to which journal this record belongs.
1259 	 */
1260 	hdr.jrh_journal_id = sc->sc_journal_id;
1261 	/* Could be less than g_journal_record_entries if called due timeout. */
1262 	hdr.jrh_nentries = MIN(sc->sc_current_count, g_journal_record_entries);
1263 	strlcpy(hdr.jrh_magic, GJ_RECORD_HEADER_MAGIC, sizeof(hdr.jrh_magic));
1264 
1265 	bioq = &sc->sc_active.jj_queue;
1266 	GJQ_LAST(sc->sc_flush_queue, pbp);
1267 
1268 	fbp = g_alloc_bio();
1269 	fbp->bio_parent = NULL;
1270 	fbp->bio_cflags = GJ_BIO_JOURNAL;
1271 	fbp->bio_offset = -1;
1272 	fbp->bio_joffset = joffset;
1273 	fbp->bio_length = pp->sectorsize;
1274 	fbp->bio_cmd = BIO_WRITE;
1275 	fbp->bio_done = g_journal_std_done;
1276 	GJQ_INSERT_AFTER(sc->sc_flush_queue, fbp, pbp);
1277 	pbp = fbp;
1278 	fbp->bio_to = pp;
1279 	GJ_LOGREQ(4, fbp, "FLUSH_OUT");
1280 	joffset += pp->sectorsize;
1281 	sc->sc_flush_count++;
1282 	if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1283 		MD5Init(&ctx);
1284 
1285 	for (i = 0; i < hdr.jrh_nentries; i++) {
1286 		bp = sc->sc_current_queue;
1287 		KASSERT(bp != NULL, ("NULL bp"));
1288 		bp->bio_to = pp;
1289 		GJ_LOGREQ(4, bp, "FLUSHED");
1290 		sc->sc_current_queue = bp->bio_next;
1291 		bp->bio_next = NULL;
1292 		sc->sc_current_count--;
1293 
1294 		/* Add to the header. */
1295 		ent = &hdr.jrh_entries[i];
1296 		ent->je_offset = bp->bio_offset;
1297 		ent->je_joffset = joffset;
1298 		ent->je_length = bp->bio_length;
1299 
1300 		data = bp->bio_data;
1301 		if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1302 			MD5Update(&ctx, data, ent->je_length);
1303 		g_reset_bio(bp);
1304 		bp->bio_cflags = GJ_BIO_JOURNAL;
1305 		bp->bio_offset = ent->je_offset;
1306 		bp->bio_joffset = ent->je_joffset;
1307 		bp->bio_length = ent->je_length;
1308 		bp->bio_data = data;
1309 		bp->bio_cmd = BIO_WRITE;
1310 		bp->bio_done = g_journal_std_done;
1311 		GJQ_INSERT_AFTER(sc->sc_flush_queue, bp, pbp);
1312 		pbp = bp;
1313 		bp->bio_to = pp;
1314 		GJ_LOGREQ(4, bp, "FLUSH_OUT");
1315 		joffset += bp->bio_length;
1316 		sc->sc_flush_count++;
1317 
1318 		/*
1319 		 * Add request to the active sc_journal_queue queue.
1320 		 * This is our cache. After journal switch we don't have to
1321 		 * read the data from the inactive journal, because we keep
1322 		 * it in memory.
1323 		 */
1324 		g_journal_insert(bioq, ent->je_offset,
1325 		    ent->je_offset + ent->je_length, ent->je_joffset, data,
1326 		    M_NOWAIT);
1327 	}
1328 
1329 	/*
1330 	 * After all requests, store valid header.
1331 	 */
1332 	data = gj_malloc(pp->sectorsize, M_WAITOK);
1333 	if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1334 		MD5Final(hash, &ctx);
1335 		bcopy(hash, hdr.jrh_sum, sizeof(hdr.jrh_sum));
1336 	}
1337 	g_journal_record_header_encode(&hdr, data);
1338 	fbp->bio_data = data;
1339 
1340 	sc->sc_journal_offset = joffset;
1341 
1342 	g_journal_check_overflow(sc);
1343 }
1344 
1345 /*
1346  * Flush request finished.
1347  */
1348 static void
1349 g_journal_flush_done(struct bio *bp)
1350 {
1351 	struct g_journal_softc *sc;
1352 	struct g_consumer *cp;
1353 
1354 	KASSERT((bp->bio_cflags & GJ_BIO_MASK) == GJ_BIO_JOURNAL,
1355 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_JOURNAL));
1356 
1357 	cp = bp->bio_from;
1358 	sc = cp->geom->softc;
1359 	sc->sc_flush_in_progress--;
1360 
1361 	if (bp->bio_error != 0) {
1362 		GJ_LOGREQ(0, bp, "[flush] Error while writing data (error=%d)",
1363 		    bp->bio_error);
1364 	}
1365 	gj_free(bp->bio_data, bp->bio_length);
1366 	GJ_LOGREQ(4, bp, "DONE");
1367 	g_destroy_bio(bp);
1368 }
1369 
1370 static void g_journal_release_delayed(struct g_journal_softc *sc);
1371 
1372 static void
1373 g_journal_flush_send(struct g_journal_softc *sc)
1374 {
1375 	struct g_consumer *cp;
1376 	struct bio *bioq, *bp, *lbp;
1377 
1378 	cp = sc->sc_jconsumer;
1379 	bioq = lbp = NULL;
1380 	while (sc->sc_flush_in_progress < g_journal_parallel_flushes) {
1381 		/* Send one flush requests to the active journal. */
1382 		bp = GJQ_FIRST(sc->sc_flush_queue);
1383 		if (bp != NULL) {
1384 			GJQ_REMOVE(sc->sc_flush_queue, bp);
1385 			sc->sc_flush_count--;
1386 			bp->bio_offset = bp->bio_joffset;
1387 			bp->bio_joffset = 0;
1388 			sc->sc_flush_in_progress++;
1389 			GJQ_INSERT_AFTER(bioq, bp, lbp);
1390 			lbp = bp;
1391 		}
1392 		/* Try to release delayed requests. */
1393 		g_journal_release_delayed(sc);
1394 		/* If there are no requests to flush, leave. */
1395 		if (GJQ_FIRST(sc->sc_flush_queue) == NULL)
1396 			break;
1397 	}
1398 	if (g_journal_do_optimize)
1399 		sc->sc_flush_in_progress += g_journal_optimize(bioq);
1400 	while ((bp = GJQ_FIRST(bioq)) != NULL) {
1401 		GJQ_REMOVE(bioq, bp);
1402 		GJ_LOGREQ(3, bp, "Flush request send");
1403 		g_io_request(bp, cp);
1404 	}
1405 }
1406 
1407 static void
1408 g_journal_add_current(struct g_journal_softc *sc, struct bio *bp)
1409 {
1410 	int n;
1411 
1412 	GJ_LOGREQ(4, bp, "CURRENT %d", sc->sc_current_count);
1413 	n = g_journal_insert_bio(&sc->sc_current_queue, bp, M_WAITOK);
1414 	sc->sc_current_count += n;
1415 	n = g_journal_optimize(sc->sc_current_queue);
1416 	sc->sc_current_count += n;
1417 	/*
1418 	 * For requests which are added to the current queue we deliver
1419 	 * response immediately.
1420 	 */
1421 	bp->bio_completed = bp->bio_length;
1422 	g_io_deliver(bp, 0);
1423 	if (sc->sc_current_count >= g_journal_record_entries) {
1424 		/*
1425 		 * Let's flush one record onto active journal provider.
1426 		 */
1427 		g_journal_flush(sc);
1428 	}
1429 }
1430 
1431 static void
1432 g_journal_release_delayed(struct g_journal_softc *sc)
1433 {
1434 	struct bio *bp;
1435 
1436 	for (;;) {
1437 		/* The flush queue is full, exit. */
1438 		if (sc->sc_flush_count >= g_journal_accept_immediately)
1439 			return;
1440 		bp = bioq_takefirst(&sc->sc_delayed_queue);
1441 		if (bp == NULL)
1442 			return;
1443 		sc->sc_delayed_count--;
1444 		g_journal_add_current(sc, bp);
1445 	}
1446 }
1447 
1448 /*
1449  * Add I/O request to the current queue. If we have enough requests for one
1450  * journal record we flush them onto active journal provider.
1451  */
1452 static void
1453 g_journal_add_request(struct g_journal_softc *sc, struct bio *bp)
1454 {
1455 
1456 	/*
1457 	 * The flush queue is full, we need to delay the request.
1458 	 */
1459 	if (sc->sc_delayed_count > 0 ||
1460 	    sc->sc_flush_count >= g_journal_accept_immediately) {
1461 		GJ_LOGREQ(4, bp, "DELAYED");
1462 		bioq_insert_tail(&sc->sc_delayed_queue, bp);
1463 		sc->sc_delayed_count++;
1464 		return;
1465 	}
1466 
1467 	KASSERT(TAILQ_EMPTY(&sc->sc_delayed_queue.queue),
1468 	    ("DELAYED queue not empty."));
1469 	g_journal_add_current(sc, bp);
1470 }
1471 
1472 static void g_journal_read_done(struct bio *bp);
1473 
1474 /*
1475  * Try to find requested data in cache.
1476  */
1477 static struct bio *
1478 g_journal_read_find(struct bio *head, int sorted, struct bio *pbp, off_t ostart,
1479     off_t oend)
1480 {
1481 	off_t cstart, cend;
1482 	struct bio *bp;
1483 
1484 	GJQ_FOREACH(head, bp) {
1485 		if (bp->bio_offset == -1)
1486 			continue;
1487 		cstart = MAX(ostart, bp->bio_offset);
1488 		cend = MIN(oend, bp->bio_offset + bp->bio_length);
1489 		if (cend <= ostart)
1490 			continue;
1491 		else if (cstart >= oend) {
1492 			if (!sorted)
1493 				continue;
1494 			else {
1495 				bp = NULL;
1496 				break;
1497 			}
1498 		}
1499 		if (bp->bio_data == NULL)
1500 			break;
1501 		GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend,
1502 		    bp);
1503 		bcopy(bp->bio_data + cstart - bp->bio_offset,
1504 		    pbp->bio_data + cstart - pbp->bio_offset, cend - cstart);
1505 		pbp->bio_completed += cend - cstart;
1506 		if (pbp->bio_completed == pbp->bio_length) {
1507 			/*
1508 			 * Cool, the whole request was in cache, deliver happy
1509 			 * message.
1510 			 */
1511 			g_io_deliver(pbp, 0);
1512 			return (pbp);
1513 		}
1514 		break;
1515 	}
1516 	return (bp);
1517 }
1518 
1519 /*
1520  * This function is used for collecting data on read.
1521  * The complexity is because parts of the data can be stored in four different
1522  * places:
1523  * - in memory - the data not yet send to the active journal provider
1524  * - in the active journal
1525  * - in the inactive journal
1526  * - in the data provider
1527  */
1528 static void
1529 g_journal_read(struct g_journal_softc *sc, struct bio *pbp, off_t ostart,
1530     off_t oend)
1531 {
1532 	struct bio *bp, *nbp, *head;
1533 	off_t cstart, cend;
1534 	u_int i, sorted = 0;
1535 
1536 	GJ_DEBUG(3, "READ: (%jd, %jd)", ostart, oend);
1537 
1538 	cstart = cend = -1;
1539 	bp = NULL;
1540 	head = NULL;
1541 	for (i = 1; i <= 5; i++) {
1542 		switch (i) {
1543 		case 1:	/* Not-yet-send data. */
1544 			head = sc->sc_current_queue;
1545 			sorted = 1;
1546 			break;
1547 		case 2: /* Skip flush queue as they are also in active queue */
1548 			continue;
1549 		case 3:	/* Active journal. */
1550 			head = sc->sc_active.jj_queue;
1551 			sorted = 1;
1552 			break;
1553 		case 4:	/* Inactive journal. */
1554 			/*
1555 			 * XXX: Here could be a race with g_journal_lowmem().
1556 			 */
1557 			head = sc->sc_inactive.jj_queue;
1558 			sorted = 1;
1559 			break;
1560 		case 5:	/* In-flight to the data provider. */
1561 			head = sc->sc_copy_queue;
1562 			sorted = 0;
1563 			break;
1564 		default:
1565 			panic("gjournal %s: i=%d", __func__, i);
1566 		}
1567 		bp = g_journal_read_find(head, sorted, pbp, ostart, oend);
1568 		if (bp == pbp) { /* Got the whole request. */
1569 			GJ_DEBUG(2, "Got the whole request from %u.", i);
1570 			return;
1571 		} else if (bp != NULL) {
1572 			cstart = MAX(ostart, bp->bio_offset);
1573 			cend = MIN(oend, bp->bio_offset + bp->bio_length);
1574 			GJ_DEBUG(2, "Got part of the request from %u (%jd-%jd).",
1575 			    i, (intmax_t)cstart, (intmax_t)cend);
1576 			break;
1577 		}
1578 	}
1579 	if (bp != NULL) {
1580 		if (bp->bio_data == NULL) {
1581 			nbp = g_duplicate_bio(pbp);
1582 			nbp->bio_cflags = GJ_BIO_READ;
1583 			nbp->bio_data =
1584 			    pbp->bio_data + cstart - pbp->bio_offset;
1585 			nbp->bio_offset =
1586 			    bp->bio_joffset + cstart - bp->bio_offset;
1587 			nbp->bio_length = cend - cstart;
1588 			nbp->bio_done = g_journal_read_done;
1589 			g_io_request(nbp, sc->sc_jconsumer);
1590 		}
1591 		/*
1592 		 * If we don't have the whole request yet, call g_journal_read()
1593 		 * recursively.
1594 		 */
1595 		if (ostart < cstart)
1596 			g_journal_read(sc, pbp, ostart, cstart);
1597 		if (oend > cend)
1598 			g_journal_read(sc, pbp, cend, oend);
1599 	} else {
1600 		/*
1601 		 * No data in memory, no data in journal.
1602 		 * Its time for asking data provider.
1603 		 */
1604 		GJ_DEBUG(3, "READ(data): (%jd, %jd)", ostart, oend);
1605 		nbp = g_duplicate_bio(pbp);
1606 		nbp->bio_cflags = GJ_BIO_READ;
1607 		nbp->bio_data = pbp->bio_data + ostart - pbp->bio_offset;
1608 		nbp->bio_offset = ostart;
1609 		nbp->bio_length = oend - ostart;
1610 		nbp->bio_done = g_journal_read_done;
1611 		g_io_request(nbp, sc->sc_dconsumer);
1612 		/* We have the whole request, return here. */
1613 		return;
1614 	}
1615 }
1616 
1617 /*
1618  * Function responsible for handling finished READ requests.
1619  * Actually, g_std_done() could be used here, the only difference is that we
1620  * log error.
1621  */
1622 static void
1623 g_journal_read_done(struct bio *bp)
1624 {
1625 	struct bio *pbp;
1626 
1627 	KASSERT(bp->bio_cflags == GJ_BIO_READ,
1628 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_READ));
1629 
1630 	pbp = bp->bio_parent;
1631 	pbp->bio_inbed++;
1632 	pbp->bio_completed += bp->bio_length;
1633 
1634 	if (bp->bio_error != 0) {
1635 		if (pbp->bio_error == 0)
1636 			pbp->bio_error = bp->bio_error;
1637 		GJ_DEBUG(0, "Error while reading data from %s (error=%d).",
1638 		    bp->bio_to->name, bp->bio_error);
1639 	}
1640 	g_destroy_bio(bp);
1641 	if (pbp->bio_children == pbp->bio_inbed &&
1642 	    pbp->bio_completed == pbp->bio_length) {
1643 		/* We're done. */
1644 		g_io_deliver(pbp, 0);
1645 	}
1646 }
1647 
1648 /*
1649  * Deactive current journal and active next one.
1650  */
1651 static void
1652 g_journal_switch(struct g_journal_softc *sc)
1653 {
1654 	struct g_provider *pp;
1655 
1656 	if (JEMPTY(sc)) {
1657 		GJ_DEBUG(3, "No need for %s switch.", sc->sc_name);
1658 		pp = LIST_FIRST(&sc->sc_geom->provider);
1659 		if (!(sc->sc_flags & GJF_DEVICE_CLEAN) && pp->acw == 0) {
1660 			sc->sc_flags |= GJF_DEVICE_CLEAN;
1661 			GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
1662 			g_journal_metadata_update(sc);
1663 		}
1664 	} else {
1665 		GJ_DEBUG(3, "Switching journal %s.", sc->sc_geom->name);
1666 
1667 		pp = sc->sc_jprovider;
1668 
1669 		sc->sc_journal_previous_id = sc->sc_journal_id;
1670 
1671 		sc->sc_journal_id = sc->sc_journal_next_id;
1672 		sc->sc_journal_next_id = arc4random();
1673 
1674 		GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1675 
1676 		g_journal_write_header(sc);
1677 
1678 		sc->sc_inactive.jj_offset = sc->sc_active.jj_offset;
1679 		sc->sc_inactive.jj_queue = sc->sc_active.jj_queue;
1680 
1681 		sc->sc_active.jj_offset =
1682 		    sc->sc_journal_offset - pp->sectorsize;
1683 		sc->sc_active.jj_queue = NULL;
1684 
1685 		/*
1686 		 * Switch is done, start copying data from the (now) inactive
1687 		 * journal to the data provider.
1688 		 */
1689 		g_journal_copy_start(sc);
1690 	}
1691 	mtx_lock(&sc->sc_mtx);
1692 	sc->sc_flags &= ~GJF_DEVICE_SWITCH;
1693 	mtx_unlock(&sc->sc_mtx);
1694 }
1695 
1696 static void
1697 g_journal_initialize(struct g_journal_softc *sc)
1698 {
1699 
1700 	sc->sc_journal_id = arc4random();
1701 	sc->sc_journal_next_id = arc4random();
1702 	sc->sc_journal_previous_id = sc->sc_journal_id;
1703 	sc->sc_journal_offset = sc->sc_jstart;
1704 	sc->sc_inactive.jj_offset = sc->sc_jstart;
1705 	g_journal_write_header(sc);
1706 	sc->sc_active.jj_offset = sc->sc_jstart;
1707 }
1708 
1709 static void
1710 g_journal_mark_as_dirty(struct g_journal_softc *sc)
1711 {
1712 	const struct g_journal_desc *desc;
1713 	int i;
1714 
1715 	GJ_DEBUG(1, "Marking file system %s as dirty.", sc->sc_name);
1716 	for (i = 0; (desc = g_journal_filesystems[i]) != NULL; i++)
1717 		desc->jd_dirty(sc->sc_dconsumer);
1718 }
1719 
1720 /*
1721  * Function read record header from the given journal.
1722  * It is very simlar to g_read_data(9), but it doesn't allocate memory for bio
1723  * and data on every call.
1724  */
1725 static int
1726 g_journal_sync_read(struct g_consumer *cp, struct bio *bp, off_t offset,
1727     void *data)
1728 {
1729 	int error;
1730 
1731 	g_reset_bio(bp);
1732 	bp->bio_cmd = BIO_READ;
1733 	bp->bio_done = NULL;
1734 	bp->bio_offset = offset;
1735 	bp->bio_length = cp->provider->sectorsize;
1736 	bp->bio_data = data;
1737 	g_io_request(bp, cp);
1738 	error = biowait(bp, "gjs_read");
1739 	return (error);
1740 }
1741 
1742 #if 0
1743 /*
1744  * Function is called when we start the journal device and we detect that
1745  * one of the journals was not fully copied.
1746  * The purpose of this function is to read all records headers from journal
1747  * and placed them in the inactive queue, so we can start journal
1748  * synchronization process and the journal provider itself.
1749  * Design decision was taken to not synchronize the whole journal here as it
1750  * can take too much time. Reading headers only and delaying synchronization
1751  * process until after journal provider is started should be the best choice.
1752  */
1753 #endif
1754 
1755 static void
1756 g_journal_sync(struct g_journal_softc *sc)
1757 {
1758 	struct g_journal_record_header rhdr;
1759 	struct g_journal_entry *ent;
1760 	struct g_journal_header jhdr;
1761 	struct g_consumer *cp;
1762 	struct bio *bp, *fbp, *tbp;
1763 	off_t joffset, offset;
1764 	u_char *buf, sum[16];
1765 	uint64_t id;
1766 	MD5_CTX ctx;
1767 	int error, found, i;
1768 
1769 	found = 0;
1770 	fbp = NULL;
1771 	cp = sc->sc_jconsumer;
1772 	bp = g_alloc_bio();
1773 	buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
1774 	offset = joffset = sc->sc_inactive.jj_offset = sc->sc_journal_offset;
1775 
1776 	GJ_DEBUG(2, "Looking for termination at %jd.", (intmax_t)joffset);
1777 
1778 	/*
1779 	 * Read and decode first journal header.
1780 	 */
1781 	error = g_journal_sync_read(cp, bp, offset, buf);
1782 	if (error != 0) {
1783 		GJ_DEBUG(0, "Error while reading journal header from %s.",
1784 		    cp->provider->name);
1785 		goto end;
1786 	}
1787 	error = g_journal_header_decode(buf, &jhdr);
1788 	if (error != 0) {
1789 		GJ_DEBUG(0, "Cannot decode journal header from %s.",
1790 		    cp->provider->name);
1791 		goto end;
1792 	}
1793 	id = sc->sc_journal_id;
1794 	if (jhdr.jh_journal_id != sc->sc_journal_id) {
1795 		GJ_DEBUG(1, "Journal ID mismatch at %jd (0x%08x != 0x%08x).",
1796 		    (intmax_t)offset, (u_int)jhdr.jh_journal_id, (u_int)id);
1797 		goto end;
1798 	}
1799 	offset += cp->provider->sectorsize;
1800 	id = sc->sc_journal_next_id = jhdr.jh_journal_next_id;
1801 
1802 	for (;;) {
1803 		/*
1804 		 * If the biggest record won't fit, look for a record header or
1805 		 * journal header from the beginning.
1806 		 */
1807 		GJ_VALIDATE_OFFSET(offset, sc);
1808 		error = g_journal_sync_read(cp, bp, offset, buf);
1809 		if (error != 0) {
1810 			/*
1811 			 * Not good. Having an error while reading header
1812 			 * means, that we cannot read next headers and in
1813 			 * consequence we cannot find termination.
1814 			 */
1815 			GJ_DEBUG(0,
1816 			    "Error while reading record header from %s.",
1817 			    cp->provider->name);
1818 			break;
1819 		}
1820 
1821 		error = g_journal_record_header_decode(buf, &rhdr);
1822 		if (error != 0) {
1823 			GJ_DEBUG(2, "Not a record header at %jd (error=%d).",
1824 			    (intmax_t)offset, error);
1825 			/*
1826 			 * This is not a record header.
1827 			 * If we are lucky, this is next journal header.
1828 			 */
1829 			error = g_journal_header_decode(buf, &jhdr);
1830 			if (error != 0) {
1831 				GJ_DEBUG(1, "Not a journal header at %jd (error=%d).",
1832 				    (intmax_t)offset, error);
1833 				/*
1834 				 * Nope, this is not journal header, which
1835 				 * bascially means that journal is not
1836 				 * terminated properly.
1837 				 */
1838 				error = ENOENT;
1839 				break;
1840 			}
1841 			/*
1842 			 * Ok. This is header of _some_ journal. Now we need to
1843 			 * verify if this is header of the _next_ journal.
1844 			 */
1845 			if (jhdr.jh_journal_id != id) {
1846 				GJ_DEBUG(1, "Journal ID mismatch at %jd "
1847 				    "(0x%08x != 0x%08x).", (intmax_t)offset,
1848 				    (u_int)jhdr.jh_journal_id, (u_int)id);
1849 				error = ENOENT;
1850 				break;
1851 			}
1852 
1853 			/* Found termination. */
1854 			found++;
1855 			GJ_DEBUG(1, "Found termination at %jd (id=0x%08x).",
1856 			    (intmax_t)offset, (u_int)id);
1857 			sc->sc_active.jj_offset = offset;
1858 			sc->sc_journal_offset =
1859 			    offset + cp->provider->sectorsize;
1860 			sc->sc_journal_id = id;
1861 			id = sc->sc_journal_next_id = jhdr.jh_journal_next_id;
1862 
1863 			while ((tbp = fbp) != NULL) {
1864 				fbp = tbp->bio_next;
1865 				GJ_LOGREQ(3, tbp, "Adding request.");
1866 				g_journal_insert_bio(&sc->sc_inactive.jj_queue,
1867 				    tbp, M_WAITOK);
1868 			}
1869 
1870 			/* Skip journal's header. */
1871 			offset += cp->provider->sectorsize;
1872 			continue;
1873 		}
1874 
1875 		/* Skip record's header. */
1876 		offset += cp->provider->sectorsize;
1877 
1878 		/*
1879 		 * Add information about every record entry to the inactive
1880 		 * queue.
1881 		 */
1882 		if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1883 			MD5Init(&ctx);
1884 		for (i = 0; i < rhdr.jrh_nentries; i++) {
1885 			ent = &rhdr.jrh_entries[i];
1886 			GJ_DEBUG(3, "Insert entry: %jd %jd.",
1887 			    (intmax_t)ent->je_offset, (intmax_t)ent->je_length);
1888 			g_journal_insert(&fbp, ent->je_offset,
1889 			    ent->je_offset + ent->je_length, ent->je_joffset,
1890 			    NULL, M_WAITOK);
1891 			if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1892 				u_char *buf2;
1893 
1894 				/*
1895 				 * TODO: Should use faster function (like
1896 				 *       g_journal_sync_read()).
1897 				 */
1898 				buf2 = g_read_data(cp, offset, ent->je_length,
1899 				    NULL);
1900 				if (buf2 == NULL)
1901 					GJ_DEBUG(0, "Cannot read data at %jd.",
1902 					    (intmax_t)offset);
1903 				else {
1904 					MD5Update(&ctx, buf2, ent->je_length);
1905 					g_free(buf2);
1906 				}
1907 			}
1908 			/* Skip entry's data. */
1909 			offset += ent->je_length;
1910 		}
1911 		if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1912 			MD5Final(sum, &ctx);
1913 			if (bcmp(sum, rhdr.jrh_sum, sizeof(rhdr.jrh_sum)) != 0) {
1914 				GJ_DEBUG(0, "MD5 hash mismatch at %jd!",
1915 				    (intmax_t)offset);
1916 			}
1917 		}
1918 	}
1919 end:
1920 	gj_free(bp->bio_data, cp->provider->sectorsize);
1921 	g_destroy_bio(bp);
1922 
1923 	/* Remove bios from unterminated journal. */
1924 	while ((tbp = fbp) != NULL) {
1925 		fbp = tbp->bio_next;
1926 		g_destroy_bio(tbp);
1927 	}
1928 
1929 	if (found < 1 && joffset > 0) {
1930 		GJ_DEBUG(0, "Journal on %s is broken/corrupted. Initializing.",
1931 		    sc->sc_name);
1932 		while ((tbp = sc->sc_inactive.jj_queue) != NULL) {
1933 			sc->sc_inactive.jj_queue = tbp->bio_next;
1934 			g_destroy_bio(tbp);
1935 		}
1936 		g_journal_initialize(sc);
1937 		g_journal_mark_as_dirty(sc);
1938 	} else {
1939 		GJ_DEBUG(0, "Journal %s consistent.", sc->sc_name);
1940 		g_journal_copy_start(sc);
1941 	}
1942 }
1943 
1944 /*
1945  * Wait for requests.
1946  * If we have requests in the current queue, flush them after 3 seconds from the
1947  * last flush. In this way we don't wait forever (or for journal switch) with
1948  * storing not full records on journal.
1949  */
1950 static void
1951 g_journal_wait(struct g_journal_softc *sc, time_t last_write)
1952 {
1953 	int error, timeout;
1954 
1955 	GJ_DEBUG(3, "%s: enter", __func__);
1956 	if (sc->sc_current_count == 0) {
1957 		if (g_journal_debug < 2)
1958 			msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", 0);
1959 		else {
1960 			/*
1961 			 * If we have debug turned on, show number of elements
1962 			 * in various queues.
1963 			 */
1964 			for (;;) {
1965 				error = msleep(sc, &sc->sc_mtx, PRIBIO,
1966 				    "gj:work", hz * 3);
1967 				if (error == 0) {
1968 					mtx_unlock(&sc->sc_mtx);
1969 					break;
1970 				}
1971 				GJ_DEBUG(3, "Report: current count=%d",
1972 				    sc->sc_current_count);
1973 				GJ_DEBUG(3, "Report: flush count=%d",
1974 				    sc->sc_flush_count);
1975 				GJ_DEBUG(3, "Report: flush in progress=%d",
1976 				    sc->sc_flush_in_progress);
1977 				GJ_DEBUG(3, "Report: copy in progress=%d",
1978 				    sc->sc_copy_in_progress);
1979 				GJ_DEBUG(3, "Report: delayed=%d",
1980 				    sc->sc_delayed_count);
1981 			}
1982 		}
1983 		GJ_DEBUG(3, "%s: exit 1", __func__);
1984 		return;
1985 	}
1986 
1987 	/*
1988 	 * Flush even not full records every 3 seconds.
1989 	 */
1990 	timeout = (last_write + 3 - time_second) * hz;
1991 	if (timeout <= 0) {
1992 		mtx_unlock(&sc->sc_mtx);
1993 		g_journal_flush(sc);
1994 		g_journal_flush_send(sc);
1995 		GJ_DEBUG(3, "%s: exit 2", __func__);
1996 		return;
1997 	}
1998 	error = msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", timeout);
1999 	if (error == EWOULDBLOCK)
2000 		g_journal_flush_send(sc);
2001 	GJ_DEBUG(3, "%s: exit 3", __func__);
2002 }
2003 
2004 /*
2005  * Worker thread.
2006  */
2007 static void
2008 g_journal_worker(void *arg)
2009 {
2010 	struct g_journal_softc *sc;
2011 	struct g_geom *gp;
2012 	struct g_provider *pp;
2013 	struct bio *bp;
2014 	time_t last_write;
2015 	int type;
2016 
2017 	thread_lock(curthread);
2018 	sched_prio(curthread, PRIBIO);
2019 	thread_unlock(curthread);
2020 
2021 	sc = arg;
2022 	type = 0;	/* gcc */
2023 
2024 	if (sc->sc_flags & GJF_DEVICE_CLEAN) {
2025 		GJ_DEBUG(0, "Journal %s clean.", sc->sc_name);
2026 		g_journal_initialize(sc);
2027 	} else {
2028 		g_journal_sync(sc);
2029 	}
2030 	/*
2031 	 * Check if we can use BIO_FLUSH.
2032 	 */
2033 	sc->sc_bio_flush = 0;
2034 	if (g_io_flush(sc->sc_jconsumer) == 0) {
2035 		sc->sc_bio_flush |= GJ_FLUSH_JOURNAL;
2036 		GJ_DEBUG(1, "BIO_FLUSH supported by %s.",
2037 		    sc->sc_jconsumer->provider->name);
2038 	} else {
2039 		GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2040 		    sc->sc_jconsumer->provider->name);
2041 	}
2042 	if (sc->sc_jconsumer != sc->sc_dconsumer) {
2043 		if (g_io_flush(sc->sc_dconsumer) == 0) {
2044 			sc->sc_bio_flush |= GJ_FLUSH_DATA;
2045 			GJ_DEBUG(1, "BIO_FLUSH supported by %s.",
2046 			    sc->sc_dconsumer->provider->name);
2047 		} else {
2048 			GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2049 			    sc->sc_dconsumer->provider->name);
2050 		}
2051 	}
2052 
2053 	gp = sc->sc_geom;
2054 	g_topology_lock();
2055 	pp = g_new_providerf(gp, "%s.journal", sc->sc_name);
2056 	pp->mediasize = sc->sc_mediasize;
2057 	/*
2058 	 * There could be a problem when data provider and journal providers
2059 	 * have different sectorsize, but such scenario is prevented on journal
2060 	 * creation.
2061 	 */
2062 	pp->sectorsize = sc->sc_sectorsize;
2063 	g_error_provider(pp, 0);
2064 	g_topology_unlock();
2065 	last_write = time_second;
2066 
2067 	if (sc->sc_rootmount != NULL) {
2068 		GJ_DEBUG(1, "root_mount_rel %p", sc->sc_rootmount);
2069 		root_mount_rel(sc->sc_rootmount);
2070 		sc->sc_rootmount = NULL;
2071 	}
2072 
2073 	for (;;) {
2074 		/* Get first request from the queue. */
2075 		mtx_lock(&sc->sc_mtx);
2076 		bp = bioq_first(&sc->sc_back_queue);
2077 		if (bp != NULL)
2078 			type = (bp->bio_cflags & GJ_BIO_MASK);
2079 		if (bp == NULL) {
2080 			bp = bioq_first(&sc->sc_regular_queue);
2081 			if (bp != NULL)
2082 				type = GJ_BIO_REGULAR;
2083 		}
2084 		if (bp == NULL) {
2085 try_switch:
2086 			if ((sc->sc_flags & GJF_DEVICE_SWITCH) ||
2087 			    (sc->sc_flags & GJF_DEVICE_DESTROY)) {
2088 				if (sc->sc_current_count > 0) {
2089 					mtx_unlock(&sc->sc_mtx);
2090 					g_journal_flush(sc);
2091 					g_journal_flush_send(sc);
2092 					continue;
2093 				}
2094 				if (sc->sc_flush_in_progress > 0)
2095 					goto sleep;
2096 				if (sc->sc_copy_in_progress > 0)
2097 					goto sleep;
2098 			}
2099 			if (sc->sc_flags & GJF_DEVICE_SWITCH) {
2100 				mtx_unlock(&sc->sc_mtx);
2101 				g_journal_switch(sc);
2102 				wakeup(&sc->sc_journal_copying);
2103 				continue;
2104 			}
2105 			if (sc->sc_flags & GJF_DEVICE_DESTROY) {
2106 				GJ_DEBUG(1, "Shutting down worker "
2107 				    "thread for %s.", gp->name);
2108 				sc->sc_worker = NULL;
2109 				wakeup(&sc->sc_worker);
2110 				mtx_unlock(&sc->sc_mtx);
2111 				kproc_exit(0);
2112 			}
2113 sleep:
2114 			g_journal_wait(sc, last_write);
2115 			continue;
2116 		}
2117 		/*
2118 		 * If we're in switch process, we need to delay all new
2119 		 * write requests until its done.
2120 		 */
2121 		if ((sc->sc_flags & GJF_DEVICE_SWITCH) &&
2122 		    type == GJ_BIO_REGULAR && bp->bio_cmd == BIO_WRITE) {
2123 			GJ_LOGREQ(2, bp, "WRITE on SWITCH");
2124 			goto try_switch;
2125 		}
2126 		if (type == GJ_BIO_REGULAR)
2127 			bioq_remove(&sc->sc_regular_queue, bp);
2128 		else
2129 			bioq_remove(&sc->sc_back_queue, bp);
2130 		mtx_unlock(&sc->sc_mtx);
2131 		switch (type) {
2132 		case GJ_BIO_REGULAR:
2133 			/* Regular request. */
2134 			switch (bp->bio_cmd) {
2135 			case BIO_READ:
2136 				g_journal_read(sc, bp, bp->bio_offset,
2137 				    bp->bio_offset + bp->bio_length);
2138 				break;
2139 			case BIO_WRITE:
2140 				last_write = time_second;
2141 				g_journal_add_request(sc, bp);
2142 				g_journal_flush_send(sc);
2143 				break;
2144 			default:
2145 				panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2146 			}
2147 			break;
2148 		case GJ_BIO_COPY:
2149 			switch (bp->bio_cmd) {
2150 			case BIO_READ:
2151 				if (g_journal_copy_read_done(bp))
2152 					g_journal_copy_send(sc);
2153 				break;
2154 			case BIO_WRITE:
2155 				g_journal_copy_write_done(bp);
2156 				g_journal_copy_send(sc);
2157 				break;
2158 			default:
2159 				panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2160 			}
2161 			break;
2162 		case GJ_BIO_JOURNAL:
2163 			g_journal_flush_done(bp);
2164 			g_journal_flush_send(sc);
2165 			break;
2166 		case GJ_BIO_READ:
2167 		default:
2168 			panic("Invalid bio (%d).", type);
2169 		}
2170 	}
2171 }
2172 
2173 static void
2174 g_journal_destroy_event(void *arg, int flags __unused)
2175 {
2176 	struct g_journal_softc *sc;
2177 
2178 	g_topology_assert();
2179 	sc = arg;
2180 	g_journal_destroy(sc);
2181 }
2182 
2183 static void
2184 g_journal_timeout(void *arg)
2185 {
2186 	struct g_journal_softc *sc;
2187 
2188 	sc = arg;
2189 	GJ_DEBUG(0, "Timeout. Journal %s cannot be completed.",
2190 	    sc->sc_geom->name);
2191 	g_post_event(g_journal_destroy_event, sc, M_NOWAIT, NULL);
2192 }
2193 
2194 static struct g_geom *
2195 g_journal_create(struct g_class *mp, struct g_provider *pp,
2196     const struct g_journal_metadata *md)
2197 {
2198 	struct g_journal_softc *sc;
2199 	struct g_geom *gp;
2200 	struct g_consumer *cp;
2201 	int error;
2202 
2203 	sc = NULL;	/* gcc */
2204 
2205 	g_topology_assert();
2206 	/*
2207 	 * There are two possibilities:
2208 	 * 1. Data and both journals are on the same provider.
2209 	 * 2. Data and journals are all on separated providers.
2210 	 */
2211 	/* Look for journal device with the same ID. */
2212 	LIST_FOREACH(gp, &mp->geom, geom) {
2213 		sc = gp->softc;
2214 		if (sc == NULL)
2215 			continue;
2216 		if (sc->sc_id == md->md_id)
2217 			break;
2218 	}
2219 	if (gp == NULL)
2220 		sc = NULL;
2221 	else if (sc != NULL && (sc->sc_type & md->md_type) != 0) {
2222 		GJ_DEBUG(1, "Journal device %u already configured.", sc->sc_id);
2223 		return (NULL);
2224 	}
2225 	if (md->md_type == 0 || (md->md_type & ~GJ_TYPE_COMPLETE) != 0) {
2226 		GJ_DEBUG(0, "Invalid type on %s.", pp->name);
2227 		return (NULL);
2228 	}
2229 	if (md->md_type & GJ_TYPE_DATA) {
2230 		GJ_DEBUG(0, "Journal %u: %s contains data.", md->md_id,
2231 		    pp->name);
2232 	}
2233 	if (md->md_type & GJ_TYPE_JOURNAL) {
2234 		GJ_DEBUG(0, "Journal %u: %s contains journal.", md->md_id,
2235 		    pp->name);
2236 	}
2237 
2238 	if (sc == NULL) {
2239 		/* Action geom. */
2240 		sc = malloc(sizeof(*sc), M_JOURNAL, M_WAITOK | M_ZERO);
2241 		sc->sc_id = md->md_id;
2242 		sc->sc_type = 0;
2243 		sc->sc_flags = 0;
2244 		sc->sc_worker = NULL;
2245 
2246 		gp = g_new_geomf(mp, "gjournal %u", sc->sc_id);
2247 		gp->start = g_journal_start;
2248 		gp->orphan = g_journal_orphan;
2249 		gp->access = g_journal_access;
2250 		gp->softc = sc;
2251 		gp->flags |= G_GEOM_VOLATILE_BIO;
2252 		sc->sc_geom = gp;
2253 
2254 		mtx_init(&sc->sc_mtx, "gjournal", NULL, MTX_DEF);
2255 
2256 		bioq_init(&sc->sc_back_queue);
2257 		bioq_init(&sc->sc_regular_queue);
2258 		bioq_init(&sc->sc_delayed_queue);
2259 		sc->sc_delayed_count = 0;
2260 		sc->sc_current_queue = NULL;
2261 		sc->sc_current_count = 0;
2262 		sc->sc_flush_queue = NULL;
2263 		sc->sc_flush_count = 0;
2264 		sc->sc_flush_in_progress = 0;
2265 		sc->sc_copy_queue = NULL;
2266 		sc->sc_copy_in_progress = 0;
2267 		sc->sc_inactive.jj_queue = NULL;
2268 		sc->sc_active.jj_queue = NULL;
2269 
2270 		sc->sc_rootmount = root_mount_hold("GJOURNAL");
2271 		GJ_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2272 
2273 		callout_init(&sc->sc_callout, 1);
2274 		if (md->md_type != GJ_TYPE_COMPLETE) {
2275 			/*
2276 			 * Journal and data are on separate providers.
2277 			 * At this point we have only one of them.
2278 			 * We setup a timeout in case the other part will not
2279 			 * appear, so we won't wait forever.
2280 			 */
2281 			callout_reset(&sc->sc_callout, 5 * hz,
2282 			    g_journal_timeout, sc);
2283 		}
2284 	}
2285 
2286 	/* Remember type of the data provider. */
2287 	if (md->md_type & GJ_TYPE_DATA)
2288 		sc->sc_orig_type = md->md_type;
2289 	sc->sc_type |= md->md_type;
2290 	cp = NULL;
2291 
2292 	if (md->md_type & GJ_TYPE_DATA) {
2293 		if (md->md_flags & GJ_FLAG_CLEAN)
2294 			sc->sc_flags |= GJF_DEVICE_CLEAN;
2295 		if (md->md_flags & GJ_FLAG_CHECKSUM)
2296 			sc->sc_flags |= GJF_DEVICE_CHECKSUM;
2297 		cp = g_new_consumer(gp);
2298 		error = g_attach(cp, pp);
2299 		KASSERT(error == 0, ("Cannot attach to %s (error=%d).",
2300 		    pp->name, error));
2301 		error = g_access(cp, 1, 1, 1);
2302 		if (error != 0) {
2303 			GJ_DEBUG(0, "Cannot access %s (error=%d).", pp->name,
2304 			    error);
2305 			g_journal_destroy(sc);
2306 			return (NULL);
2307 		}
2308 		sc->sc_dconsumer = cp;
2309 		sc->sc_mediasize = pp->mediasize - pp->sectorsize;
2310 		sc->sc_sectorsize = pp->sectorsize;
2311 		sc->sc_jstart = md->md_jstart;
2312 		sc->sc_jend = md->md_jend;
2313 		if (md->md_provider[0] != '\0')
2314 			sc->sc_flags |= GJF_DEVICE_HARDCODED;
2315 		sc->sc_journal_offset = md->md_joffset;
2316 		sc->sc_journal_id = md->md_jid;
2317 		sc->sc_journal_previous_id = md->md_jid;
2318 	}
2319 	if (md->md_type & GJ_TYPE_JOURNAL) {
2320 		if (cp == NULL) {
2321 			cp = g_new_consumer(gp);
2322 			error = g_attach(cp, pp);
2323 			KASSERT(error == 0, ("Cannot attach to %s (error=%d).",
2324 			    pp->name, error));
2325 			error = g_access(cp, 1, 1, 1);
2326 			if (error != 0) {
2327 				GJ_DEBUG(0, "Cannot access %s (error=%d).",
2328 				    pp->name, error);
2329 				g_journal_destroy(sc);
2330 				return (NULL);
2331 			}
2332 		} else {
2333 			/*
2334 			 * Journal is on the same provider as data, which means
2335 			 * that data provider ends where journal starts.
2336 			 */
2337 			sc->sc_mediasize = md->md_jstart;
2338 		}
2339 		sc->sc_jconsumer = cp;
2340 	}
2341 
2342 	/* Start switcher kproc if needed. */
2343 	if (g_journal_switcher_proc == NULL)
2344 		g_journal_start_switcher(mp);
2345 
2346 	if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE) {
2347 		/* Journal is not complete yet. */
2348 		return (gp);
2349 	} else {
2350 		/* Journal complete, cancel timeout. */
2351 		callout_drain(&sc->sc_callout);
2352 	}
2353 
2354 	error = kproc_create(g_journal_worker, sc, &sc->sc_worker, 0, 0,
2355 	    "g_journal %s", sc->sc_name);
2356 	if (error != 0) {
2357 		GJ_DEBUG(0, "Cannot create worker thread for %s.journal.",
2358 		    sc->sc_name);
2359 		g_journal_destroy(sc);
2360 		return (NULL);
2361 	}
2362 
2363 	return (gp);
2364 }
2365 
2366 static void
2367 g_journal_destroy_consumer(void *arg, int flags __unused)
2368 {
2369 	struct g_consumer *cp;
2370 
2371 	g_topology_assert();
2372 	cp = arg;
2373 	g_detach(cp);
2374 	g_destroy_consumer(cp);
2375 }
2376 
2377 static int
2378 g_journal_destroy(struct g_journal_softc *sc)
2379 {
2380 	struct g_geom *gp;
2381 	struct g_provider *pp;
2382 	struct g_consumer *cp;
2383 
2384 	g_topology_assert();
2385 
2386 	if (sc == NULL)
2387 		return (ENXIO);
2388 
2389 	gp = sc->sc_geom;
2390 	pp = LIST_FIRST(&gp->provider);
2391 	if (pp != NULL) {
2392 		if (pp->acr != 0 || pp->acw != 0 || pp->ace != 0) {
2393 			GJ_DEBUG(1, "Device %s is still open (r%dw%de%d).",
2394 			    pp->name, pp->acr, pp->acw, pp->ace);
2395 			return (EBUSY);
2396 		}
2397 		g_error_provider(pp, ENXIO);
2398 
2399 		g_journal_flush(sc);
2400 		g_journal_flush_send(sc);
2401 		g_journal_switch(sc);
2402 	}
2403 
2404 	sc->sc_flags |= (GJF_DEVICE_DESTROY | GJF_DEVICE_CLEAN);
2405 
2406 	g_topology_unlock();
2407 
2408 	if (sc->sc_rootmount != NULL) {
2409 		GJ_DEBUG(1, "root_mount_rel %p", sc->sc_rootmount);
2410 		root_mount_rel(sc->sc_rootmount);
2411 		sc->sc_rootmount = NULL;
2412 	}
2413 
2414 	callout_drain(&sc->sc_callout);
2415 	mtx_lock(&sc->sc_mtx);
2416 	wakeup(sc);
2417 	while (sc->sc_worker != NULL)
2418 		msleep(&sc->sc_worker, &sc->sc_mtx, PRIBIO, "gj:destroy", 0);
2419 	mtx_unlock(&sc->sc_mtx);
2420 
2421 	if (pp != NULL) {
2422 		GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
2423 		g_journal_metadata_update(sc);
2424 		g_topology_lock();
2425 		g_wither_provider(pp, ENXIO);
2426 	} else {
2427 		g_topology_lock();
2428 	}
2429 	mtx_destroy(&sc->sc_mtx);
2430 
2431 	if (sc->sc_current_count != 0) {
2432 		GJ_DEBUG(0, "Warning! Number of current requests %d.",
2433 		    sc->sc_current_count);
2434 	}
2435 
2436 	gp->softc = NULL;
2437 	LIST_FOREACH(cp, &gp->consumer, consumer) {
2438 		if (cp->acr + cp->acw + cp->ace > 0)
2439 			g_access(cp, -1, -1, -1);
2440 		/*
2441 		 * We keep all consumers open for writting, so if I'll detach
2442 		 * and destroy consumer here, I'll get providers for taste, so
2443 		 * journal will be started again.
2444 		 * Sending an event here, prevents this from happening.
2445 		 */
2446 		g_post_event(g_journal_destroy_consumer, cp, M_WAITOK, NULL);
2447 	}
2448 	g_wither_geom(gp, ENXIO);
2449 	free(sc, M_JOURNAL);
2450 	return (0);
2451 }
2452 
2453 static void
2454 g_journal_taste_orphan(struct g_consumer *cp)
2455 {
2456 
2457 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2458 	    cp->provider->name));
2459 }
2460 
2461 static struct g_geom *
2462 g_journal_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2463 {
2464 	struct g_journal_metadata md;
2465 	struct g_consumer *cp;
2466 	struct g_geom *gp;
2467 	int error;
2468 
2469 	g_topology_assert();
2470 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2471 	GJ_DEBUG(2, "Tasting %s.", pp->name);
2472 	if (pp->geom->class == mp)
2473 		return (NULL);
2474 
2475 	gp = g_new_geomf(mp, "journal:taste");
2476 	/* This orphan function should be never called. */
2477 	gp->orphan = g_journal_taste_orphan;
2478 	cp = g_new_consumer(gp);
2479 	g_attach(cp, pp);
2480 	error = g_journal_metadata_read(cp, &md);
2481 	g_detach(cp);
2482 	g_destroy_consumer(cp);
2483 	g_destroy_geom(gp);
2484 	if (error != 0)
2485 		return (NULL);
2486 	gp = NULL;
2487 
2488 	if (md.md_provider[0] != '\0' &&
2489 	    !g_compare_names(md.md_provider, pp->name))
2490 		return (NULL);
2491 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
2492 		return (NULL);
2493 	if (g_journal_debug >= 2)
2494 		journal_metadata_dump(&md);
2495 
2496 	gp = g_journal_create(mp, pp, &md);
2497 	return (gp);
2498 }
2499 
2500 static struct g_journal_softc *
2501 g_journal_find_device(struct g_class *mp, const char *name)
2502 {
2503 	struct g_journal_softc *sc;
2504 	struct g_geom *gp;
2505 	struct g_provider *pp;
2506 
2507 	if (strncmp(name, "/dev/", 5) == 0)
2508 		name += 5;
2509 	LIST_FOREACH(gp, &mp->geom, geom) {
2510 		sc = gp->softc;
2511 		if (sc == NULL)
2512 			continue;
2513 		if (sc->sc_flags & GJF_DEVICE_DESTROY)
2514 			continue;
2515 		if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2516 			continue;
2517 		pp = LIST_FIRST(&gp->provider);
2518 		if (strcmp(sc->sc_name, name) == 0)
2519 			return (sc);
2520 		if (pp != NULL && strcmp(pp->name, name) == 0)
2521 			return (sc);
2522 	}
2523 	return (NULL);
2524 }
2525 
2526 static void
2527 g_journal_ctl_destroy(struct gctl_req *req, struct g_class *mp)
2528 {
2529 	struct g_journal_softc *sc;
2530 	const char *name;
2531 	char param[16];
2532 	int *nargs;
2533 	int error, i;
2534 
2535 	g_topology_assert();
2536 
2537 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
2538 	if (nargs == NULL) {
2539 		gctl_error(req, "No '%s' argument.", "nargs");
2540 		return;
2541 	}
2542 	if (*nargs <= 0) {
2543 		gctl_error(req, "Missing device(s).");
2544 		return;
2545 	}
2546 
2547 	for (i = 0; i < *nargs; i++) {
2548 		snprintf(param, sizeof(param), "arg%d", i);
2549 		name = gctl_get_asciiparam(req, param);
2550 		if (name == NULL) {
2551 			gctl_error(req, "No 'arg%d' argument.", i);
2552 			return;
2553 		}
2554 		sc = g_journal_find_device(mp, name);
2555 		if (sc == NULL) {
2556 			gctl_error(req, "No such device: %s.", name);
2557 			return;
2558 		}
2559 		error = g_journal_destroy(sc);
2560 		if (error != 0) {
2561 			gctl_error(req, "Cannot destroy device %s (error=%d).",
2562 			    LIST_FIRST(&sc->sc_geom->provider)->name, error);
2563 			return;
2564 		}
2565 	}
2566 }
2567 
2568 static void
2569 g_journal_ctl_sync(struct gctl_req *req __unused, struct g_class *mp __unused)
2570 {
2571 
2572 	g_topology_assert();
2573 	g_topology_unlock();
2574 	g_journal_sync_requested++;
2575 	wakeup(&g_journal_switcher_state);
2576 	while (g_journal_sync_requested > 0)
2577 		tsleep(&g_journal_sync_requested, PRIBIO, "j:sreq", hz / 2);
2578 	g_topology_lock();
2579 }
2580 
2581 static void
2582 g_journal_config(struct gctl_req *req, struct g_class *mp, const char *verb)
2583 {
2584 	uint32_t *version;
2585 
2586 	g_topology_assert();
2587 
2588 	version = gctl_get_paraml(req, "version", sizeof(*version));
2589 	if (version == NULL) {
2590 		gctl_error(req, "No '%s' argument.", "version");
2591 		return;
2592 	}
2593 	if (*version != G_JOURNAL_VERSION) {
2594 		gctl_error(req, "Userland and kernel parts are out of sync.");
2595 		return;
2596 	}
2597 
2598 	if (strcmp(verb, "destroy") == 0 || strcmp(verb, "stop") == 0) {
2599 		g_journal_ctl_destroy(req, mp);
2600 		return;
2601 	} else if (strcmp(verb, "sync") == 0) {
2602 		g_journal_ctl_sync(req, mp);
2603 		return;
2604 	}
2605 
2606 	gctl_error(req, "Unknown verb.");
2607 }
2608 
2609 static void
2610 g_journal_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
2611     struct g_consumer *cp, struct g_provider *pp)
2612 {
2613 	struct g_journal_softc *sc;
2614 
2615 	g_topology_assert();
2616 
2617 	sc = gp->softc;
2618 	if (sc == NULL)
2619 		return;
2620 	if (pp != NULL) {
2621 		/* Nothing here. */
2622 	} else if (cp != NULL) {
2623 		int first = 1;
2624 
2625 		sbuf_printf(sb, "%s<Role>", indent);
2626 		if (cp == sc->sc_dconsumer) {
2627 			sbuf_cat(sb, "Data");
2628 			first = 0;
2629 		}
2630 		if (cp == sc->sc_jconsumer) {
2631 			if (!first)
2632 				sbuf_cat(sb, ",");
2633 			sbuf_cat(sb, "Journal");
2634 		}
2635 		sbuf_cat(sb, "</Role>\n");
2636 		if (cp == sc->sc_jconsumer) {
2637 			sbuf_printf(sb, "<Jstart>%jd</Jstart>\n",
2638 			    (intmax_t)sc->sc_jstart);
2639 			sbuf_printf(sb, "<Jend>%jd</Jend>\n",
2640 			    (intmax_t)sc->sc_jend);
2641 		}
2642 	} else {
2643 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
2644 	}
2645 }
2646 
2647 static eventhandler_tag g_journal_event_shutdown = NULL;
2648 static eventhandler_tag g_journal_event_lowmem = NULL;
2649 
2650 static void
2651 g_journal_shutdown(void *arg, int howto __unused)
2652 {
2653 	struct g_class *mp;
2654 	struct g_geom *gp, *gp2;
2655 
2656 	if (panicstr != NULL)
2657 		return;
2658 	mp = arg;
2659 	g_topology_lock();
2660 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
2661 		if (gp->softc == NULL)
2662 			continue;
2663 		GJ_DEBUG(0, "Shutting down geom %s.", gp->name);
2664 		g_journal_destroy(gp->softc);
2665 	}
2666 	g_topology_unlock();
2667 }
2668 
2669 /*
2670  * Free cached requests from inactive queue in case of low memory.
2671  * We free GJ_FREE_AT_ONCE elements at once.
2672  */
2673 #define	GJ_FREE_AT_ONCE	4
2674 static void
2675 g_journal_lowmem(void *arg, int howto __unused)
2676 {
2677 	struct g_journal_softc *sc;
2678 	struct g_class *mp;
2679 	struct g_geom *gp;
2680 	struct bio *bp;
2681 	u_int nfree = GJ_FREE_AT_ONCE;
2682 
2683 	g_journal_stats_low_mem++;
2684 	mp = arg;
2685 	g_topology_lock();
2686 	LIST_FOREACH(gp, &mp->geom, geom) {
2687 		sc = gp->softc;
2688 		if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY))
2689 			continue;
2690 		mtx_lock(&sc->sc_mtx);
2691 		for (bp = sc->sc_inactive.jj_queue; nfree > 0 && bp != NULL;
2692 		    nfree--, bp = bp->bio_next) {
2693 			/*
2694 			 * This is safe to free the bio_data, because:
2695 			 * 1. If bio_data is NULL it will be read from the
2696 			 *    inactive journal.
2697 			 * 2. If bp is sent down, it is first removed from the
2698 			 *    inactive queue, so it's impossible to free the
2699 			 *    data from under in-flight bio.
2700 			 * On the other hand, freeing elements from the active
2701 			 * queue, is not safe.
2702 			 */
2703 			if (bp->bio_data != NULL) {
2704 				GJ_DEBUG(2, "Freeing data from %s.",
2705 				    sc->sc_name);
2706 				gj_free(bp->bio_data, bp->bio_length);
2707 				bp->bio_data = NULL;
2708 			}
2709 		}
2710 		mtx_unlock(&sc->sc_mtx);
2711 		if (nfree == 0)
2712 			break;
2713 	}
2714 	g_topology_unlock();
2715 }
2716 
2717 static void g_journal_switcher(void *arg);
2718 
2719 static void
2720 g_journal_init(struct g_class *mp)
2721 {
2722 
2723 	/* Pick a conservative value if provided value sucks. */
2724 	if (g_journal_cache_divisor <= 0 ||
2725 	    (vm_kmem_size / g_journal_cache_divisor == 0)) {
2726 		g_journal_cache_divisor = 5;
2727 	}
2728 	if (g_journal_cache_limit > 0) {
2729 		g_journal_cache_limit = vm_kmem_size / g_journal_cache_divisor;
2730 		g_journal_cache_low =
2731 		    (g_journal_cache_limit / 100) * g_journal_cache_switch;
2732 	}
2733 	g_journal_event_shutdown = EVENTHANDLER_REGISTER(shutdown_post_sync,
2734 	    g_journal_shutdown, mp, EVENTHANDLER_PRI_FIRST);
2735 	if (g_journal_event_shutdown == NULL)
2736 		GJ_DEBUG(0, "Warning! Cannot register shutdown event.");
2737 	g_journal_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem,
2738 	    g_journal_lowmem, mp, EVENTHANDLER_PRI_FIRST);
2739 	if (g_journal_event_lowmem == NULL)
2740 		GJ_DEBUG(0, "Warning! Cannot register lowmem event.");
2741 }
2742 
2743 static void
2744 g_journal_fini(struct g_class *mp)
2745 {
2746 
2747 	if (g_journal_event_shutdown != NULL) {
2748 		EVENTHANDLER_DEREGISTER(shutdown_post_sync,
2749 		    g_journal_event_shutdown);
2750 	}
2751 	if (g_journal_event_lowmem != NULL)
2752 		EVENTHANDLER_DEREGISTER(vm_lowmem, g_journal_event_lowmem);
2753 	g_journal_stop_switcher();
2754 }
2755 
2756 DECLARE_GEOM_CLASS(g_journal_class, g_journal);
2757 
2758 static const struct g_journal_desc *
2759 g_journal_find_desc(const char *fstype)
2760 {
2761 	const struct g_journal_desc *desc;
2762 	int i;
2763 
2764 	for (desc = g_journal_filesystems[i = 0]; desc != NULL;
2765 	     desc = g_journal_filesystems[++i]) {
2766 		if (strcmp(desc->jd_fstype, fstype) == 0)
2767 			break;
2768 	}
2769 	return (desc);
2770 }
2771 
2772 static void
2773 g_journal_switch_wait(struct g_journal_softc *sc)
2774 {
2775 	struct bintime bt;
2776 
2777 	mtx_assert(&sc->sc_mtx, MA_OWNED);
2778 	if (g_journal_debug >= 2) {
2779 		if (sc->sc_flush_in_progress > 0) {
2780 			GJ_DEBUG(2, "%d requests flushing.",
2781 			    sc->sc_flush_in_progress);
2782 		}
2783 		if (sc->sc_copy_in_progress > 0) {
2784 			GJ_DEBUG(2, "%d requests copying.",
2785 			    sc->sc_copy_in_progress);
2786 		}
2787 		if (sc->sc_flush_count > 0) {
2788 			GJ_DEBUG(2, "%d requests to flush.",
2789 			    sc->sc_flush_count);
2790 		}
2791 		if (sc->sc_delayed_count > 0) {
2792 			GJ_DEBUG(2, "%d requests delayed.",
2793 			    sc->sc_delayed_count);
2794 		}
2795 	}
2796 	g_journal_stats_switches++;
2797 	if (sc->sc_copy_in_progress > 0)
2798 		g_journal_stats_wait_for_copy++;
2799 	GJ_TIMER_START(1, &bt);
2800 	sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2801 	sc->sc_flags |= GJF_DEVICE_SWITCH;
2802 	wakeup(sc);
2803 	while (sc->sc_flags & GJF_DEVICE_SWITCH) {
2804 		msleep(&sc->sc_journal_copying, &sc->sc_mtx, PRIBIO,
2805 		    "gj:switch", 0);
2806 	}
2807 	GJ_TIMER_STOP(1, &bt, "Switch time of %s", sc->sc_name);
2808 }
2809 
2810 static void
2811 g_journal_do_switch(struct g_class *classp)
2812 {
2813 	struct g_journal_softc *sc;
2814 	const struct g_journal_desc *desc;
2815 	struct g_geom *gp;
2816 	struct mount *mp;
2817 	struct bintime bt;
2818 	char *mountpoint;
2819 	int error, save;
2820 
2821 	g_topology_lock();
2822 	LIST_FOREACH(gp, &classp->geom, geom) {
2823 		sc = gp->softc;
2824 		if (sc == NULL)
2825 			continue;
2826 		if (sc->sc_flags & GJF_DEVICE_DESTROY)
2827 			continue;
2828 		if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2829 			continue;
2830 		mtx_lock(&sc->sc_mtx);
2831 		sc->sc_flags |= GJF_DEVICE_BEFORE_SWITCH;
2832 		mtx_unlock(&sc->sc_mtx);
2833 	}
2834 	g_topology_unlock();
2835 
2836 	mtx_lock(&mountlist_mtx);
2837 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2838 		if (mp->mnt_gjprovider == NULL)
2839 			continue;
2840 		if (mp->mnt_flag & MNT_RDONLY)
2841 			continue;
2842 		desc = g_journal_find_desc(mp->mnt_stat.f_fstypename);
2843 		if (desc == NULL)
2844 			continue;
2845 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
2846 			continue;
2847 		/* mtx_unlock(&mountlist_mtx) was done inside vfs_busy() */
2848 
2849 		g_topology_lock();
2850 		sc = g_journal_find_device(classp, mp->mnt_gjprovider);
2851 		g_topology_unlock();
2852 
2853 		if (sc == NULL) {
2854 			GJ_DEBUG(0, "Cannot find journal geom for %s.",
2855 			    mp->mnt_gjprovider);
2856 			goto next;
2857 		} else if (JEMPTY(sc)) {
2858 			mtx_lock(&sc->sc_mtx);
2859 			sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2860 			mtx_unlock(&sc->sc_mtx);
2861 			GJ_DEBUG(3, "No need for %s switch.", sc->sc_name);
2862 			goto next;
2863 		}
2864 
2865 		mountpoint = mp->mnt_stat.f_mntonname;
2866 
2867 		error = vn_start_write(NULL, &mp, V_WAIT);
2868 		if (error != 0) {
2869 			GJ_DEBUG(0, "vn_start_write(%s) failed (error=%d).",
2870 			    mountpoint, error);
2871 			goto next;
2872 		}
2873 
2874 		save = curthread_pflags_set(TDP_SYNCIO);
2875 
2876 		GJ_TIMER_START(1, &bt);
2877 		vfs_msync(mp, MNT_NOWAIT);
2878 		GJ_TIMER_STOP(1, &bt, "Msync time of %s", mountpoint);
2879 
2880 		GJ_TIMER_START(1, &bt);
2881 		error = VFS_SYNC(mp, MNT_NOWAIT);
2882 		if (error == 0)
2883 			GJ_TIMER_STOP(1, &bt, "Sync time of %s", mountpoint);
2884 		else {
2885 			GJ_DEBUG(0, "Cannot sync file system %s (error=%d).",
2886 			    mountpoint, error);
2887 		}
2888 
2889 		curthread_pflags_restore(save);
2890 
2891 		vn_finished_write(mp);
2892 
2893 		if (error != 0)
2894 			goto next;
2895 
2896 		/*
2897 		 * Send BIO_FLUSH before freezing the file system, so it can be
2898 		 * faster after the freeze.
2899 		 */
2900 		GJ_TIMER_START(1, &bt);
2901 		g_journal_flush_cache(sc);
2902 		GJ_TIMER_STOP(1, &bt, "BIO_FLUSH time of %s", sc->sc_name);
2903 
2904 		GJ_TIMER_START(1, &bt);
2905 		error = vfs_write_suspend(mp, VS_SKIP_UNMOUNT);
2906 		GJ_TIMER_STOP(1, &bt, "Suspend time of %s", mountpoint);
2907 		if (error != 0) {
2908 			GJ_DEBUG(0, "Cannot suspend file system %s (error=%d).",
2909 			    mountpoint, error);
2910 			goto next;
2911 		}
2912 
2913 		error = desc->jd_clean(mp);
2914 		if (error != 0)
2915 			goto next;
2916 
2917 		mtx_lock(&sc->sc_mtx);
2918 		g_journal_switch_wait(sc);
2919 		mtx_unlock(&sc->sc_mtx);
2920 
2921 		vfs_write_resume(mp, 0);
2922 next:
2923 		mtx_lock(&mountlist_mtx);
2924 		vfs_unbusy(mp);
2925 	}
2926 	mtx_unlock(&mountlist_mtx);
2927 
2928 	sc = NULL;
2929 	for (;;) {
2930 		g_topology_lock();
2931 		LIST_FOREACH(gp, &g_journal_class.geom, geom) {
2932 			sc = gp->softc;
2933 			if (sc == NULL)
2934 				continue;
2935 			mtx_lock(&sc->sc_mtx);
2936 			if ((sc->sc_type & GJ_TYPE_COMPLETE) == GJ_TYPE_COMPLETE &&
2937 			    !(sc->sc_flags & GJF_DEVICE_DESTROY) &&
2938 			    (sc->sc_flags & GJF_DEVICE_BEFORE_SWITCH)) {
2939 				break;
2940 			}
2941 			mtx_unlock(&sc->sc_mtx);
2942 			sc = NULL;
2943 		}
2944 		g_topology_unlock();
2945 		if (sc == NULL)
2946 			break;
2947 		mtx_assert(&sc->sc_mtx, MA_OWNED);
2948 		g_journal_switch_wait(sc);
2949 		mtx_unlock(&sc->sc_mtx);
2950 	}
2951 }
2952 
2953 static void
2954 g_journal_start_switcher(struct g_class *mp)
2955 {
2956 	int error;
2957 
2958 	g_topology_assert();
2959 	MPASS(g_journal_switcher_proc == NULL);
2960 	g_journal_switcher_state = GJ_SWITCHER_WORKING;
2961 	error = kproc_create(g_journal_switcher, mp, &g_journal_switcher_proc,
2962 	    0, 0, "g_journal switcher");
2963 	KASSERT(error == 0, ("Cannot create switcher thread."));
2964 }
2965 
2966 static void
2967 g_journal_stop_switcher(void)
2968 {
2969 	g_topology_assert();
2970 	MPASS(g_journal_switcher_proc != NULL);
2971 	g_journal_switcher_state = GJ_SWITCHER_DIE;
2972 	wakeup(&g_journal_switcher_state);
2973 	while (g_journal_switcher_state != GJ_SWITCHER_DIED)
2974 		tsleep(&g_journal_switcher_state, PRIBIO, "jfini:wait", hz / 5);
2975 	GJ_DEBUG(1, "Switcher died.");
2976 	g_journal_switcher_proc = NULL;
2977 }
2978 
2979 /*
2980  * TODO: Kill switcher thread on last geom destruction?
2981  */
2982 static void
2983 g_journal_switcher(void *arg)
2984 {
2985 	struct g_class *mp;
2986 	struct bintime bt;
2987 	int error;
2988 
2989 	mp = arg;
2990 	curthread->td_pflags |= TDP_NORUNNINGBUF;
2991 	for (;;) {
2992 		g_journal_switcher_wokenup = 0;
2993 		error = tsleep(&g_journal_switcher_state, PRIBIO, "jsw:wait",
2994 		    g_journal_switch_time * hz);
2995 		if (g_journal_switcher_state == GJ_SWITCHER_DIE) {
2996 			g_journal_switcher_state = GJ_SWITCHER_DIED;
2997 			GJ_DEBUG(1, "Switcher exiting.");
2998 			wakeup(&g_journal_switcher_state);
2999 			kproc_exit(0);
3000 		}
3001 		if (error == 0 && g_journal_sync_requested == 0) {
3002 			GJ_DEBUG(1, "Out of cache, force switch (used=%jd "
3003 			    "limit=%jd).", (intmax_t)g_journal_cache_used,
3004 			    (intmax_t)g_journal_cache_limit);
3005 		}
3006 		GJ_TIMER_START(1, &bt);
3007 		g_journal_do_switch(mp);
3008 		GJ_TIMER_STOP(1, &bt, "Entire switch time");
3009 		if (g_journal_sync_requested > 0) {
3010 			g_journal_sync_requested = 0;
3011 			wakeup(&g_journal_sync_requested);
3012 		}
3013 	}
3014 }
3015