xref: /freebsd/sys/geom/journal/g_journal.c (revision 74d9553e43cfafc29448d0bb836916aa21dea0de)
1 /*-
2  * Copyright (c) 2005-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/limits.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/bio.h>
38 #include <sys/sysctl.h>
39 #include <sys/malloc.h>
40 #include <sys/mount.h>
41 #include <sys/eventhandler.h>
42 #include <sys/proc.h>
43 #include <sys/kthread.h>
44 #include <sys/sched.h>
45 #include <sys/taskqueue.h>
46 #include <sys/vnode.h>
47 #include <sys/sbuf.h>
48 #ifdef GJ_MEMDEBUG
49 #include <sys/stack.h>
50 #include <sys/kdb.h>
51 #endif
52 #include <vm/vm.h>
53 #include <vm/vm_kern.h>
54 #include <geom/geom.h>
55 
56 #include <geom/journal/g_journal.h>
57 
58 FEATURE(geom_journal, "GEOM journaling support");
59 
60 /*
61  * On-disk journal format:
62  *
63  * JH - Journal header
64  * RH - Record header
65  *
66  * %%%%%% ****** +------+ +------+     ****** +------+     %%%%%%
67  * % JH % * RH * | Data | | Data | ... * RH * | Data | ... % JH % ...
68  * %%%%%% ****** +------+ +------+     ****** +------+     %%%%%%
69  *
70  */
71 
72 CTASSERT(sizeof(struct g_journal_header) <= 512);
73 CTASSERT(sizeof(struct g_journal_record_header) <= 512);
74 
75 static MALLOC_DEFINE(M_JOURNAL, "journal_data", "GEOM_JOURNAL Data");
76 static struct mtx g_journal_cache_mtx;
77 MTX_SYSINIT(g_journal_cache, &g_journal_cache_mtx, "cache usage", MTX_DEF);
78 
79 const struct g_journal_desc *g_journal_filesystems[] = {
80 	&g_journal_ufs,
81 	NULL
82 };
83 
84 SYSCTL_DECL(_kern_geom);
85 
86 int g_journal_debug = 0;
87 static u_int g_journal_switch_time = 10;
88 static u_int g_journal_force_switch = 70;
89 static u_int g_journal_parallel_flushes = 16;
90 static u_int g_journal_parallel_copies = 16;
91 static u_int g_journal_accept_immediately = 64;
92 static u_int g_journal_record_entries = GJ_RECORD_HEADER_NENTRIES;
93 static u_int g_journal_do_optimize = 1;
94 
95 static SYSCTL_NODE(_kern_geom, OID_AUTO, journal, CTLFLAG_RW, 0,
96     "GEOM_JOURNAL stuff");
97 SYSCTL_INT(_kern_geom_journal, OID_AUTO, debug, CTLFLAG_RWTUN, &g_journal_debug, 0,
98     "Debug level");
99 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, switch_time, CTLFLAG_RW,
100     &g_journal_switch_time, 0, "Switch journals every N seconds");
101 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, force_switch, CTLFLAG_RW,
102     &g_journal_force_switch, 0, "Force switch when journal is N% full");
103 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_flushes, CTLFLAG_RW,
104     &g_journal_parallel_flushes, 0,
105     "Number of flush I/O requests to send in parallel");
106 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, accept_immediately, CTLFLAG_RW,
107     &g_journal_accept_immediately, 0,
108     "Number of I/O requests accepted immediately");
109 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_copies, CTLFLAG_RW,
110     &g_journal_parallel_copies, 0,
111     "Number of copy I/O requests to send in parallel");
112 static int
113 g_journal_record_entries_sysctl(SYSCTL_HANDLER_ARGS)
114 {
115 	u_int entries;
116 	int error;
117 
118 	entries = g_journal_record_entries;
119 	error = sysctl_handle_int(oidp, &entries, 0, req);
120 	if (error != 0 || req->newptr == NULL)
121 		return (error);
122 	if (entries < 1 || entries > GJ_RECORD_HEADER_NENTRIES)
123 		return (EINVAL);
124 	g_journal_record_entries = entries;
125 	return (0);
126 }
127 SYSCTL_PROC(_kern_geom_journal, OID_AUTO, record_entries,
128     CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_record_entries_sysctl, "I",
129     "Maximum number of entires in one journal record");
130 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, optimize, CTLFLAG_RW,
131     &g_journal_do_optimize, 0, "Try to combine bios on flush and copy");
132 
133 static u_int g_journal_cache_used = 0;
134 static u_int g_journal_cache_limit = 64 * 1024 * 1024;
135 static u_int g_journal_cache_divisor = 2;
136 static u_int g_journal_cache_switch = 90;
137 static u_int g_journal_cache_misses = 0;
138 static u_int g_journal_cache_alloc_failures = 0;
139 static u_int g_journal_cache_low = 0;
140 
141 static SYSCTL_NODE(_kern_geom_journal, OID_AUTO, cache, CTLFLAG_RW, 0,
142     "GEOM_JOURNAL cache");
143 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, used, CTLFLAG_RD,
144     &g_journal_cache_used, 0, "Number of allocated bytes");
145 static int
146 g_journal_cache_limit_sysctl(SYSCTL_HANDLER_ARGS)
147 {
148 	u_int limit;
149 	int error;
150 
151 	limit = g_journal_cache_limit;
152 	error = sysctl_handle_int(oidp, &limit, 0, req);
153 	if (error != 0 || req->newptr == NULL)
154 		return (error);
155 	g_journal_cache_limit = limit;
156 	g_journal_cache_low = (limit / 100) * g_journal_cache_switch;
157 	return (0);
158 }
159 SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, limit,
160     CTLTYPE_UINT | CTLFLAG_RWTUN, NULL, 0, g_journal_cache_limit_sysctl, "I",
161     "Maximum number of allocated bytes");
162 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, divisor, CTLFLAG_RDTUN,
163     &g_journal_cache_divisor, 0,
164     "(kmem_size / kern.geom.journal.cache.divisor) == cache size");
165 static int
166 g_journal_cache_switch_sysctl(SYSCTL_HANDLER_ARGS)
167 {
168 	u_int cswitch;
169 	int error;
170 
171 	cswitch = g_journal_cache_switch;
172 	error = sysctl_handle_int(oidp, &cswitch, 0, req);
173 	if (error != 0 || req->newptr == NULL)
174 		return (error);
175 	if (cswitch > 100)
176 		return (EINVAL);
177 	g_journal_cache_switch = cswitch;
178 	g_journal_cache_low = (g_journal_cache_limit / 100) * cswitch;
179 	return (0);
180 }
181 SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, switch,
182     CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_cache_switch_sysctl, "I",
183     "Force switch when we hit this percent of cache use");
184 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, misses, CTLFLAG_RW,
185     &g_journal_cache_misses, 0, "Number of cache misses");
186 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, alloc_failures, CTLFLAG_RW,
187     &g_journal_cache_alloc_failures, 0, "Memory allocation failures");
188 
189 static u_long g_journal_stats_bytes_skipped = 0;
190 static u_long g_journal_stats_combined_ios = 0;
191 static u_long g_journal_stats_switches = 0;
192 static u_long g_journal_stats_wait_for_copy = 0;
193 static u_long g_journal_stats_journal_full = 0;
194 static u_long g_journal_stats_low_mem = 0;
195 
196 static SYSCTL_NODE(_kern_geom_journal, OID_AUTO, stats, CTLFLAG_RW, 0,
197     "GEOM_JOURNAL statistics");
198 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, skipped_bytes, CTLFLAG_RW,
199     &g_journal_stats_bytes_skipped, 0, "Number of skipped bytes");
200 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, combined_ios, CTLFLAG_RW,
201     &g_journal_stats_combined_ios, 0, "Number of combined I/O requests");
202 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, switches, CTLFLAG_RW,
203     &g_journal_stats_switches, 0, "Number of journal switches");
204 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, wait_for_copy, CTLFLAG_RW,
205     &g_journal_stats_wait_for_copy, 0, "Wait for journal copy on switch");
206 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, journal_full, CTLFLAG_RW,
207     &g_journal_stats_journal_full, 0,
208     "Number of times journal was almost full.");
209 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, low_mem, CTLFLAG_RW,
210     &g_journal_stats_low_mem, 0, "Number of times low_mem hook was called.");
211 
212 static g_taste_t g_journal_taste;
213 static g_ctl_req_t g_journal_config;
214 static g_dumpconf_t g_journal_dumpconf;
215 static g_init_t g_journal_init;
216 static g_fini_t g_journal_fini;
217 
218 struct g_class g_journal_class = {
219 	.name = G_JOURNAL_CLASS_NAME,
220 	.version = G_VERSION,
221 	.taste = g_journal_taste,
222 	.ctlreq = g_journal_config,
223 	.dumpconf = g_journal_dumpconf,
224 	.init = g_journal_init,
225 	.fini = g_journal_fini
226 };
227 
228 static int g_journal_destroy(struct g_journal_softc *sc);
229 static void g_journal_metadata_update(struct g_journal_softc *sc);
230 static void g_journal_start_switcher(struct g_class *mp);
231 static void g_journal_stop_switcher(void);
232 static void g_journal_switch_wait(struct g_journal_softc *sc);
233 
234 #define	GJ_SWITCHER_WORKING	0
235 #define	GJ_SWITCHER_DIE		1
236 #define	GJ_SWITCHER_DIED	2
237 static struct proc *g_journal_switcher_proc = NULL;
238 static int g_journal_switcher_state = GJ_SWITCHER_WORKING;
239 static int g_journal_switcher_wokenup = 0;
240 static int g_journal_sync_requested = 0;
241 
242 #ifdef GJ_MEMDEBUG
243 struct meminfo {
244 	size_t		mi_size;
245 	struct stack	mi_stack;
246 };
247 #endif
248 
249 /*
250  * We use our own malloc/realloc/free funtions, so we can collect statistics
251  * and force journal switch when we're running out of cache.
252  */
253 static void *
254 gj_malloc(size_t size, int flags)
255 {
256 	void *p;
257 #ifdef GJ_MEMDEBUG
258 	struct meminfo *mi;
259 #endif
260 
261 	mtx_lock(&g_journal_cache_mtx);
262 	if (g_journal_cache_limit > 0 && !g_journal_switcher_wokenup &&
263 	    g_journal_cache_used + size > g_journal_cache_low) {
264 		GJ_DEBUG(1, "No cache, waking up the switcher.");
265 		g_journal_switcher_wokenup = 1;
266 		wakeup(&g_journal_switcher_state);
267 	}
268 	if ((flags & M_NOWAIT) && g_journal_cache_limit > 0 &&
269 	    g_journal_cache_used + size > g_journal_cache_limit) {
270 		mtx_unlock(&g_journal_cache_mtx);
271 		g_journal_cache_alloc_failures++;
272 		return (NULL);
273 	}
274 	g_journal_cache_used += size;
275 	mtx_unlock(&g_journal_cache_mtx);
276 	flags &= ~M_NOWAIT;
277 #ifndef GJ_MEMDEBUG
278 	p = malloc(size, M_JOURNAL, flags | M_WAITOK);
279 #else
280 	mi = malloc(sizeof(*mi) + size, M_JOURNAL, flags | M_WAITOK);
281 	p = (u_char *)mi + sizeof(*mi);
282 	mi->mi_size = size;
283 	stack_save(&mi->mi_stack);
284 #endif
285 	return (p);
286 }
287 
288 static void
289 gj_free(void *p, size_t size)
290 {
291 #ifdef GJ_MEMDEBUG
292 	struct meminfo *mi;
293 #endif
294 
295 	KASSERT(p != NULL, ("p=NULL"));
296 	KASSERT(size > 0, ("size=0"));
297 	mtx_lock(&g_journal_cache_mtx);
298 	KASSERT(g_journal_cache_used >= size, ("Freeing too much?"));
299 	g_journal_cache_used -= size;
300 	mtx_unlock(&g_journal_cache_mtx);
301 #ifdef GJ_MEMDEBUG
302 	mi = p = (void *)((u_char *)p - sizeof(*mi));
303 	if (mi->mi_size != size) {
304 		printf("GJOURNAL: Size mismatch! %zu != %zu\n", size,
305 		    mi->mi_size);
306 		printf("GJOURNAL: Alloc backtrace:\n");
307 		stack_print(&mi->mi_stack);
308 		printf("GJOURNAL: Free backtrace:\n");
309 		kdb_backtrace();
310 	}
311 #endif
312 	free(p, M_JOURNAL);
313 }
314 
315 static void *
316 gj_realloc(void *p, size_t size, size_t oldsize)
317 {
318 	void *np;
319 
320 #ifndef GJ_MEMDEBUG
321 	mtx_lock(&g_journal_cache_mtx);
322 	g_journal_cache_used -= oldsize;
323 	g_journal_cache_used += size;
324 	mtx_unlock(&g_journal_cache_mtx);
325 	np = realloc(p, size, M_JOURNAL, M_WAITOK);
326 #else
327 	np = gj_malloc(size, M_WAITOK);
328 	bcopy(p, np, MIN(oldsize, size));
329 	gj_free(p, oldsize);
330 #endif
331 	return (np);
332 }
333 
334 static void
335 g_journal_check_overflow(struct g_journal_softc *sc)
336 {
337 	off_t length, used;
338 
339 	if ((sc->sc_active.jj_offset < sc->sc_inactive.jj_offset &&
340 	     sc->sc_journal_offset >= sc->sc_inactive.jj_offset) ||
341 	    (sc->sc_active.jj_offset > sc->sc_inactive.jj_offset &&
342 	     sc->sc_journal_offset >= sc->sc_inactive.jj_offset &&
343 	     sc->sc_journal_offset < sc->sc_active.jj_offset)) {
344 		panic("Journal overflow "
345 		    "(id = %u joffset=%jd active=%jd inactive=%jd)",
346 		    (unsigned)sc->sc_id,
347 		    (intmax_t)sc->sc_journal_offset,
348 		    (intmax_t)sc->sc_active.jj_offset,
349 		    (intmax_t)sc->sc_inactive.jj_offset);
350 	}
351 	if (sc->sc_active.jj_offset < sc->sc_inactive.jj_offset) {
352 		length = sc->sc_inactive.jj_offset - sc->sc_active.jj_offset;
353 		used = sc->sc_journal_offset - sc->sc_active.jj_offset;
354 	} else {
355 		length = sc->sc_jend - sc->sc_active.jj_offset;
356 		length += sc->sc_inactive.jj_offset - sc->sc_jstart;
357 		if (sc->sc_journal_offset >= sc->sc_active.jj_offset)
358 			used = sc->sc_journal_offset - sc->sc_active.jj_offset;
359 		else {
360 			used = sc->sc_jend - sc->sc_active.jj_offset;
361 			used += sc->sc_journal_offset - sc->sc_jstart;
362 		}
363 	}
364 	/* Already woken up? */
365 	if (g_journal_switcher_wokenup)
366 		return;
367 	/*
368 	 * If the active journal takes more than g_journal_force_switch precent
369 	 * of free journal space, we force journal switch.
370 	 */
371 	KASSERT(length > 0,
372 	    ("length=%jd used=%jd active=%jd inactive=%jd joffset=%jd",
373 	    (intmax_t)length, (intmax_t)used,
374 	    (intmax_t)sc->sc_active.jj_offset,
375 	    (intmax_t)sc->sc_inactive.jj_offset,
376 	    (intmax_t)sc->sc_journal_offset));
377 	if ((used * 100) / length > g_journal_force_switch) {
378 		g_journal_stats_journal_full++;
379 		GJ_DEBUG(1, "Journal %s %jd%% full, forcing journal switch.",
380 		    sc->sc_name, (used * 100) / length);
381 		mtx_lock(&g_journal_cache_mtx);
382 		g_journal_switcher_wokenup = 1;
383 		wakeup(&g_journal_switcher_state);
384 		mtx_unlock(&g_journal_cache_mtx);
385 	}
386 }
387 
388 static void
389 g_journal_orphan(struct g_consumer *cp)
390 {
391 	struct g_journal_softc *sc;
392 	char name[256];
393 	int error;
394 
395 	g_topology_assert();
396 	sc = cp->geom->softc;
397 	strlcpy(name, cp->provider->name, sizeof(name));
398 	GJ_DEBUG(0, "Lost provider %s.", name);
399 	if (sc == NULL)
400 		return;
401 	error = g_journal_destroy(sc);
402 	if (error == 0)
403 		GJ_DEBUG(0, "Journal %s destroyed.", name);
404 	else {
405 		GJ_DEBUG(0, "Cannot destroy journal %s (error=%d). "
406 		    "Destroy it manually after last close.", sc->sc_name,
407 		    error);
408 	}
409 }
410 
411 static int
412 g_journal_access(struct g_provider *pp, int acr, int acw, int ace)
413 {
414 	struct g_journal_softc *sc;
415 	int dcr, dcw, dce;
416 
417 	g_topology_assert();
418 	GJ_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name,
419 	    acr, acw, ace);
420 
421 	dcr = pp->acr + acr;
422 	dcw = pp->acw + acw;
423 	dce = pp->ace + ace;
424 
425 	sc = pp->geom->softc;
426 	if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY)) {
427 		if (acr <= 0 && acw <= 0 && ace <= 0)
428 			return (0);
429 		else
430 			return (ENXIO);
431 	}
432 	if (pp->acw == 0 && dcw > 0) {
433 		GJ_DEBUG(1, "Marking %s as dirty.", sc->sc_name);
434 		sc->sc_flags &= ~GJF_DEVICE_CLEAN;
435 		g_topology_unlock();
436 		g_journal_metadata_update(sc);
437 		g_topology_lock();
438 	} /* else if (pp->acw == 0 && dcw > 0 && JEMPTY(sc)) {
439 		GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
440 		sc->sc_flags |= GJF_DEVICE_CLEAN;
441 		g_topology_unlock();
442 		g_journal_metadata_update(sc);
443 		g_topology_lock();
444 	} */
445 	return (0);
446 }
447 
448 static void
449 g_journal_header_encode(struct g_journal_header *hdr, u_char *data)
450 {
451 
452 	bcopy(GJ_HEADER_MAGIC, data, sizeof(GJ_HEADER_MAGIC));
453 	data += sizeof(GJ_HEADER_MAGIC);
454 	le32enc(data, hdr->jh_journal_id);
455 	data += 4;
456 	le32enc(data, hdr->jh_journal_next_id);
457 }
458 
459 static int
460 g_journal_header_decode(const u_char *data, struct g_journal_header *hdr)
461 {
462 
463 	bcopy(data, hdr->jh_magic, sizeof(hdr->jh_magic));
464 	data += sizeof(hdr->jh_magic);
465 	if (bcmp(hdr->jh_magic, GJ_HEADER_MAGIC, sizeof(GJ_HEADER_MAGIC)) != 0)
466 		return (EINVAL);
467 	hdr->jh_journal_id = le32dec(data);
468 	data += 4;
469 	hdr->jh_journal_next_id = le32dec(data);
470 	return (0);
471 }
472 
473 static void
474 g_journal_flush_cache(struct g_journal_softc *sc)
475 {
476 	struct bintime bt;
477 	int error;
478 
479 	if (sc->sc_bio_flush == 0)
480 		return;
481 	GJ_TIMER_START(1, &bt);
482 	if (sc->sc_bio_flush & GJ_FLUSH_JOURNAL) {
483 		error = g_io_flush(sc->sc_jconsumer);
484 		GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.",
485 		    sc->sc_jconsumer->provider->name, error);
486 	}
487 	if (sc->sc_bio_flush & GJ_FLUSH_DATA) {
488 		/*
489 		 * TODO: This could be called in parallel with the
490 		 *       previous call.
491 		 */
492 		error = g_io_flush(sc->sc_dconsumer);
493 		GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.",
494 		    sc->sc_dconsumer->provider->name, error);
495 	}
496 	GJ_TIMER_STOP(1, &bt, "Cache flush time");
497 }
498 
499 static int
500 g_journal_write_header(struct g_journal_softc *sc)
501 {
502 	struct g_journal_header hdr;
503 	struct g_consumer *cp;
504 	u_char *buf;
505 	int error;
506 
507 	cp = sc->sc_jconsumer;
508 	buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
509 
510 	strlcpy(hdr.jh_magic, GJ_HEADER_MAGIC, sizeof(hdr.jh_magic));
511 	hdr.jh_journal_id = sc->sc_journal_id;
512 	hdr.jh_journal_next_id = sc->sc_journal_next_id;
513 	g_journal_header_encode(&hdr, buf);
514 	error = g_write_data(cp, sc->sc_journal_offset, buf,
515 	    cp->provider->sectorsize);
516 	/* if (error == 0) */
517 	sc->sc_journal_offset += cp->provider->sectorsize;
518 
519 	gj_free(buf, cp->provider->sectorsize);
520 	return (error);
521 }
522 
523 /*
524  * Every journal record has a header and data following it.
525  * Functions below are used to decode the header before storing it to
526  * little endian and to encode it after reading to system endianness.
527  */
528 static void
529 g_journal_record_header_encode(struct g_journal_record_header *hdr,
530     u_char *data)
531 {
532 	struct g_journal_entry *ent;
533 	u_int i;
534 
535 	bcopy(GJ_RECORD_HEADER_MAGIC, data, sizeof(GJ_RECORD_HEADER_MAGIC));
536 	data += sizeof(GJ_RECORD_HEADER_MAGIC);
537 	le32enc(data, hdr->jrh_journal_id);
538 	data += 8;
539 	le16enc(data, hdr->jrh_nentries);
540 	data += 2;
541 	bcopy(hdr->jrh_sum, data, sizeof(hdr->jrh_sum));
542 	data += 8;
543 	for (i = 0; i < hdr->jrh_nentries; i++) {
544 		ent = &hdr->jrh_entries[i];
545 		le64enc(data, ent->je_joffset);
546 		data += 8;
547 		le64enc(data, ent->je_offset);
548 		data += 8;
549 		le64enc(data, ent->je_length);
550 		data += 8;
551 	}
552 }
553 
554 static int
555 g_journal_record_header_decode(const u_char *data,
556     struct g_journal_record_header *hdr)
557 {
558 	struct g_journal_entry *ent;
559 	u_int i;
560 
561 	bcopy(data, hdr->jrh_magic, sizeof(hdr->jrh_magic));
562 	data += sizeof(hdr->jrh_magic);
563 	if (strcmp(hdr->jrh_magic, GJ_RECORD_HEADER_MAGIC) != 0)
564 		return (EINVAL);
565 	hdr->jrh_journal_id = le32dec(data);
566 	data += 8;
567 	hdr->jrh_nentries = le16dec(data);
568 	data += 2;
569 	if (hdr->jrh_nentries > GJ_RECORD_HEADER_NENTRIES)
570 		return (EINVAL);
571 	bcopy(data, hdr->jrh_sum, sizeof(hdr->jrh_sum));
572 	data += 8;
573 	for (i = 0; i < hdr->jrh_nentries; i++) {
574 		ent = &hdr->jrh_entries[i];
575 		ent->je_joffset = le64dec(data);
576 		data += 8;
577 		ent->je_offset = le64dec(data);
578 		data += 8;
579 		ent->je_length = le64dec(data);
580 		data += 8;
581 	}
582 	return (0);
583 }
584 
585 /*
586  * Function reads metadata from a provider (via the given consumer), decodes
587  * it to system endianness and verifies its correctness.
588  */
589 static int
590 g_journal_metadata_read(struct g_consumer *cp, struct g_journal_metadata *md)
591 {
592 	struct g_provider *pp;
593 	u_char *buf;
594 	int error;
595 
596 	g_topology_assert();
597 
598 	error = g_access(cp, 1, 0, 0);
599 	if (error != 0)
600 		return (error);
601 	pp = cp->provider;
602 	g_topology_unlock();
603 	/* Metadata is stored in last sector. */
604 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
605 	    &error);
606 	g_topology_lock();
607 	g_access(cp, -1, 0, 0);
608 	if (buf == NULL) {
609 		GJ_DEBUG(1, "Cannot read metadata from %s (error=%d).",
610 		    cp->provider->name, error);
611 		return (error);
612 	}
613 
614 	/* Decode metadata. */
615 	error = journal_metadata_decode(buf, md);
616 	g_free(buf);
617 	/* Is this is gjournal provider at all? */
618 	if (strcmp(md->md_magic, G_JOURNAL_MAGIC) != 0)
619 		return (EINVAL);
620 	/*
621 	 * Are we able to handle this version of metadata?
622 	 * We only maintain backward compatibility.
623 	 */
624 	if (md->md_version > G_JOURNAL_VERSION) {
625 		GJ_DEBUG(0,
626 		    "Kernel module is too old to handle metadata from %s.",
627 		    cp->provider->name);
628 		return (EINVAL);
629 	}
630 	/* Is checksum correct? */
631 	if (error != 0) {
632 		GJ_DEBUG(0, "MD5 metadata hash mismatch for provider %s.",
633 		    cp->provider->name);
634 		return (error);
635 	}
636 	return (0);
637 }
638 
639 /*
640  * Two functions below are responsible for updating metadata.
641  * Only metadata on the data provider is updated (we need to update
642  * information about active journal in there).
643  */
644 static void
645 g_journal_metadata_done(struct bio *bp)
646 {
647 
648 	/*
649 	 * There is not much we can do on error except informing about it.
650 	 */
651 	if (bp->bio_error != 0) {
652 		GJ_LOGREQ(0, bp, "Cannot update metadata (error=%d).",
653 		    bp->bio_error);
654 	} else {
655 		GJ_LOGREQ(2, bp, "Metadata updated.");
656 	}
657 	gj_free(bp->bio_data, bp->bio_length);
658 	g_destroy_bio(bp);
659 }
660 
661 static void
662 g_journal_metadata_update(struct g_journal_softc *sc)
663 {
664 	struct g_journal_metadata md;
665 	struct g_consumer *cp;
666 	struct bio *bp;
667 	u_char *sector;
668 
669 	cp = sc->sc_dconsumer;
670 	sector = gj_malloc(cp->provider->sectorsize, M_WAITOK);
671 	strlcpy(md.md_magic, G_JOURNAL_MAGIC, sizeof(md.md_magic));
672 	md.md_version = G_JOURNAL_VERSION;
673 	md.md_id = sc->sc_id;
674 	md.md_type = sc->sc_orig_type;
675 	md.md_jstart = sc->sc_jstart;
676 	md.md_jend = sc->sc_jend;
677 	md.md_joffset = sc->sc_inactive.jj_offset;
678 	md.md_jid = sc->sc_journal_previous_id;
679 	md.md_flags = 0;
680 	if (sc->sc_flags & GJF_DEVICE_CLEAN)
681 		md.md_flags |= GJ_FLAG_CLEAN;
682 
683 	if (sc->sc_flags & GJF_DEVICE_HARDCODED)
684 		strlcpy(md.md_provider, sc->sc_name, sizeof(md.md_provider));
685 	else
686 		bzero(md.md_provider, sizeof(md.md_provider));
687 	md.md_provsize = cp->provider->mediasize;
688 	journal_metadata_encode(&md, sector);
689 
690 	/*
691 	 * Flush the cache, so we know all data are on disk.
692 	 * We write here informations like "journal is consistent", so we need
693 	 * to be sure it is. Without BIO_FLUSH here, we can end up in situation
694 	 * where metadata is stored on disk, but not all data.
695 	 */
696 	g_journal_flush_cache(sc);
697 
698 	bp = g_alloc_bio();
699 	bp->bio_offset = cp->provider->mediasize - cp->provider->sectorsize;
700 	bp->bio_length = cp->provider->sectorsize;
701 	bp->bio_data = sector;
702 	bp->bio_cmd = BIO_WRITE;
703 	if (!(sc->sc_flags & GJF_DEVICE_DESTROY)) {
704 		bp->bio_done = g_journal_metadata_done;
705 		g_io_request(bp, cp);
706 	} else {
707 		bp->bio_done = NULL;
708 		g_io_request(bp, cp);
709 		biowait(bp, "gjmdu");
710 		g_journal_metadata_done(bp);
711 	}
712 
713 	/*
714 	 * Be sure metadata reached the disk.
715 	 */
716 	g_journal_flush_cache(sc);
717 }
718 
719 /*
720  * This is where the I/O request comes from the GEOM.
721  */
722 static void
723 g_journal_start(struct bio *bp)
724 {
725 	struct g_journal_softc *sc;
726 
727 	sc = bp->bio_to->geom->softc;
728 	GJ_LOGREQ(3, bp, "Request received.");
729 
730 	switch (bp->bio_cmd) {
731 	case BIO_READ:
732 	case BIO_WRITE:
733 		mtx_lock(&sc->sc_mtx);
734 		bioq_insert_tail(&sc->sc_regular_queue, bp);
735 		wakeup(sc);
736 		mtx_unlock(&sc->sc_mtx);
737 		return;
738 	case BIO_GETATTR:
739 		if (strcmp(bp->bio_attribute, "GJOURNAL::provider") == 0) {
740 			strlcpy(bp->bio_data, bp->bio_to->name, bp->bio_length);
741 			bp->bio_completed = strlen(bp->bio_to->name) + 1;
742 			g_io_deliver(bp, 0);
743 			return;
744 		}
745 		/* FALLTHROUGH */
746 	case BIO_DELETE:
747 	default:
748 		g_io_deliver(bp, EOPNOTSUPP);
749 		return;
750 	}
751 }
752 
753 static void
754 g_journal_std_done(struct bio *bp)
755 {
756 	struct g_journal_softc *sc;
757 
758 	sc = bp->bio_from->geom->softc;
759 	mtx_lock(&sc->sc_mtx);
760 	bioq_insert_tail(&sc->sc_back_queue, bp);
761 	wakeup(sc);
762 	mtx_unlock(&sc->sc_mtx);
763 }
764 
765 static struct bio *
766 g_journal_new_bio(off_t start, off_t end, off_t joffset, u_char *data,
767     int flags)
768 {
769 	struct bio *bp;
770 
771 	bp = g_alloc_bio();
772 	bp->bio_offset = start;
773 	bp->bio_joffset = joffset;
774 	bp->bio_length = end - start;
775 	bp->bio_cmd = BIO_WRITE;
776 	bp->bio_done = g_journal_std_done;
777 	if (data == NULL)
778 		bp->bio_data = NULL;
779 	else {
780 		bp->bio_data = gj_malloc(bp->bio_length, flags);
781 		if (bp->bio_data != NULL)
782 			bcopy(data, bp->bio_data, bp->bio_length);
783 	}
784 	return (bp);
785 }
786 
787 #define	g_journal_insert_bio(head, bp, flags)				\
788 	g_journal_insert((head), (bp)->bio_offset,			\
789 		(bp)->bio_offset + (bp)->bio_length, (bp)->bio_joffset,	\
790 		(bp)->bio_data, flags)
791 /*
792  * The function below does a lot more than just inserting bio to the queue.
793  * It keeps the queue sorted by offset and ensures that there are no doubled
794  * data (it combines bios where ranges overlap).
795  *
796  * The function returns the number of bios inserted (as bio can be splitted).
797  */
798 static int
799 g_journal_insert(struct bio **head, off_t nstart, off_t nend, off_t joffset,
800     u_char *data, int flags)
801 {
802 	struct bio *nbp, *cbp, *pbp;
803 	off_t cstart, cend;
804 	u_char *tmpdata;
805 	int n;
806 
807 	GJ_DEBUG(3, "INSERT(%p): (%jd, %jd, %jd)", *head, nstart, nend,
808 	    joffset);
809 	n = 0;
810 	pbp = NULL;
811 	GJQ_FOREACH(*head, cbp) {
812 		cstart = cbp->bio_offset;
813 		cend = cbp->bio_offset + cbp->bio_length;
814 
815 		if (nstart >= cend) {
816 			/*
817 			 *  +-------------+
818 			 *  |             |
819 			 *  |   current   |  +-------------+
820 			 *  |     bio     |  |             |
821 			 *  |             |  |     new     |
822 			 *  +-------------+  |     bio     |
823 			 *                   |             |
824 			 *                   +-------------+
825 			 */
826 			GJ_DEBUG(3, "INSERT(%p): 1", *head);
827 		} else if (nend <= cstart) {
828 			/*
829 			 *                   +-------------+
830 			 *                   |             |
831 			 *  +-------------+  |   current   |
832 			 *  |             |  |     bio     |
833 			 *  |     new     |  |             |
834 			 *  |     bio     |  +-------------+
835 			 *  |             |
836 			 *  +-------------+
837 			 */
838 			nbp = g_journal_new_bio(nstart, nend, joffset, data,
839 			    flags);
840 			if (pbp == NULL)
841 				*head = nbp;
842 			else
843 				pbp->bio_next = nbp;
844 			nbp->bio_next = cbp;
845 			n++;
846 			GJ_DEBUG(3, "INSERT(%p): 2 (nbp=%p pbp=%p)", *head, nbp,
847 			    pbp);
848 			goto end;
849 		} else if (nstart <= cstart && nend >= cend) {
850 			/*
851 			 *      +-------------+      +-------------+
852 			 *      | current bio |      | current bio |
853 			 *  +---+-------------+---+  +-------------+---+
854 			 *  |   |             |   |  |             |   |
855 			 *  |   |             |   |  |             |   |
856 			 *  |   +-------------+   |  +-------------+   |
857 			 *  |       new bio       |  |     new bio     |
858 			 *  +---------------------+  +-----------------+
859 			 *
860 			 *      +-------------+  +-------------+
861 			 *      | current bio |  | current bio |
862 			 *  +---+-------------+  +-------------+
863 			 *  |   |             |  |             |
864 			 *  |   |             |  |             |
865 			 *  |   +-------------+  +-------------+
866 			 *  |     new bio     |  |   new bio   |
867 			 *  +-----------------+  +-------------+
868 			 */
869 			g_journal_stats_bytes_skipped += cbp->bio_length;
870 			cbp->bio_offset = nstart;
871 			cbp->bio_joffset = joffset;
872 			cbp->bio_length = cend - nstart;
873 			if (cbp->bio_data != NULL) {
874 				gj_free(cbp->bio_data, cend - cstart);
875 				cbp->bio_data = NULL;
876 			}
877 			if (data != NULL) {
878 				cbp->bio_data = gj_malloc(cbp->bio_length,
879 				    flags);
880 				if (cbp->bio_data != NULL) {
881 					bcopy(data, cbp->bio_data,
882 					    cbp->bio_length);
883 				}
884 				data += cend - nstart;
885 			}
886 			joffset += cend - nstart;
887 			nstart = cend;
888 			GJ_DEBUG(3, "INSERT(%p): 3 (cbp=%p)", *head, cbp);
889 		} else if (nstart > cstart && nend >= cend) {
890 			/*
891 			 *  +-----------------+  +-------------+
892 			 *  |   current bio   |  | current bio |
893 			 *  |   +-------------+  |   +---------+---+
894 			 *  |   |             |  |   |         |   |
895 			 *  |   |             |  |   |         |   |
896 			 *  +---+-------------+  +---+---------+   |
897 			 *      |   new bio   |      |   new bio   |
898 			 *      +-------------+      +-------------+
899 			 */
900 			g_journal_stats_bytes_skipped += cend - nstart;
901 			nbp = g_journal_new_bio(nstart, cend, joffset, data,
902 			    flags);
903 			nbp->bio_next = cbp->bio_next;
904 			cbp->bio_next = nbp;
905 			cbp->bio_length = nstart - cstart;
906 			if (cbp->bio_data != NULL) {
907 				cbp->bio_data = gj_realloc(cbp->bio_data,
908 				    cbp->bio_length, cend - cstart);
909 			}
910 			if (data != NULL)
911 				data += cend - nstart;
912 			joffset += cend - nstart;
913 			nstart = cend;
914 			n++;
915 			GJ_DEBUG(3, "INSERT(%p): 4 (cbp=%p)", *head, cbp);
916 		} else if (nstart > cstart && nend < cend) {
917 			/*
918 			 *  +---------------------+
919 			 *  |     current bio     |
920 			 *  |   +-------------+   |
921 			 *  |   |             |   |
922 			 *  |   |             |   |
923 			 *  +---+-------------+---+
924 			 *      |   new bio   |
925 			 *      +-------------+
926 			 */
927 			g_journal_stats_bytes_skipped += nend - nstart;
928 			nbp = g_journal_new_bio(nstart, nend, joffset, data,
929 			    flags);
930 			nbp->bio_next = cbp->bio_next;
931 			cbp->bio_next = nbp;
932 			if (cbp->bio_data == NULL)
933 				tmpdata = NULL;
934 			else
935 				tmpdata = cbp->bio_data + nend - cstart;
936 			nbp = g_journal_new_bio(nend, cend,
937 			    cbp->bio_joffset + nend - cstart, tmpdata, flags);
938 			nbp->bio_next = ((struct bio *)cbp->bio_next)->bio_next;
939 			((struct bio *)cbp->bio_next)->bio_next = nbp;
940 			cbp->bio_length = nstart - cstart;
941 			if (cbp->bio_data != NULL) {
942 				cbp->bio_data = gj_realloc(cbp->bio_data,
943 				    cbp->bio_length, cend - cstart);
944 			}
945 			n += 2;
946 			GJ_DEBUG(3, "INSERT(%p): 5 (cbp=%p)", *head, cbp);
947 			goto end;
948 		} else if (nstart <= cstart && nend < cend) {
949 			/*
950 			 *  +-----------------+      +-------------+
951 			 *  |   current bio   |      | current bio |
952 			 *  +-------------+   |  +---+---------+   |
953 			 *  |             |   |  |   |         |   |
954 			 *  |             |   |  |   |         |   |
955 			 *  +-------------+---+  |   +---------+---+
956 			 *  |   new bio   |      |   new bio   |
957 			 *  +-------------+      +-------------+
958 			 */
959 			g_journal_stats_bytes_skipped += nend - nstart;
960 			nbp = g_journal_new_bio(nstart, nend, joffset, data,
961 			    flags);
962 			if (pbp == NULL)
963 				*head = nbp;
964 			else
965 				pbp->bio_next = nbp;
966 			nbp->bio_next = cbp;
967 			cbp->bio_offset = nend;
968 			cbp->bio_length = cend - nend;
969 			cbp->bio_joffset += nend - cstart;
970 			tmpdata = cbp->bio_data;
971 			if (tmpdata != NULL) {
972 				cbp->bio_data = gj_malloc(cbp->bio_length,
973 				    flags);
974 				if (cbp->bio_data != NULL) {
975 					bcopy(tmpdata + nend - cstart,
976 					    cbp->bio_data, cbp->bio_length);
977 				}
978 				gj_free(tmpdata, cend - cstart);
979 			}
980 			n++;
981 			GJ_DEBUG(3, "INSERT(%p): 6 (cbp=%p)", *head, cbp);
982 			goto end;
983 		}
984 		if (nstart == nend)
985 			goto end;
986 		pbp = cbp;
987 	}
988 	nbp = g_journal_new_bio(nstart, nend, joffset, data, flags);
989 	if (pbp == NULL)
990 		*head = nbp;
991 	else
992 		pbp->bio_next = nbp;
993 	nbp->bio_next = NULL;
994 	n++;
995 	GJ_DEBUG(3, "INSERT(%p): 8 (nbp=%p pbp=%p)", *head, nbp, pbp);
996 end:
997 	if (g_journal_debug >= 3) {
998 		GJQ_FOREACH(*head, cbp) {
999 			GJ_DEBUG(3, "ELEMENT: %p (%jd, %jd, %jd, %p)", cbp,
1000 			    (intmax_t)cbp->bio_offset,
1001 			    (intmax_t)cbp->bio_length,
1002 			    (intmax_t)cbp->bio_joffset, cbp->bio_data);
1003 		}
1004 		GJ_DEBUG(3, "INSERT(%p): DONE %d", *head, n);
1005 	}
1006 	return (n);
1007 }
1008 
1009 /*
1010  * The function combines neighbour bios trying to squeeze as much data as
1011  * possible into one bio.
1012  *
1013  * The function returns the number of bios combined (negative value).
1014  */
1015 static int
1016 g_journal_optimize(struct bio *head)
1017 {
1018 	struct bio *cbp, *pbp;
1019 	int n;
1020 
1021 	n = 0;
1022 	pbp = NULL;
1023 	GJQ_FOREACH(head, cbp) {
1024 		/* Skip bios which has to be read first. */
1025 		if (cbp->bio_data == NULL) {
1026 			pbp = NULL;
1027 			continue;
1028 		}
1029 		/* There is no previous bio yet. */
1030 		if (pbp == NULL) {
1031 			pbp = cbp;
1032 			continue;
1033 		}
1034 		/* Is this a neighbour bio? */
1035 		if (pbp->bio_offset + pbp->bio_length != cbp->bio_offset) {
1036 			/* Be sure that bios queue is sorted. */
1037 			KASSERT(pbp->bio_offset + pbp->bio_length < cbp->bio_offset,
1038 			    ("poffset=%jd plength=%jd coffset=%jd",
1039 			    (intmax_t)pbp->bio_offset,
1040 			    (intmax_t)pbp->bio_length,
1041 			    (intmax_t)cbp->bio_offset));
1042 			pbp = cbp;
1043 			continue;
1044 		}
1045 		/* Be sure we don't end up with too big bio. */
1046 		if (pbp->bio_length + cbp->bio_length > MAXPHYS) {
1047 			pbp = cbp;
1048 			continue;
1049 		}
1050 		/* Ok, we can join bios. */
1051 		GJ_LOGREQ(4, pbp, "Join: ");
1052 		GJ_LOGREQ(4, cbp, "and: ");
1053 		pbp->bio_data = gj_realloc(pbp->bio_data,
1054 		    pbp->bio_length + cbp->bio_length, pbp->bio_length);
1055 		bcopy(cbp->bio_data, pbp->bio_data + pbp->bio_length,
1056 		    cbp->bio_length);
1057 		gj_free(cbp->bio_data, cbp->bio_length);
1058 		pbp->bio_length += cbp->bio_length;
1059 		pbp->bio_next = cbp->bio_next;
1060 		g_destroy_bio(cbp);
1061 		cbp = pbp;
1062 		g_journal_stats_combined_ios++;
1063 		n--;
1064 		GJ_LOGREQ(4, pbp, "Got: ");
1065 	}
1066 	return (n);
1067 }
1068 
1069 /*
1070  * TODO: Update comment.
1071  * These are functions responsible for copying one portion of data from journal
1072  * to the destination provider.
1073  * The order goes like this:
1074  * 1. Read the header, which contains informations about data blocks
1075  *    following it.
1076  * 2. Read the data blocks from the journal.
1077  * 3. Write the data blocks on the data provider.
1078  *
1079  * g_journal_copy_start()
1080  * g_journal_copy_done() - got finished write request, logs potential errors.
1081  */
1082 
1083 /*
1084  * When there is no data in cache, this function is used to read it.
1085  */
1086 static void
1087 g_journal_read_first(struct g_journal_softc *sc, struct bio *bp)
1088 {
1089 	struct bio *cbp;
1090 
1091 	/*
1092 	 * We were short in memory, so data was freed.
1093 	 * In that case we need to read it back from journal.
1094 	 */
1095 	cbp = g_alloc_bio();
1096 	cbp->bio_cflags = bp->bio_cflags;
1097 	cbp->bio_parent = bp;
1098 	cbp->bio_offset = bp->bio_joffset;
1099 	cbp->bio_length = bp->bio_length;
1100 	cbp->bio_data = gj_malloc(bp->bio_length, M_WAITOK);
1101 	cbp->bio_cmd = BIO_READ;
1102 	cbp->bio_done = g_journal_std_done;
1103 	GJ_LOGREQ(4, cbp, "READ FIRST");
1104 	g_io_request(cbp, sc->sc_jconsumer);
1105 	g_journal_cache_misses++;
1106 }
1107 
1108 static void
1109 g_journal_copy_send(struct g_journal_softc *sc)
1110 {
1111 	struct bio *bioq, *bp, *lbp;
1112 
1113 	bioq = lbp = NULL;
1114 	mtx_lock(&sc->sc_mtx);
1115 	for (; sc->sc_copy_in_progress < g_journal_parallel_copies;) {
1116 		bp = GJQ_FIRST(sc->sc_inactive.jj_queue);
1117 		if (bp == NULL)
1118 			break;
1119 		GJQ_REMOVE(sc->sc_inactive.jj_queue, bp);
1120 		sc->sc_copy_in_progress++;
1121 		GJQ_INSERT_AFTER(bioq, bp, lbp);
1122 		lbp = bp;
1123 	}
1124 	mtx_unlock(&sc->sc_mtx);
1125 	if (g_journal_do_optimize)
1126 		sc->sc_copy_in_progress += g_journal_optimize(bioq);
1127 	while ((bp = GJQ_FIRST(bioq)) != NULL) {
1128 		GJQ_REMOVE(bioq, bp);
1129 		GJQ_INSERT_HEAD(sc->sc_copy_queue, bp);
1130 		bp->bio_cflags = GJ_BIO_COPY;
1131 		if (bp->bio_data == NULL)
1132 			g_journal_read_first(sc, bp);
1133 		else {
1134 			bp->bio_joffset = 0;
1135 			GJ_LOGREQ(4, bp, "SEND");
1136 			g_io_request(bp, sc->sc_dconsumer);
1137 		}
1138 	}
1139 }
1140 
1141 static void
1142 g_journal_copy_start(struct g_journal_softc *sc)
1143 {
1144 
1145 	/*
1146 	 * Remember in metadata that we're starting to copy journaled data
1147 	 * to the data provider.
1148 	 * In case of power failure, we will copy these data once again on boot.
1149 	 */
1150 	if (!sc->sc_journal_copying) {
1151 		sc->sc_journal_copying = 1;
1152 		GJ_DEBUG(1, "Starting copy of journal.");
1153 		g_journal_metadata_update(sc);
1154 	}
1155 	g_journal_copy_send(sc);
1156 }
1157 
1158 /*
1159  * Data block has been read from the journal provider.
1160  */
1161 static int
1162 g_journal_copy_read_done(struct bio *bp)
1163 {
1164 	struct g_journal_softc *sc;
1165 	struct g_consumer *cp;
1166 	struct bio *pbp;
1167 
1168 	KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1169 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1170 
1171 	sc = bp->bio_from->geom->softc;
1172 	pbp = bp->bio_parent;
1173 
1174 	if (bp->bio_error != 0) {
1175 		GJ_DEBUG(0, "Error while reading data from %s (error=%d).",
1176 		    bp->bio_to->name, bp->bio_error);
1177 		/*
1178 		 * We will not be able to deliver WRITE request as well.
1179 		 */
1180 		gj_free(bp->bio_data, bp->bio_length);
1181 		g_destroy_bio(pbp);
1182 		g_destroy_bio(bp);
1183 		sc->sc_copy_in_progress--;
1184 		return (1);
1185 	}
1186 	pbp->bio_data = bp->bio_data;
1187 	cp = sc->sc_dconsumer;
1188 	g_io_request(pbp, cp);
1189 	GJ_LOGREQ(4, bp, "READ DONE");
1190 	g_destroy_bio(bp);
1191 	return (0);
1192 }
1193 
1194 /*
1195  * Data block has been written to the data provider.
1196  */
1197 static void
1198 g_journal_copy_write_done(struct bio *bp)
1199 {
1200 	struct g_journal_softc *sc;
1201 
1202 	KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1203 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1204 
1205 	sc = bp->bio_from->geom->softc;
1206 	sc->sc_copy_in_progress--;
1207 
1208 	if (bp->bio_error != 0) {
1209 		GJ_LOGREQ(0, bp, "[copy] Error while writing data (error=%d)",
1210 		    bp->bio_error);
1211 	}
1212 	GJQ_REMOVE(sc->sc_copy_queue, bp);
1213 	gj_free(bp->bio_data, bp->bio_length);
1214 	GJ_LOGREQ(4, bp, "DONE");
1215 	g_destroy_bio(bp);
1216 
1217 	if (sc->sc_copy_in_progress == 0) {
1218 		/*
1219 		 * This was the last write request for this journal.
1220 		 */
1221 		GJ_DEBUG(1, "Data has been copied.");
1222 		sc->sc_journal_copying = 0;
1223 	}
1224 }
1225 
1226 static void g_journal_flush_done(struct bio *bp);
1227 
1228 /*
1229  * Flush one record onto active journal provider.
1230  */
1231 static void
1232 g_journal_flush(struct g_journal_softc *sc)
1233 {
1234 	struct g_journal_record_header hdr;
1235 	struct g_journal_entry *ent;
1236 	struct g_provider *pp;
1237 	struct bio **bioq;
1238 	struct bio *bp, *fbp, *pbp;
1239 	off_t joffset, size;
1240 	u_char *data, hash[16];
1241 	MD5_CTX ctx;
1242 	u_int i;
1243 
1244 	if (sc->sc_current_count == 0)
1245 		return;
1246 
1247 	size = 0;
1248 	pp = sc->sc_jprovider;
1249 	GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1250 	joffset = sc->sc_journal_offset;
1251 
1252 	GJ_DEBUG(2, "Storing %d journal entries on %s at %jd.",
1253 	    sc->sc_current_count, pp->name, (intmax_t)joffset);
1254 
1255 	/*
1256 	 * Store 'journal id', so we know to which journal this record belongs.
1257 	 */
1258 	hdr.jrh_journal_id = sc->sc_journal_id;
1259 	/* Could be less than g_journal_record_entries if called due timeout. */
1260 	hdr.jrh_nentries = MIN(sc->sc_current_count, g_journal_record_entries);
1261 	strlcpy(hdr.jrh_magic, GJ_RECORD_HEADER_MAGIC, sizeof(hdr.jrh_magic));
1262 
1263 	bioq = &sc->sc_active.jj_queue;
1264 	pbp = sc->sc_flush_queue;
1265 
1266 	fbp = g_alloc_bio();
1267 	fbp->bio_parent = NULL;
1268 	fbp->bio_cflags = GJ_BIO_JOURNAL;
1269 	fbp->bio_offset = -1;
1270 	fbp->bio_joffset = joffset;
1271 	fbp->bio_length = pp->sectorsize;
1272 	fbp->bio_cmd = BIO_WRITE;
1273 	fbp->bio_done = g_journal_std_done;
1274 	GJQ_INSERT_AFTER(sc->sc_flush_queue, fbp, pbp);
1275 	pbp = fbp;
1276 	fbp->bio_to = pp;
1277 	GJ_LOGREQ(4, fbp, "FLUSH_OUT");
1278 	joffset += pp->sectorsize;
1279 	sc->sc_flush_count++;
1280 	if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1281 		MD5Init(&ctx);
1282 
1283 	for (i = 0; i < hdr.jrh_nentries; i++) {
1284 		bp = sc->sc_current_queue;
1285 		KASSERT(bp != NULL, ("NULL bp"));
1286 		bp->bio_to = pp;
1287 		GJ_LOGREQ(4, bp, "FLUSHED");
1288 		sc->sc_current_queue = bp->bio_next;
1289 		bp->bio_next = NULL;
1290 		sc->sc_current_count--;
1291 
1292 		/* Add to the header. */
1293 		ent = &hdr.jrh_entries[i];
1294 		ent->je_offset = bp->bio_offset;
1295 		ent->je_joffset = joffset;
1296 		ent->je_length = bp->bio_length;
1297 		size += ent->je_length;
1298 
1299 		data = bp->bio_data;
1300 		if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1301 			MD5Update(&ctx, data, ent->je_length);
1302 		g_reset_bio(bp);
1303 		bp->bio_cflags = GJ_BIO_JOURNAL;
1304 		bp->bio_offset = ent->je_offset;
1305 		bp->bio_joffset = ent->je_joffset;
1306 		bp->bio_length = ent->je_length;
1307 		bp->bio_data = data;
1308 		bp->bio_cmd = BIO_WRITE;
1309 		bp->bio_done = g_journal_std_done;
1310 		GJQ_INSERT_AFTER(sc->sc_flush_queue, bp, pbp);
1311 		pbp = bp;
1312 		bp->bio_to = pp;
1313 		GJ_LOGREQ(4, bp, "FLUSH_OUT");
1314 		joffset += bp->bio_length;
1315 		sc->sc_flush_count++;
1316 
1317 		/*
1318 		 * Add request to the active sc_journal_queue queue.
1319 		 * This is our cache. After journal switch we don't have to
1320 		 * read the data from the inactive journal, because we keep
1321 		 * it in memory.
1322 		 */
1323 		g_journal_insert(bioq, ent->je_offset,
1324 		    ent->je_offset + ent->je_length, ent->je_joffset, data,
1325 		    M_NOWAIT);
1326 	}
1327 
1328 	/*
1329 	 * After all requests, store valid header.
1330 	 */
1331 	data = gj_malloc(pp->sectorsize, M_WAITOK);
1332 	if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1333 		MD5Final(hash, &ctx);
1334 		bcopy(hash, hdr.jrh_sum, sizeof(hdr.jrh_sum));
1335 	}
1336 	g_journal_record_header_encode(&hdr, data);
1337 	fbp->bio_data = data;
1338 
1339 	sc->sc_journal_offset = joffset;
1340 
1341 	g_journal_check_overflow(sc);
1342 }
1343 
1344 /*
1345  * Flush request finished.
1346  */
1347 static void
1348 g_journal_flush_done(struct bio *bp)
1349 {
1350 	struct g_journal_softc *sc;
1351 	struct g_consumer *cp;
1352 
1353 	KASSERT((bp->bio_cflags & GJ_BIO_MASK) == GJ_BIO_JOURNAL,
1354 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_JOURNAL));
1355 
1356 	cp = bp->bio_from;
1357 	sc = cp->geom->softc;
1358 	sc->sc_flush_in_progress--;
1359 
1360 	if (bp->bio_error != 0) {
1361 		GJ_LOGREQ(0, bp, "[flush] Error while writing data (error=%d)",
1362 		    bp->bio_error);
1363 	}
1364 	gj_free(bp->bio_data, bp->bio_length);
1365 	GJ_LOGREQ(4, bp, "DONE");
1366 	g_destroy_bio(bp);
1367 }
1368 
1369 static void g_journal_release_delayed(struct g_journal_softc *sc);
1370 
1371 static void
1372 g_journal_flush_send(struct g_journal_softc *sc)
1373 {
1374 	struct g_consumer *cp;
1375 	struct bio *bioq, *bp, *lbp;
1376 
1377 	cp = sc->sc_jconsumer;
1378 	bioq = lbp = NULL;
1379 	while (sc->sc_flush_in_progress < g_journal_parallel_flushes) {
1380 		/* Send one flush requests to the active journal. */
1381 		bp = GJQ_FIRST(sc->sc_flush_queue);
1382 		if (bp != NULL) {
1383 			GJQ_REMOVE(sc->sc_flush_queue, bp);
1384 			sc->sc_flush_count--;
1385 			bp->bio_offset = bp->bio_joffset;
1386 			bp->bio_joffset = 0;
1387 			sc->sc_flush_in_progress++;
1388 			GJQ_INSERT_AFTER(bioq, bp, lbp);
1389 			lbp = bp;
1390 		}
1391 		/* Try to release delayed requests. */
1392 		g_journal_release_delayed(sc);
1393 		/* If there are no requests to flush, leave. */
1394 		if (GJQ_FIRST(sc->sc_flush_queue) == NULL)
1395 			break;
1396 	}
1397 	if (g_journal_do_optimize)
1398 		sc->sc_flush_in_progress += g_journal_optimize(bioq);
1399 	while ((bp = GJQ_FIRST(bioq)) != NULL) {
1400 		GJQ_REMOVE(bioq, bp);
1401 		GJ_LOGREQ(3, bp, "Flush request send");
1402 		g_io_request(bp, cp);
1403 	}
1404 }
1405 
1406 static void
1407 g_journal_add_current(struct g_journal_softc *sc, struct bio *bp)
1408 {
1409 	int n;
1410 
1411 	GJ_LOGREQ(4, bp, "CURRENT %d", sc->sc_current_count);
1412 	n = g_journal_insert_bio(&sc->sc_current_queue, bp, M_WAITOK);
1413 	sc->sc_current_count += n;
1414 	n = g_journal_optimize(sc->sc_current_queue);
1415 	sc->sc_current_count += n;
1416 	/*
1417 	 * For requests which are added to the current queue we deliver
1418 	 * response immediately.
1419 	 */
1420 	bp->bio_completed = bp->bio_length;
1421 	g_io_deliver(bp, 0);
1422 	if (sc->sc_current_count >= g_journal_record_entries) {
1423 		/*
1424 		 * Let's flush one record onto active journal provider.
1425 		 */
1426 		g_journal_flush(sc);
1427 	}
1428 }
1429 
1430 static void
1431 g_journal_release_delayed(struct g_journal_softc *sc)
1432 {
1433 	struct bio *bp;
1434 
1435 	for (;;) {
1436 		/* The flush queue is full, exit. */
1437 		if (sc->sc_flush_count >= g_journal_accept_immediately)
1438 			return;
1439 		bp = bioq_takefirst(&sc->sc_delayed_queue);
1440 		if (bp == NULL)
1441 			return;
1442 		sc->sc_delayed_count--;
1443 		g_journal_add_current(sc, bp);
1444 	}
1445 }
1446 
1447 /*
1448  * Add I/O request to the current queue. If we have enough requests for one
1449  * journal record we flush them onto active journal provider.
1450  */
1451 static void
1452 g_journal_add_request(struct g_journal_softc *sc, struct bio *bp)
1453 {
1454 
1455 	/*
1456 	 * The flush queue is full, we need to delay the request.
1457 	 */
1458 	if (sc->sc_delayed_count > 0 ||
1459 	    sc->sc_flush_count >= g_journal_accept_immediately) {
1460 		GJ_LOGREQ(4, bp, "DELAYED");
1461 		bioq_insert_tail(&sc->sc_delayed_queue, bp);
1462 		sc->sc_delayed_count++;
1463 		return;
1464 	}
1465 
1466 	KASSERT(TAILQ_EMPTY(&sc->sc_delayed_queue.queue),
1467 	    ("DELAYED queue not empty."));
1468 	g_journal_add_current(sc, bp);
1469 }
1470 
1471 static void g_journal_read_done(struct bio *bp);
1472 
1473 /*
1474  * Try to find requested data in cache.
1475  */
1476 static struct bio *
1477 g_journal_read_find(struct bio *head, int sorted, struct bio *pbp, off_t ostart,
1478     off_t oend)
1479 {
1480 	off_t cstart, cend;
1481 	struct bio *bp;
1482 
1483 	GJQ_FOREACH(head, bp) {
1484 		if (bp->bio_offset == -1)
1485 			continue;
1486 		cstart = MAX(ostart, bp->bio_offset);
1487 		cend = MIN(oend, bp->bio_offset + bp->bio_length);
1488 		if (cend <= ostart)
1489 			continue;
1490 		else if (cstart >= oend) {
1491 			if (!sorted)
1492 				continue;
1493 			else {
1494 				bp = NULL;
1495 				break;
1496 			}
1497 		}
1498 		if (bp->bio_data == NULL)
1499 			break;
1500 		GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend,
1501 		    bp);
1502 		bcopy(bp->bio_data + cstart - bp->bio_offset,
1503 		    pbp->bio_data + cstart - pbp->bio_offset, cend - cstart);
1504 		pbp->bio_completed += cend - cstart;
1505 		if (pbp->bio_completed == pbp->bio_length) {
1506 			/*
1507 			 * Cool, the whole request was in cache, deliver happy
1508 			 * message.
1509 			 */
1510 			g_io_deliver(pbp, 0);
1511 			return (pbp);
1512 		}
1513 		break;
1514 	}
1515 	return (bp);
1516 }
1517 
1518 /*
1519  * Try to find requested data in cache.
1520  */
1521 static struct bio *
1522 g_journal_read_queue_find(struct bio_queue *head, struct bio *pbp, off_t ostart,
1523     off_t oend)
1524 {
1525 	off_t cstart, cend;
1526 	struct bio *bp;
1527 
1528 	TAILQ_FOREACH(bp, head, bio_queue) {
1529 		cstart = MAX(ostart, bp->bio_offset);
1530 		cend = MIN(oend, bp->bio_offset + bp->bio_length);
1531 		if (cend <= ostart)
1532 			continue;
1533 		else if (cstart >= oend)
1534 			continue;
1535 		KASSERT(bp->bio_data != NULL,
1536 		    ("%s: bio_data == NULL", __func__));
1537 		GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend,
1538 		    bp);
1539 		bcopy(bp->bio_data + cstart - bp->bio_offset,
1540 		    pbp->bio_data + cstart - pbp->bio_offset, cend - cstart);
1541 		pbp->bio_completed += cend - cstart;
1542 		if (pbp->bio_completed == pbp->bio_length) {
1543 			/*
1544 			 * Cool, the whole request was in cache, deliver happy
1545 			 * message.
1546 			 */
1547 			g_io_deliver(pbp, 0);
1548 			return (pbp);
1549 		}
1550 		break;
1551 	}
1552 	return (bp);
1553 }
1554 
1555 /*
1556  * This function is used for colecting data on read.
1557  * The complexity is because parts of the data can be stored in four different
1558  * places:
1559  * - in delayed requests
1560  * - in memory - the data not yet send to the active journal provider
1561  * - in requests which are going to be sent to the active journal
1562  * - in the active journal
1563  * - in the inactive journal
1564  * - in the data provider
1565  */
1566 static void
1567 g_journal_read(struct g_journal_softc *sc, struct bio *pbp, off_t ostart,
1568     off_t oend)
1569 {
1570 	struct bio *bp, *nbp, *head;
1571 	off_t cstart, cend;
1572 	u_int i, sorted = 0;
1573 
1574 	GJ_DEBUG(3, "READ: (%jd, %jd)", ostart, oend);
1575 
1576 	cstart = cend = -1;
1577 	bp = NULL;
1578 	head = NULL;
1579 	for (i = 0; i <= 5; i++) {
1580 		switch (i) {
1581 		case 0:	/* Delayed requests. */
1582 			head = NULL;
1583 			sorted = 0;
1584 			break;
1585 		case 1:	/* Not-yet-send data. */
1586 			head = sc->sc_current_queue;
1587 			sorted = 1;
1588 			break;
1589 		case 2:	/* In-flight to the active journal. */
1590 			head = sc->sc_flush_queue;
1591 			sorted = 0;
1592 			break;
1593 		case 3:	/* Active journal. */
1594 			head = sc->sc_active.jj_queue;
1595 			sorted = 1;
1596 			break;
1597 		case 4:	/* Inactive journal. */
1598 			/*
1599 			 * XXX: Here could be a race with g_journal_lowmem().
1600 			 */
1601 			head = sc->sc_inactive.jj_queue;
1602 			sorted = 1;
1603 			break;
1604 		case 5:	/* In-flight to the data provider. */
1605 			head = sc->sc_copy_queue;
1606 			sorted = 0;
1607 			break;
1608 		default:
1609 			panic("gjournal %s: i=%d", __func__, i);
1610 		}
1611 		if (i == 0)
1612 			bp = g_journal_read_queue_find(&sc->sc_delayed_queue.queue, pbp, ostart, oend);
1613 		else
1614 			bp = g_journal_read_find(head, sorted, pbp, ostart, oend);
1615 		if (bp == pbp) { /* Got the whole request. */
1616 			GJ_DEBUG(2, "Got the whole request from %u.", i);
1617 			return;
1618 		} else if (bp != NULL) {
1619 			cstart = MAX(ostart, bp->bio_offset);
1620 			cend = MIN(oend, bp->bio_offset + bp->bio_length);
1621 			GJ_DEBUG(2, "Got part of the request from %u (%jd-%jd).",
1622 			    i, (intmax_t)cstart, (intmax_t)cend);
1623 			break;
1624 		}
1625 	}
1626 	if (bp != NULL) {
1627 		if (bp->bio_data == NULL) {
1628 			nbp = g_duplicate_bio(pbp);
1629 			nbp->bio_cflags = GJ_BIO_READ;
1630 			nbp->bio_data =
1631 			    pbp->bio_data + cstart - pbp->bio_offset;
1632 			nbp->bio_offset =
1633 			    bp->bio_joffset + cstart - bp->bio_offset;
1634 			nbp->bio_length = cend - cstart;
1635 			nbp->bio_done = g_journal_read_done;
1636 			g_io_request(nbp, sc->sc_jconsumer);
1637 		}
1638 		/*
1639 		 * If we don't have the whole request yet, call g_journal_read()
1640 		 * recursively.
1641 		 */
1642 		if (ostart < cstart)
1643 			g_journal_read(sc, pbp, ostart, cstart);
1644 		if (oend > cend)
1645 			g_journal_read(sc, pbp, cend, oend);
1646 	} else {
1647 		/*
1648 		 * No data in memory, no data in journal.
1649 		 * Its time for asking data provider.
1650 		 */
1651 		GJ_DEBUG(3, "READ(data): (%jd, %jd)", ostart, oend);
1652 		nbp = g_duplicate_bio(pbp);
1653 		nbp->bio_cflags = GJ_BIO_READ;
1654 		nbp->bio_data = pbp->bio_data + ostart - pbp->bio_offset;
1655 		nbp->bio_offset = ostart;
1656 		nbp->bio_length = oend - ostart;
1657 		nbp->bio_done = g_journal_read_done;
1658 		g_io_request(nbp, sc->sc_dconsumer);
1659 		/* We have the whole request, return here. */
1660 		return;
1661 	}
1662 }
1663 
1664 /*
1665  * Function responsible for handling finished READ requests.
1666  * Actually, g_std_done() could be used here, the only difference is that we
1667  * log error.
1668  */
1669 static void
1670 g_journal_read_done(struct bio *bp)
1671 {
1672 	struct bio *pbp;
1673 
1674 	KASSERT(bp->bio_cflags == GJ_BIO_READ,
1675 	    ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_READ));
1676 
1677 	pbp = bp->bio_parent;
1678 	pbp->bio_inbed++;
1679 	pbp->bio_completed += bp->bio_length;
1680 
1681 	if (bp->bio_error != 0) {
1682 		if (pbp->bio_error == 0)
1683 			pbp->bio_error = bp->bio_error;
1684 		GJ_DEBUG(0, "Error while reading data from %s (error=%d).",
1685 		    bp->bio_to->name, bp->bio_error);
1686 	}
1687 	g_destroy_bio(bp);
1688 	if (pbp->bio_children == pbp->bio_inbed &&
1689 	    pbp->bio_completed == pbp->bio_length) {
1690 		/* We're done. */
1691 		g_io_deliver(pbp, 0);
1692 	}
1693 }
1694 
1695 /*
1696  * Deactive current journal and active next one.
1697  */
1698 static void
1699 g_journal_switch(struct g_journal_softc *sc)
1700 {
1701 	struct g_provider *pp;
1702 
1703 	if (JEMPTY(sc)) {
1704 		GJ_DEBUG(3, "No need for %s switch.", sc->sc_name);
1705 		pp = LIST_FIRST(&sc->sc_geom->provider);
1706 		if (!(sc->sc_flags & GJF_DEVICE_CLEAN) && pp->acw == 0) {
1707 			sc->sc_flags |= GJF_DEVICE_CLEAN;
1708 			GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
1709 			g_journal_metadata_update(sc);
1710 		}
1711 	} else {
1712 		GJ_DEBUG(3, "Switching journal %s.", sc->sc_geom->name);
1713 
1714 		pp = sc->sc_jprovider;
1715 
1716 		sc->sc_journal_previous_id = sc->sc_journal_id;
1717 
1718 		sc->sc_journal_id = sc->sc_journal_next_id;
1719 		sc->sc_journal_next_id = arc4random();
1720 
1721 		GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1722 
1723 		g_journal_write_header(sc);
1724 
1725 		sc->sc_inactive.jj_offset = sc->sc_active.jj_offset;
1726 		sc->sc_inactive.jj_queue = sc->sc_active.jj_queue;
1727 
1728 		sc->sc_active.jj_offset =
1729 		    sc->sc_journal_offset - pp->sectorsize;
1730 		sc->sc_active.jj_queue = NULL;
1731 
1732 		/*
1733 		 * Switch is done, start copying data from the (now) inactive
1734 		 * journal to the data provider.
1735 		 */
1736 		g_journal_copy_start(sc);
1737 	}
1738 	mtx_lock(&sc->sc_mtx);
1739 	sc->sc_flags &= ~GJF_DEVICE_SWITCH;
1740 	mtx_unlock(&sc->sc_mtx);
1741 }
1742 
1743 static void
1744 g_journal_initialize(struct g_journal_softc *sc)
1745 {
1746 
1747 	sc->sc_journal_id = arc4random();
1748 	sc->sc_journal_next_id = arc4random();
1749 	sc->sc_journal_previous_id = sc->sc_journal_id;
1750 	sc->sc_journal_offset = sc->sc_jstart;
1751 	sc->sc_inactive.jj_offset = sc->sc_jstart;
1752 	g_journal_write_header(sc);
1753 	sc->sc_active.jj_offset = sc->sc_jstart;
1754 }
1755 
1756 static void
1757 g_journal_mark_as_dirty(struct g_journal_softc *sc)
1758 {
1759 	const struct g_journal_desc *desc;
1760 	int i;
1761 
1762 	GJ_DEBUG(1, "Marking file system %s as dirty.", sc->sc_name);
1763 	for (i = 0; (desc = g_journal_filesystems[i]) != NULL; i++)
1764 		desc->jd_dirty(sc->sc_dconsumer);
1765 }
1766 
1767 /*
1768  * Function read record header from the given journal.
1769  * It is very simlar to g_read_data(9), but it doesn't allocate memory for bio
1770  * and data on every call.
1771  */
1772 static int
1773 g_journal_sync_read(struct g_consumer *cp, struct bio *bp, off_t offset,
1774     void *data)
1775 {
1776 	int error;
1777 
1778 	g_reset_bio(bp);
1779 	bp->bio_cmd = BIO_READ;
1780 	bp->bio_done = NULL;
1781 	bp->bio_offset = offset;
1782 	bp->bio_length = cp->provider->sectorsize;
1783 	bp->bio_data = data;
1784 	g_io_request(bp, cp);
1785 	error = biowait(bp, "gjs_read");
1786 	return (error);
1787 }
1788 
1789 #if 0
1790 /*
1791  * Function is called when we start the journal device and we detect that
1792  * one of the journals was not fully copied.
1793  * The purpose of this function is to read all records headers from journal
1794  * and placed them in the inactive queue, so we can start journal
1795  * synchronization process and the journal provider itself.
1796  * Design decision was taken to not synchronize the whole journal here as it
1797  * can take too much time. Reading headers only and delaying synchronization
1798  * process until after journal provider is started should be the best choice.
1799  */
1800 #endif
1801 
1802 static void
1803 g_journal_sync(struct g_journal_softc *sc)
1804 {
1805 	struct g_journal_record_header rhdr;
1806 	struct g_journal_entry *ent;
1807 	struct g_journal_header jhdr;
1808 	struct g_consumer *cp;
1809 	struct bio *bp, *fbp, *tbp;
1810 	off_t joffset, offset;
1811 	u_char *buf, sum[16];
1812 	uint64_t id;
1813 	MD5_CTX ctx;
1814 	int error, found, i;
1815 
1816 	found = 0;
1817 	fbp = NULL;
1818 	cp = sc->sc_jconsumer;
1819 	bp = g_alloc_bio();
1820 	buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
1821 	offset = joffset = sc->sc_inactive.jj_offset = sc->sc_journal_offset;
1822 
1823 	GJ_DEBUG(2, "Looking for termination at %jd.", (intmax_t)joffset);
1824 
1825 	/*
1826 	 * Read and decode first journal header.
1827 	 */
1828 	error = g_journal_sync_read(cp, bp, offset, buf);
1829 	if (error != 0) {
1830 		GJ_DEBUG(0, "Error while reading journal header from %s.",
1831 		    cp->provider->name);
1832 		goto end;
1833 	}
1834 	error = g_journal_header_decode(buf, &jhdr);
1835 	if (error != 0) {
1836 		GJ_DEBUG(0, "Cannot decode journal header from %s.",
1837 		    cp->provider->name);
1838 		goto end;
1839 	}
1840 	id = sc->sc_journal_id;
1841 	if (jhdr.jh_journal_id != sc->sc_journal_id) {
1842 		GJ_DEBUG(1, "Journal ID mismatch at %jd (0x%08x != 0x%08x).",
1843 		    (intmax_t)offset, (u_int)jhdr.jh_journal_id, (u_int)id);
1844 		goto end;
1845 	}
1846 	offset += cp->provider->sectorsize;
1847 	id = sc->sc_journal_next_id = jhdr.jh_journal_next_id;
1848 
1849 	for (;;) {
1850 		/*
1851 		 * If the biggest record won't fit, look for a record header or
1852 		 * journal header from the beginning.
1853 		 */
1854 		GJ_VALIDATE_OFFSET(offset, sc);
1855 		error = g_journal_sync_read(cp, bp, offset, buf);
1856 		if (error != 0) {
1857 			/*
1858 			 * Not good. Having an error while reading header
1859 			 * means, that we cannot read next headers and in
1860 			 * consequence we cannot find termination.
1861 			 */
1862 			GJ_DEBUG(0,
1863 			    "Error while reading record header from %s.",
1864 			    cp->provider->name);
1865 			break;
1866 		}
1867 
1868 		error = g_journal_record_header_decode(buf, &rhdr);
1869 		if (error != 0) {
1870 			GJ_DEBUG(2, "Not a record header at %jd (error=%d).",
1871 			    (intmax_t)offset, error);
1872 			/*
1873 			 * This is not a record header.
1874 			 * If we are lucky, this is next journal header.
1875 			 */
1876 			error = g_journal_header_decode(buf, &jhdr);
1877 			if (error != 0) {
1878 				GJ_DEBUG(1, "Not a journal header at %jd (error=%d).",
1879 				    (intmax_t)offset, error);
1880 				/*
1881 				 * Nope, this is not journal header, which
1882 				 * bascially means that journal is not
1883 				 * terminated properly.
1884 				 */
1885 				error = ENOENT;
1886 				break;
1887 			}
1888 			/*
1889 			 * Ok. This is header of _some_ journal. Now we need to
1890 			 * verify if this is header of the _next_ journal.
1891 			 */
1892 			if (jhdr.jh_journal_id != id) {
1893 				GJ_DEBUG(1, "Journal ID mismatch at %jd "
1894 				    "(0x%08x != 0x%08x).", (intmax_t)offset,
1895 				    (u_int)jhdr.jh_journal_id, (u_int)id);
1896 				error = ENOENT;
1897 				break;
1898 			}
1899 
1900 			/* Found termination. */
1901 			found++;
1902 			GJ_DEBUG(1, "Found termination at %jd (id=0x%08x).",
1903 			    (intmax_t)offset, (u_int)id);
1904 			sc->sc_active.jj_offset = offset;
1905 			sc->sc_journal_offset =
1906 			    offset + cp->provider->sectorsize;
1907 			sc->sc_journal_id = id;
1908 			id = sc->sc_journal_next_id = jhdr.jh_journal_next_id;
1909 
1910 			while ((tbp = fbp) != NULL) {
1911 				fbp = tbp->bio_next;
1912 				GJ_LOGREQ(3, tbp, "Adding request.");
1913 				g_journal_insert_bio(&sc->sc_inactive.jj_queue,
1914 				    tbp, M_WAITOK);
1915 			}
1916 
1917 			/* Skip journal's header. */
1918 			offset += cp->provider->sectorsize;
1919 			continue;
1920 		}
1921 
1922 		/* Skip record's header. */
1923 		offset += cp->provider->sectorsize;
1924 
1925 		/*
1926 		 * Add information about every record entry to the inactive
1927 		 * queue.
1928 		 */
1929 		if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1930 			MD5Init(&ctx);
1931 		for (i = 0; i < rhdr.jrh_nentries; i++) {
1932 			ent = &rhdr.jrh_entries[i];
1933 			GJ_DEBUG(3, "Insert entry: %jd %jd.",
1934 			    (intmax_t)ent->je_offset, (intmax_t)ent->je_length);
1935 			g_journal_insert(&fbp, ent->je_offset,
1936 			    ent->je_offset + ent->je_length, ent->je_joffset,
1937 			    NULL, M_WAITOK);
1938 			if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1939 				u_char *buf2;
1940 
1941 				/*
1942 				 * TODO: Should use faster function (like
1943 				 *       g_journal_sync_read()).
1944 				 */
1945 				buf2 = g_read_data(cp, offset, ent->je_length,
1946 				    NULL);
1947 				if (buf2 == NULL)
1948 					GJ_DEBUG(0, "Cannot read data at %jd.",
1949 					    (intmax_t)offset);
1950 				else {
1951 					MD5Update(&ctx, buf2, ent->je_length);
1952 					g_free(buf2);
1953 				}
1954 			}
1955 			/* Skip entry's data. */
1956 			offset += ent->je_length;
1957 		}
1958 		if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1959 			MD5Final(sum, &ctx);
1960 			if (bcmp(sum, rhdr.jrh_sum, sizeof(rhdr.jrh_sum)) != 0) {
1961 				GJ_DEBUG(0, "MD5 hash mismatch at %jd!",
1962 				    (intmax_t)offset);
1963 			}
1964 		}
1965 	}
1966 end:
1967 	gj_free(bp->bio_data, cp->provider->sectorsize);
1968 	g_destroy_bio(bp);
1969 
1970 	/* Remove bios from unterminated journal. */
1971 	while ((tbp = fbp) != NULL) {
1972 		fbp = tbp->bio_next;
1973 		g_destroy_bio(tbp);
1974 	}
1975 
1976 	if (found < 1 && joffset > 0) {
1977 		GJ_DEBUG(0, "Journal on %s is broken/corrupted. Initializing.",
1978 		    sc->sc_name);
1979 		while ((tbp = sc->sc_inactive.jj_queue) != NULL) {
1980 			sc->sc_inactive.jj_queue = tbp->bio_next;
1981 			g_destroy_bio(tbp);
1982 		}
1983 		g_journal_initialize(sc);
1984 		g_journal_mark_as_dirty(sc);
1985 	} else {
1986 		GJ_DEBUG(0, "Journal %s consistent.", sc->sc_name);
1987 		g_journal_copy_start(sc);
1988 	}
1989 }
1990 
1991 /*
1992  * Wait for requests.
1993  * If we have requests in the current queue, flush them after 3 seconds from the
1994  * last flush. In this way we don't wait forever (or for journal switch) with
1995  * storing not full records on journal.
1996  */
1997 static void
1998 g_journal_wait(struct g_journal_softc *sc, time_t last_write)
1999 {
2000 	int error, timeout;
2001 
2002 	GJ_DEBUG(3, "%s: enter", __func__);
2003 	if (sc->sc_current_count == 0) {
2004 		if (g_journal_debug < 2)
2005 			msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", 0);
2006 		else {
2007 			/*
2008 			 * If we have debug turned on, show number of elements
2009 			 * in various queues.
2010 			 */
2011 			for (;;) {
2012 				error = msleep(sc, &sc->sc_mtx, PRIBIO,
2013 				    "gj:work", hz * 3);
2014 				if (error == 0) {
2015 					mtx_unlock(&sc->sc_mtx);
2016 					break;
2017 				}
2018 				GJ_DEBUG(3, "Report: current count=%d",
2019 				    sc->sc_current_count);
2020 				GJ_DEBUG(3, "Report: flush count=%d",
2021 				    sc->sc_flush_count);
2022 				GJ_DEBUG(3, "Report: flush in progress=%d",
2023 				    sc->sc_flush_in_progress);
2024 				GJ_DEBUG(3, "Report: copy in progress=%d",
2025 				    sc->sc_copy_in_progress);
2026 				GJ_DEBUG(3, "Report: delayed=%d",
2027 				    sc->sc_delayed_count);
2028 			}
2029 		}
2030 		GJ_DEBUG(3, "%s: exit 1", __func__);
2031 		return;
2032 	}
2033 
2034 	/*
2035 	 * Flush even not full records every 3 seconds.
2036 	 */
2037 	timeout = (last_write + 3 - time_second) * hz;
2038 	if (timeout <= 0) {
2039 		mtx_unlock(&sc->sc_mtx);
2040 		g_journal_flush(sc);
2041 		g_journal_flush_send(sc);
2042 		GJ_DEBUG(3, "%s: exit 2", __func__);
2043 		return;
2044 	}
2045 	error = msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", timeout);
2046 	if (error == EWOULDBLOCK)
2047 		g_journal_flush_send(sc);
2048 	GJ_DEBUG(3, "%s: exit 3", __func__);
2049 }
2050 
2051 /*
2052  * Worker thread.
2053  */
2054 static void
2055 g_journal_worker(void *arg)
2056 {
2057 	struct g_journal_softc *sc;
2058 	struct g_geom *gp;
2059 	struct g_provider *pp;
2060 	struct bio *bp;
2061 	time_t last_write;
2062 	int type;
2063 
2064 	thread_lock(curthread);
2065 	sched_prio(curthread, PRIBIO);
2066 	thread_unlock(curthread);
2067 
2068 	sc = arg;
2069 	type = 0;	/* gcc */
2070 
2071 	if (sc->sc_flags & GJF_DEVICE_CLEAN) {
2072 		GJ_DEBUG(0, "Journal %s clean.", sc->sc_name);
2073 		g_journal_initialize(sc);
2074 	} else {
2075 		g_journal_sync(sc);
2076 	}
2077 	/*
2078 	 * Check if we can use BIO_FLUSH.
2079 	 */
2080 	sc->sc_bio_flush = 0;
2081 	if (g_io_flush(sc->sc_jconsumer) == 0) {
2082 		sc->sc_bio_flush |= GJ_FLUSH_JOURNAL;
2083 		GJ_DEBUG(1, "BIO_FLUSH supported by %s.",
2084 		    sc->sc_jconsumer->provider->name);
2085 	} else {
2086 		GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2087 		    sc->sc_jconsumer->provider->name);
2088 	}
2089 	if (sc->sc_jconsumer != sc->sc_dconsumer) {
2090 		if (g_io_flush(sc->sc_dconsumer) == 0) {
2091 			sc->sc_bio_flush |= GJ_FLUSH_DATA;
2092 			GJ_DEBUG(1, "BIO_FLUSH supported by %s.",
2093 			    sc->sc_dconsumer->provider->name);
2094 		} else {
2095 			GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2096 			    sc->sc_dconsumer->provider->name);
2097 		}
2098 	}
2099 
2100 	gp = sc->sc_geom;
2101 	g_topology_lock();
2102 	pp = g_new_providerf(gp, "%s.journal", sc->sc_name);
2103 	pp->mediasize = sc->sc_mediasize;
2104 	/*
2105 	 * There could be a problem when data provider and journal providers
2106 	 * have different sectorsize, but such scenario is prevented on journal
2107 	 * creation.
2108 	 */
2109 	pp->sectorsize = sc->sc_sectorsize;
2110 	g_error_provider(pp, 0);
2111 	g_topology_unlock();
2112 	last_write = time_second;
2113 
2114 	if (sc->sc_rootmount != NULL) {
2115 		GJ_DEBUG(1, "root_mount_rel %p", sc->sc_rootmount);
2116 		root_mount_rel(sc->sc_rootmount);
2117 		sc->sc_rootmount = NULL;
2118 	}
2119 
2120 	for (;;) {
2121 		/* Get first request from the queue. */
2122 		mtx_lock(&sc->sc_mtx);
2123 		bp = bioq_first(&sc->sc_back_queue);
2124 		if (bp != NULL)
2125 			type = (bp->bio_cflags & GJ_BIO_MASK);
2126 		if (bp == NULL) {
2127 			bp = bioq_first(&sc->sc_regular_queue);
2128 			if (bp != NULL)
2129 				type = GJ_BIO_REGULAR;
2130 		}
2131 		if (bp == NULL) {
2132 try_switch:
2133 			if ((sc->sc_flags & GJF_DEVICE_SWITCH) ||
2134 			    (sc->sc_flags & GJF_DEVICE_DESTROY)) {
2135 				if (sc->sc_current_count > 0) {
2136 					mtx_unlock(&sc->sc_mtx);
2137 					g_journal_flush(sc);
2138 					g_journal_flush_send(sc);
2139 					continue;
2140 				}
2141 				if (sc->sc_flush_in_progress > 0)
2142 					goto sleep;
2143 				if (sc->sc_copy_in_progress > 0)
2144 					goto sleep;
2145 			}
2146 			if (sc->sc_flags & GJF_DEVICE_SWITCH) {
2147 				mtx_unlock(&sc->sc_mtx);
2148 				g_journal_switch(sc);
2149 				wakeup(&sc->sc_journal_copying);
2150 				continue;
2151 			}
2152 			if (sc->sc_flags & GJF_DEVICE_DESTROY) {
2153 				GJ_DEBUG(1, "Shutting down worker "
2154 				    "thread for %s.", gp->name);
2155 				sc->sc_worker = NULL;
2156 				wakeup(&sc->sc_worker);
2157 				mtx_unlock(&sc->sc_mtx);
2158 				kproc_exit(0);
2159 			}
2160 sleep:
2161 			g_journal_wait(sc, last_write);
2162 			continue;
2163 		}
2164 		/*
2165 		 * If we're in switch process, we need to delay all new
2166 		 * write requests until its done.
2167 		 */
2168 		if ((sc->sc_flags & GJF_DEVICE_SWITCH) &&
2169 		    type == GJ_BIO_REGULAR && bp->bio_cmd == BIO_WRITE) {
2170 			GJ_LOGREQ(2, bp, "WRITE on SWITCH");
2171 			goto try_switch;
2172 		}
2173 		if (type == GJ_BIO_REGULAR)
2174 			bioq_remove(&sc->sc_regular_queue, bp);
2175 		else
2176 			bioq_remove(&sc->sc_back_queue, bp);
2177 		mtx_unlock(&sc->sc_mtx);
2178 		switch (type) {
2179 		case GJ_BIO_REGULAR:
2180 			/* Regular request. */
2181 			switch (bp->bio_cmd) {
2182 			case BIO_READ:
2183 				g_journal_read(sc, bp, bp->bio_offset,
2184 				    bp->bio_offset + bp->bio_length);
2185 				break;
2186 			case BIO_WRITE:
2187 				last_write = time_second;
2188 				g_journal_add_request(sc, bp);
2189 				g_journal_flush_send(sc);
2190 				break;
2191 			default:
2192 				panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2193 			}
2194 			break;
2195 		case GJ_BIO_COPY:
2196 			switch (bp->bio_cmd) {
2197 			case BIO_READ:
2198 				if (g_journal_copy_read_done(bp))
2199 					g_journal_copy_send(sc);
2200 				break;
2201 			case BIO_WRITE:
2202 				g_journal_copy_write_done(bp);
2203 				g_journal_copy_send(sc);
2204 				break;
2205 			default:
2206 				panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2207 			}
2208 			break;
2209 		case GJ_BIO_JOURNAL:
2210 			g_journal_flush_done(bp);
2211 			g_journal_flush_send(sc);
2212 			break;
2213 		case GJ_BIO_READ:
2214 		default:
2215 			panic("Invalid bio (%d).", type);
2216 		}
2217 	}
2218 }
2219 
2220 static void
2221 g_journal_destroy_event(void *arg, int flags __unused)
2222 {
2223 	struct g_journal_softc *sc;
2224 
2225 	g_topology_assert();
2226 	sc = arg;
2227 	g_journal_destroy(sc);
2228 }
2229 
2230 static void
2231 g_journal_timeout(void *arg)
2232 {
2233 	struct g_journal_softc *sc;
2234 
2235 	sc = arg;
2236 	GJ_DEBUG(0, "Timeout. Journal %s cannot be completed.",
2237 	    sc->sc_geom->name);
2238 	g_post_event(g_journal_destroy_event, sc, M_NOWAIT, NULL);
2239 }
2240 
2241 static struct g_geom *
2242 g_journal_create(struct g_class *mp, struct g_provider *pp,
2243     const struct g_journal_metadata *md)
2244 {
2245 	struct g_journal_softc *sc;
2246 	struct g_geom *gp;
2247 	struct g_consumer *cp;
2248 	int error;
2249 
2250 	sc = NULL;	/* gcc */
2251 
2252 	g_topology_assert();
2253 	/*
2254 	 * There are two possibilities:
2255 	 * 1. Data and both journals are on the same provider.
2256 	 * 2. Data and journals are all on separated providers.
2257 	 */
2258 	/* Look for journal device with the same ID. */
2259 	LIST_FOREACH(gp, &mp->geom, geom) {
2260 		sc = gp->softc;
2261 		if (sc == NULL)
2262 			continue;
2263 		if (sc->sc_id == md->md_id)
2264 			break;
2265 	}
2266 	if (gp == NULL)
2267 		sc = NULL;
2268 	else if (sc != NULL && (sc->sc_type & md->md_type) != 0) {
2269 		GJ_DEBUG(1, "Journal device %u already configured.", sc->sc_id);
2270 		return (NULL);
2271 	}
2272 	if (md->md_type == 0 || (md->md_type & ~GJ_TYPE_COMPLETE) != 0) {
2273 		GJ_DEBUG(0, "Invalid type on %s.", pp->name);
2274 		return (NULL);
2275 	}
2276 	if (md->md_type & GJ_TYPE_DATA) {
2277 		GJ_DEBUG(0, "Journal %u: %s contains data.", md->md_id,
2278 		    pp->name);
2279 	}
2280 	if (md->md_type & GJ_TYPE_JOURNAL) {
2281 		GJ_DEBUG(0, "Journal %u: %s contains journal.", md->md_id,
2282 		    pp->name);
2283 	}
2284 
2285 	if (sc == NULL) {
2286 		/* Action geom. */
2287 		sc = malloc(sizeof(*sc), M_JOURNAL, M_WAITOK | M_ZERO);
2288 		sc->sc_id = md->md_id;
2289 		sc->sc_type = 0;
2290 		sc->sc_flags = 0;
2291 		sc->sc_worker = NULL;
2292 
2293 		gp = g_new_geomf(mp, "gjournal %u", sc->sc_id);
2294 		gp->start = g_journal_start;
2295 		gp->orphan = g_journal_orphan;
2296 		gp->access = g_journal_access;
2297 		gp->softc = sc;
2298 		gp->flags |= G_GEOM_VOLATILE_BIO;
2299 		sc->sc_geom = gp;
2300 
2301 		mtx_init(&sc->sc_mtx, "gjournal", NULL, MTX_DEF);
2302 
2303 		bioq_init(&sc->sc_back_queue);
2304 		bioq_init(&sc->sc_regular_queue);
2305 		bioq_init(&sc->sc_delayed_queue);
2306 		sc->sc_delayed_count = 0;
2307 		sc->sc_current_queue = NULL;
2308 		sc->sc_current_count = 0;
2309 		sc->sc_flush_queue = NULL;
2310 		sc->sc_flush_count = 0;
2311 		sc->sc_flush_in_progress = 0;
2312 		sc->sc_copy_queue = NULL;
2313 		sc->sc_copy_in_progress = 0;
2314 		sc->sc_inactive.jj_queue = NULL;
2315 		sc->sc_active.jj_queue = NULL;
2316 
2317 		sc->sc_rootmount = root_mount_hold("GJOURNAL");
2318 		GJ_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2319 
2320 		callout_init(&sc->sc_callout, 1);
2321 		if (md->md_type != GJ_TYPE_COMPLETE) {
2322 			/*
2323 			 * Journal and data are on separate providers.
2324 			 * At this point we have only one of them.
2325 			 * We setup a timeout in case the other part will not
2326 			 * appear, so we won't wait forever.
2327 			 */
2328 			callout_reset(&sc->sc_callout, 5 * hz,
2329 			    g_journal_timeout, sc);
2330 		}
2331 	}
2332 
2333 	/* Remember type of the data provider. */
2334 	if (md->md_type & GJ_TYPE_DATA)
2335 		sc->sc_orig_type = md->md_type;
2336 	sc->sc_type |= md->md_type;
2337 	cp = NULL;
2338 
2339 	if (md->md_type & GJ_TYPE_DATA) {
2340 		if (md->md_flags & GJ_FLAG_CLEAN)
2341 			sc->sc_flags |= GJF_DEVICE_CLEAN;
2342 		if (md->md_flags & GJ_FLAG_CHECKSUM)
2343 			sc->sc_flags |= GJF_DEVICE_CHECKSUM;
2344 		cp = g_new_consumer(gp);
2345 		error = g_attach(cp, pp);
2346 		KASSERT(error == 0, ("Cannot attach to %s (error=%d).",
2347 		    pp->name, error));
2348 		error = g_access(cp, 1, 1, 1);
2349 		if (error != 0) {
2350 			GJ_DEBUG(0, "Cannot access %s (error=%d).", pp->name,
2351 			    error);
2352 			g_journal_destroy(sc);
2353 			return (NULL);
2354 		}
2355 		sc->sc_dconsumer = cp;
2356 		sc->sc_mediasize = pp->mediasize - pp->sectorsize;
2357 		sc->sc_sectorsize = pp->sectorsize;
2358 		sc->sc_jstart = md->md_jstart;
2359 		sc->sc_jend = md->md_jend;
2360 		if (md->md_provider[0] != '\0')
2361 			sc->sc_flags |= GJF_DEVICE_HARDCODED;
2362 		sc->sc_journal_offset = md->md_joffset;
2363 		sc->sc_journal_id = md->md_jid;
2364 		sc->sc_journal_previous_id = md->md_jid;
2365 	}
2366 	if (md->md_type & GJ_TYPE_JOURNAL) {
2367 		if (cp == NULL) {
2368 			cp = g_new_consumer(gp);
2369 			error = g_attach(cp, pp);
2370 			KASSERT(error == 0, ("Cannot attach to %s (error=%d).",
2371 			    pp->name, error));
2372 			error = g_access(cp, 1, 1, 1);
2373 			if (error != 0) {
2374 				GJ_DEBUG(0, "Cannot access %s (error=%d).",
2375 				    pp->name, error);
2376 				g_journal_destroy(sc);
2377 				return (NULL);
2378 			}
2379 		} else {
2380 			/*
2381 			 * Journal is on the same provider as data, which means
2382 			 * that data provider ends where journal starts.
2383 			 */
2384 			sc->sc_mediasize = md->md_jstart;
2385 		}
2386 		sc->sc_jconsumer = cp;
2387 	}
2388 
2389 	/* Start switcher kproc if needed. */
2390 	if (g_journal_switcher_proc == NULL)
2391 		g_journal_start_switcher(mp);
2392 
2393 	if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE) {
2394 		/* Journal is not complete yet. */
2395 		return (gp);
2396 	} else {
2397 		/* Journal complete, cancel timeout. */
2398 		callout_drain(&sc->sc_callout);
2399 	}
2400 
2401 	error = kproc_create(g_journal_worker, sc, &sc->sc_worker, 0, 0,
2402 	    "g_journal %s", sc->sc_name);
2403 	if (error != 0) {
2404 		GJ_DEBUG(0, "Cannot create worker thread for %s.journal.",
2405 		    sc->sc_name);
2406 		g_journal_destroy(sc);
2407 		return (NULL);
2408 	}
2409 
2410 	return (gp);
2411 }
2412 
2413 static void
2414 g_journal_destroy_consumer(void *arg, int flags __unused)
2415 {
2416 	struct g_consumer *cp;
2417 
2418 	g_topology_assert();
2419 	cp = arg;
2420 	g_detach(cp);
2421 	g_destroy_consumer(cp);
2422 }
2423 
2424 static int
2425 g_journal_destroy(struct g_journal_softc *sc)
2426 {
2427 	struct g_geom *gp;
2428 	struct g_provider *pp;
2429 	struct g_consumer *cp;
2430 
2431 	g_topology_assert();
2432 
2433 	if (sc == NULL)
2434 		return (ENXIO);
2435 
2436 	gp = sc->sc_geom;
2437 	pp = LIST_FIRST(&gp->provider);
2438 	if (pp != NULL) {
2439 		if (pp->acr != 0 || pp->acw != 0 || pp->ace != 0) {
2440 			GJ_DEBUG(1, "Device %s is still open (r%dw%de%d).",
2441 			    pp->name, pp->acr, pp->acw, pp->ace);
2442 			return (EBUSY);
2443 		}
2444 		g_error_provider(pp, ENXIO);
2445 
2446 		g_journal_flush(sc);
2447 		g_journal_flush_send(sc);
2448 		g_journal_switch(sc);
2449 	}
2450 
2451 	sc->sc_flags |= (GJF_DEVICE_DESTROY | GJF_DEVICE_CLEAN);
2452 
2453 	g_topology_unlock();
2454 
2455 	if (sc->sc_rootmount != NULL) {
2456 		GJ_DEBUG(1, "root_mount_rel %p", sc->sc_rootmount);
2457 		root_mount_rel(sc->sc_rootmount);
2458 		sc->sc_rootmount = NULL;
2459 	}
2460 
2461 	callout_drain(&sc->sc_callout);
2462 	mtx_lock(&sc->sc_mtx);
2463 	wakeup(sc);
2464 	while (sc->sc_worker != NULL)
2465 		msleep(&sc->sc_worker, &sc->sc_mtx, PRIBIO, "gj:destroy", 0);
2466 	mtx_unlock(&sc->sc_mtx);
2467 
2468 	if (pp != NULL) {
2469 		GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
2470 		g_journal_metadata_update(sc);
2471 		g_topology_lock();
2472 		g_wither_provider(pp, ENXIO);
2473 	} else {
2474 		g_topology_lock();
2475 	}
2476 	mtx_destroy(&sc->sc_mtx);
2477 
2478 	if (sc->sc_current_count != 0) {
2479 		GJ_DEBUG(0, "Warning! Number of current requests %d.",
2480 		    sc->sc_current_count);
2481 	}
2482 
2483 	LIST_FOREACH(cp, &gp->consumer, consumer) {
2484 		if (cp->acr + cp->acw + cp->ace > 0)
2485 			g_access(cp, -1, -1, -1);
2486 		/*
2487 		 * We keep all consumers open for writting, so if I'll detach
2488 		 * and destroy consumer here, I'll get providers for taste, so
2489 		 * journal will be started again.
2490 		 * Sending an event here, prevents this from happening.
2491 		 */
2492 		g_post_event(g_journal_destroy_consumer, cp, M_WAITOK, NULL);
2493 	}
2494 	gp->softc = NULL;
2495 	g_wither_geom(gp, ENXIO);
2496 	free(sc, M_JOURNAL);
2497 	return (0);
2498 }
2499 
2500 static void
2501 g_journal_taste_orphan(struct g_consumer *cp)
2502 {
2503 
2504 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2505 	    cp->provider->name));
2506 }
2507 
2508 static struct g_geom *
2509 g_journal_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2510 {
2511 	struct g_journal_metadata md;
2512 	struct g_consumer *cp;
2513 	struct g_geom *gp;
2514 	int error;
2515 
2516 	g_topology_assert();
2517 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2518 	GJ_DEBUG(2, "Tasting %s.", pp->name);
2519 	if (pp->geom->class == mp)
2520 		return (NULL);
2521 
2522 	gp = g_new_geomf(mp, "journal:taste");
2523 	/* This orphan function should be never called. */
2524 	gp->orphan = g_journal_taste_orphan;
2525 	cp = g_new_consumer(gp);
2526 	g_attach(cp, pp);
2527 	error = g_journal_metadata_read(cp, &md);
2528 	g_detach(cp);
2529 	g_destroy_consumer(cp);
2530 	g_destroy_geom(gp);
2531 	if (error != 0)
2532 		return (NULL);
2533 	gp = NULL;
2534 
2535 	if (md.md_provider[0] != '\0' &&
2536 	    !g_compare_names(md.md_provider, pp->name))
2537 		return (NULL);
2538 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
2539 		return (NULL);
2540 	if (g_journal_debug >= 2)
2541 		journal_metadata_dump(&md);
2542 
2543 	gp = g_journal_create(mp, pp, &md);
2544 	return (gp);
2545 }
2546 
2547 static struct g_journal_softc *
2548 g_journal_find_device(struct g_class *mp, const char *name)
2549 {
2550 	struct g_journal_softc *sc;
2551 	struct g_geom *gp;
2552 	struct g_provider *pp;
2553 
2554 	if (strncmp(name, "/dev/", 5) == 0)
2555 		name += 5;
2556 	LIST_FOREACH(gp, &mp->geom, geom) {
2557 		sc = gp->softc;
2558 		if (sc == NULL)
2559 			continue;
2560 		if (sc->sc_flags & GJF_DEVICE_DESTROY)
2561 			continue;
2562 		if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2563 			continue;
2564 		pp = LIST_FIRST(&gp->provider);
2565 		if (strcmp(sc->sc_name, name) == 0)
2566 			return (sc);
2567 		if (pp != NULL && strcmp(pp->name, name) == 0)
2568 			return (sc);
2569 	}
2570 	return (NULL);
2571 }
2572 
2573 static void
2574 g_journal_ctl_destroy(struct gctl_req *req, struct g_class *mp)
2575 {
2576 	struct g_journal_softc *sc;
2577 	const char *name;
2578 	char param[16];
2579 	int *nargs;
2580 	int error, i;
2581 
2582 	g_topology_assert();
2583 
2584 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
2585 	if (nargs == NULL) {
2586 		gctl_error(req, "No '%s' argument.", "nargs");
2587 		return;
2588 	}
2589 	if (*nargs <= 0) {
2590 		gctl_error(req, "Missing device(s).");
2591 		return;
2592 	}
2593 
2594 	for (i = 0; i < *nargs; i++) {
2595 		snprintf(param, sizeof(param), "arg%d", i);
2596 		name = gctl_get_asciiparam(req, param);
2597 		if (name == NULL) {
2598 			gctl_error(req, "No 'arg%d' argument.", i);
2599 			return;
2600 		}
2601 		sc = g_journal_find_device(mp, name);
2602 		if (sc == NULL) {
2603 			gctl_error(req, "No such device: %s.", name);
2604 			return;
2605 		}
2606 		error = g_journal_destroy(sc);
2607 		if (error != 0) {
2608 			gctl_error(req, "Cannot destroy device %s (error=%d).",
2609 			    LIST_FIRST(&sc->sc_geom->provider)->name, error);
2610 			return;
2611 		}
2612 	}
2613 }
2614 
2615 static void
2616 g_journal_ctl_sync(struct gctl_req *req __unused, struct g_class *mp __unused)
2617 {
2618 
2619 	g_topology_assert();
2620 	g_topology_unlock();
2621 	g_journal_sync_requested++;
2622 	wakeup(&g_journal_switcher_state);
2623 	while (g_journal_sync_requested > 0)
2624 		tsleep(&g_journal_sync_requested, PRIBIO, "j:sreq", hz / 2);
2625 	g_topology_lock();
2626 }
2627 
2628 static void
2629 g_journal_config(struct gctl_req *req, struct g_class *mp, const char *verb)
2630 {
2631 	uint32_t *version;
2632 
2633 	g_topology_assert();
2634 
2635 	version = gctl_get_paraml(req, "version", sizeof(*version));
2636 	if (version == NULL) {
2637 		gctl_error(req, "No '%s' argument.", "version");
2638 		return;
2639 	}
2640 	if (*version != G_JOURNAL_VERSION) {
2641 		gctl_error(req, "Userland and kernel parts are out of sync.");
2642 		return;
2643 	}
2644 
2645 	if (strcmp(verb, "destroy") == 0 || strcmp(verb, "stop") == 0) {
2646 		g_journal_ctl_destroy(req, mp);
2647 		return;
2648 	} else if (strcmp(verb, "sync") == 0) {
2649 		g_journal_ctl_sync(req, mp);
2650 		return;
2651 	}
2652 
2653 	gctl_error(req, "Unknown verb.");
2654 }
2655 
2656 static void
2657 g_journal_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
2658     struct g_consumer *cp, struct g_provider *pp)
2659 {
2660 	struct g_journal_softc *sc;
2661 
2662 	g_topology_assert();
2663 
2664 	sc = gp->softc;
2665 	if (sc == NULL)
2666 		return;
2667 	if (pp != NULL) {
2668 		/* Nothing here. */
2669 	} else if (cp != NULL) {
2670 		int first = 1;
2671 
2672 		sbuf_printf(sb, "%s<Role>", indent);
2673 		if (cp == sc->sc_dconsumer) {
2674 			sbuf_printf(sb, "Data");
2675 			first = 0;
2676 		}
2677 		if (cp == sc->sc_jconsumer) {
2678 			if (!first)
2679 				sbuf_printf(sb, ",");
2680 			sbuf_printf(sb, "Journal");
2681 		}
2682 		sbuf_printf(sb, "</Role>\n");
2683 		if (cp == sc->sc_jconsumer) {
2684 			sbuf_printf(sb, "<Jstart>%jd</Jstart>\n",
2685 			    (intmax_t)sc->sc_jstart);
2686 			sbuf_printf(sb, "<Jend>%jd</Jend>\n",
2687 			    (intmax_t)sc->sc_jend);
2688 		}
2689 	} else {
2690 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
2691 	}
2692 }
2693 
2694 static eventhandler_tag g_journal_event_shutdown = NULL;
2695 static eventhandler_tag g_journal_event_lowmem = NULL;
2696 
2697 static void
2698 g_journal_shutdown(void *arg, int howto __unused)
2699 {
2700 	struct g_class *mp;
2701 	struct g_geom *gp, *gp2;
2702 
2703 	if (panicstr != NULL)
2704 		return;
2705 	mp = arg;
2706 	g_topology_lock();
2707 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
2708 		if (gp->softc == NULL)
2709 			continue;
2710 		GJ_DEBUG(0, "Shutting down geom %s.", gp->name);
2711 		g_journal_destroy(gp->softc);
2712 	}
2713 	g_topology_unlock();
2714 }
2715 
2716 /*
2717  * Free cached requests from inactive queue in case of low memory.
2718  * We free GJ_FREE_AT_ONCE elements at once.
2719  */
2720 #define	GJ_FREE_AT_ONCE	4
2721 static void
2722 g_journal_lowmem(void *arg, int howto __unused)
2723 {
2724 	struct g_journal_softc *sc;
2725 	struct g_class *mp;
2726 	struct g_geom *gp;
2727 	struct bio *bp;
2728 	u_int nfree = GJ_FREE_AT_ONCE;
2729 
2730 	g_journal_stats_low_mem++;
2731 	mp = arg;
2732 	g_topology_lock();
2733 	LIST_FOREACH(gp, &mp->geom, geom) {
2734 		sc = gp->softc;
2735 		if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY))
2736 			continue;
2737 		mtx_lock(&sc->sc_mtx);
2738 		for (bp = sc->sc_inactive.jj_queue; nfree > 0 && bp != NULL;
2739 		    nfree--, bp = bp->bio_next) {
2740 			/*
2741 			 * This is safe to free the bio_data, because:
2742 			 * 1. If bio_data is NULL it will be read from the
2743 			 *    inactive journal.
2744 			 * 2. If bp is sent down, it is first removed from the
2745 			 *    inactive queue, so it's impossible to free the
2746 			 *    data from under in-flight bio.
2747 			 * On the other hand, freeing elements from the active
2748 			 * queue, is not safe.
2749 			 */
2750 			if (bp->bio_data != NULL) {
2751 				GJ_DEBUG(2, "Freeing data from %s.",
2752 				    sc->sc_name);
2753 				gj_free(bp->bio_data, bp->bio_length);
2754 				bp->bio_data = NULL;
2755 			}
2756 		}
2757 		mtx_unlock(&sc->sc_mtx);
2758 		if (nfree == 0)
2759 			break;
2760 	}
2761 	g_topology_unlock();
2762 }
2763 
2764 static void g_journal_switcher(void *arg);
2765 
2766 static void
2767 g_journal_init(struct g_class *mp)
2768 {
2769 
2770 	/* Pick a conservative value if provided value sucks. */
2771 	if (g_journal_cache_divisor <= 0 ||
2772 	    (vm_kmem_size / g_journal_cache_divisor == 0)) {
2773 		g_journal_cache_divisor = 5;
2774 	}
2775 	if (g_journal_cache_limit > 0) {
2776 		g_journal_cache_limit = vm_kmem_size / g_journal_cache_divisor;
2777 		g_journal_cache_low =
2778 		    (g_journal_cache_limit / 100) * g_journal_cache_switch;
2779 	}
2780 	g_journal_event_shutdown = EVENTHANDLER_REGISTER(shutdown_post_sync,
2781 	    g_journal_shutdown, mp, EVENTHANDLER_PRI_FIRST);
2782 	if (g_journal_event_shutdown == NULL)
2783 		GJ_DEBUG(0, "Warning! Cannot register shutdown event.");
2784 	g_journal_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem,
2785 	    g_journal_lowmem, mp, EVENTHANDLER_PRI_FIRST);
2786 	if (g_journal_event_lowmem == NULL)
2787 		GJ_DEBUG(0, "Warning! Cannot register lowmem event.");
2788 }
2789 
2790 static void
2791 g_journal_fini(struct g_class *mp)
2792 {
2793 
2794 	if (g_journal_event_shutdown != NULL) {
2795 		EVENTHANDLER_DEREGISTER(shutdown_post_sync,
2796 		    g_journal_event_shutdown);
2797 	}
2798 	if (g_journal_event_lowmem != NULL)
2799 		EVENTHANDLER_DEREGISTER(vm_lowmem, g_journal_event_lowmem);
2800 	g_journal_stop_switcher();
2801 }
2802 
2803 DECLARE_GEOM_CLASS(g_journal_class, g_journal);
2804 
2805 static const struct g_journal_desc *
2806 g_journal_find_desc(const char *fstype)
2807 {
2808 	const struct g_journal_desc *desc;
2809 	int i;
2810 
2811 	for (desc = g_journal_filesystems[i = 0]; desc != NULL;
2812 	     desc = g_journal_filesystems[++i]) {
2813 		if (strcmp(desc->jd_fstype, fstype) == 0)
2814 			break;
2815 	}
2816 	return (desc);
2817 }
2818 
2819 static void
2820 g_journal_switch_wait(struct g_journal_softc *sc)
2821 {
2822 	struct bintime bt;
2823 
2824 	mtx_assert(&sc->sc_mtx, MA_OWNED);
2825 	if (g_journal_debug >= 2) {
2826 		if (sc->sc_flush_in_progress > 0) {
2827 			GJ_DEBUG(2, "%d requests flushing.",
2828 			    sc->sc_flush_in_progress);
2829 		}
2830 		if (sc->sc_copy_in_progress > 0) {
2831 			GJ_DEBUG(2, "%d requests copying.",
2832 			    sc->sc_copy_in_progress);
2833 		}
2834 		if (sc->sc_flush_count > 0) {
2835 			GJ_DEBUG(2, "%d requests to flush.",
2836 			    sc->sc_flush_count);
2837 		}
2838 		if (sc->sc_delayed_count > 0) {
2839 			GJ_DEBUG(2, "%d requests delayed.",
2840 			    sc->sc_delayed_count);
2841 		}
2842 	}
2843 	g_journal_stats_switches++;
2844 	if (sc->sc_copy_in_progress > 0)
2845 		g_journal_stats_wait_for_copy++;
2846 	GJ_TIMER_START(1, &bt);
2847 	sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2848 	sc->sc_flags |= GJF_DEVICE_SWITCH;
2849 	wakeup(sc);
2850 	while (sc->sc_flags & GJF_DEVICE_SWITCH) {
2851 		msleep(&sc->sc_journal_copying, &sc->sc_mtx, PRIBIO,
2852 		    "gj:switch", 0);
2853 	}
2854 	GJ_TIMER_STOP(1, &bt, "Switch time of %s", sc->sc_name);
2855 }
2856 
2857 static void
2858 g_journal_do_switch(struct g_class *classp)
2859 {
2860 	struct g_journal_softc *sc;
2861 	const struct g_journal_desc *desc;
2862 	struct g_geom *gp;
2863 	struct mount *mp;
2864 	struct bintime bt;
2865 	char *mountpoint;
2866 	int error, save;
2867 
2868 	g_topology_lock();
2869 	LIST_FOREACH(gp, &classp->geom, geom) {
2870 		sc = gp->softc;
2871 		if (sc == NULL)
2872 			continue;
2873 		if (sc->sc_flags & GJF_DEVICE_DESTROY)
2874 			continue;
2875 		if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2876 			continue;
2877 		mtx_lock(&sc->sc_mtx);
2878 		sc->sc_flags |= GJF_DEVICE_BEFORE_SWITCH;
2879 		mtx_unlock(&sc->sc_mtx);
2880 	}
2881 	g_topology_unlock();
2882 
2883 	mtx_lock(&mountlist_mtx);
2884 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2885 		if (mp->mnt_gjprovider == NULL)
2886 			continue;
2887 		if (mp->mnt_flag & MNT_RDONLY)
2888 			continue;
2889 		desc = g_journal_find_desc(mp->mnt_stat.f_fstypename);
2890 		if (desc == NULL)
2891 			continue;
2892 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
2893 			continue;
2894 		/* mtx_unlock(&mountlist_mtx) was done inside vfs_busy() */
2895 
2896 		g_topology_lock();
2897 		sc = g_journal_find_device(classp, mp->mnt_gjprovider);
2898 		g_topology_unlock();
2899 
2900 		if (sc == NULL) {
2901 			GJ_DEBUG(0, "Cannot find journal geom for %s.",
2902 			    mp->mnt_gjprovider);
2903 			goto next;
2904 		} else if (JEMPTY(sc)) {
2905 			mtx_lock(&sc->sc_mtx);
2906 			sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2907 			mtx_unlock(&sc->sc_mtx);
2908 			GJ_DEBUG(3, "No need for %s switch.", sc->sc_name);
2909 			goto next;
2910 		}
2911 
2912 		mountpoint = mp->mnt_stat.f_mntonname;
2913 
2914 		error = vn_start_write(NULL, &mp, V_WAIT);
2915 		if (error != 0) {
2916 			GJ_DEBUG(0, "vn_start_write(%s) failed (error=%d).",
2917 			    mountpoint, error);
2918 			goto next;
2919 		}
2920 
2921 		save = curthread_pflags_set(TDP_SYNCIO);
2922 
2923 		GJ_TIMER_START(1, &bt);
2924 		vfs_msync(mp, MNT_NOWAIT);
2925 		GJ_TIMER_STOP(1, &bt, "Msync time of %s", mountpoint);
2926 
2927 		GJ_TIMER_START(1, &bt);
2928 		error = VFS_SYNC(mp, MNT_NOWAIT);
2929 		if (error == 0)
2930 			GJ_TIMER_STOP(1, &bt, "Sync time of %s", mountpoint);
2931 		else {
2932 			GJ_DEBUG(0, "Cannot sync file system %s (error=%d).",
2933 			    mountpoint, error);
2934 		}
2935 
2936 		curthread_pflags_restore(save);
2937 
2938 		vn_finished_write(mp);
2939 
2940 		if (error != 0)
2941 			goto next;
2942 
2943 		/*
2944 		 * Send BIO_FLUSH before freezing the file system, so it can be
2945 		 * faster after the freeze.
2946 		 */
2947 		GJ_TIMER_START(1, &bt);
2948 		g_journal_flush_cache(sc);
2949 		GJ_TIMER_STOP(1, &bt, "BIO_FLUSH time of %s", sc->sc_name);
2950 
2951 		GJ_TIMER_START(1, &bt);
2952 		error = vfs_write_suspend(mp, VS_SKIP_UNMOUNT);
2953 		GJ_TIMER_STOP(1, &bt, "Suspend time of %s", mountpoint);
2954 		if (error != 0) {
2955 			GJ_DEBUG(0, "Cannot suspend file system %s (error=%d).",
2956 			    mountpoint, error);
2957 			goto next;
2958 		}
2959 
2960 		error = desc->jd_clean(mp);
2961 		if (error != 0)
2962 			goto next;
2963 
2964 		mtx_lock(&sc->sc_mtx);
2965 		g_journal_switch_wait(sc);
2966 		mtx_unlock(&sc->sc_mtx);
2967 
2968 		vfs_write_resume(mp, 0);
2969 next:
2970 		mtx_lock(&mountlist_mtx);
2971 		vfs_unbusy(mp);
2972 	}
2973 	mtx_unlock(&mountlist_mtx);
2974 
2975 	sc = NULL;
2976 	for (;;) {
2977 		g_topology_lock();
2978 		LIST_FOREACH(gp, &g_journal_class.geom, geom) {
2979 			sc = gp->softc;
2980 			if (sc == NULL)
2981 				continue;
2982 			mtx_lock(&sc->sc_mtx);
2983 			if ((sc->sc_type & GJ_TYPE_COMPLETE) == GJ_TYPE_COMPLETE &&
2984 			    !(sc->sc_flags & GJF_DEVICE_DESTROY) &&
2985 			    (sc->sc_flags & GJF_DEVICE_BEFORE_SWITCH)) {
2986 				break;
2987 			}
2988 			mtx_unlock(&sc->sc_mtx);
2989 			sc = NULL;
2990 		}
2991 		g_topology_unlock();
2992 		if (sc == NULL)
2993 			break;
2994 		mtx_assert(&sc->sc_mtx, MA_OWNED);
2995 		g_journal_switch_wait(sc);
2996 		mtx_unlock(&sc->sc_mtx);
2997 	}
2998 }
2999 
3000 static void
3001 g_journal_start_switcher(struct g_class *mp)
3002 {
3003 	int error;
3004 
3005 	g_topology_assert();
3006 	MPASS(g_journal_switcher_proc == NULL);
3007 	g_journal_switcher_state = GJ_SWITCHER_WORKING;
3008 	error = kproc_create(g_journal_switcher, mp, &g_journal_switcher_proc,
3009 	    0, 0, "g_journal switcher");
3010 	KASSERT(error == 0, ("Cannot create switcher thread."));
3011 }
3012 
3013 static void
3014 g_journal_stop_switcher(void)
3015 {
3016 	g_topology_assert();
3017 	MPASS(g_journal_switcher_proc != NULL);
3018 	g_journal_switcher_state = GJ_SWITCHER_DIE;
3019 	wakeup(&g_journal_switcher_state);
3020 	while (g_journal_switcher_state != GJ_SWITCHER_DIED)
3021 		tsleep(&g_journal_switcher_state, PRIBIO, "jfini:wait", hz / 5);
3022 	GJ_DEBUG(1, "Switcher died.");
3023 	g_journal_switcher_proc = NULL;
3024 }
3025 
3026 /*
3027  * TODO: Kill switcher thread on last geom destruction?
3028  */
3029 static void
3030 g_journal_switcher(void *arg)
3031 {
3032 	struct g_class *mp;
3033 	struct bintime bt;
3034 	int error;
3035 
3036 	mp = arg;
3037 	curthread->td_pflags |= TDP_NORUNNINGBUF;
3038 	for (;;) {
3039 		g_journal_switcher_wokenup = 0;
3040 		error = tsleep(&g_journal_switcher_state, PRIBIO, "jsw:wait",
3041 		    g_journal_switch_time * hz);
3042 		if (g_journal_switcher_state == GJ_SWITCHER_DIE) {
3043 			g_journal_switcher_state = GJ_SWITCHER_DIED;
3044 			GJ_DEBUG(1, "Switcher exiting.");
3045 			wakeup(&g_journal_switcher_state);
3046 			kproc_exit(0);
3047 		}
3048 		if (error == 0 && g_journal_sync_requested == 0) {
3049 			GJ_DEBUG(1, "Out of cache, force switch (used=%u "
3050 			    "limit=%u).", g_journal_cache_used,
3051 			    g_journal_cache_limit);
3052 		}
3053 		GJ_TIMER_START(1, &bt);
3054 		g_journal_do_switch(mp);
3055 		GJ_TIMER_STOP(1, &bt, "Entire switch time");
3056 		if (g_journal_sync_requested > 0) {
3057 			g_journal_sync_requested = 0;
3058 			wakeup(&g_journal_sync_requested);
3059 		}
3060 	}
3061 }
3062