2 * Copyright (c) 2005-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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.
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
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
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>
36 #include <sys/mutex.h>
38 #include <sys/sysctl.h>
39 #include <sys/malloc.h>
40 #include <sys/mount.h>
41 #include <sys/eventhandler.h>
43 #include <sys/kthread.h>
44 #include <sys/sched.h>
45 #include <sys/taskqueue.h>
46 #include <sys/vnode.h>
49 #include <sys/stack.h>
53 #include <vm/vm_kern.h>
54 #include <geom/geom.h>
56 #include <geom/journal/g_journal.h>
58 FEATURE(geom_journal, "GEOM journaling support");
61 * On-disk journal format:
66 * %%%%%% ****** +------+ +------+ ****** +------+ %%%%%%
67 * % JH % * RH * | Data | | Data | ... * RH * | Data | ... % JH % ...
68 * %%%%%% ****** +------+ +------+ ****** +------+ %%%%%%
72 CTASSERT(sizeof(struct g_journal_header) <= 512);
73 CTASSERT(sizeof(struct g_journal_record_header) <= 512);
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);
79 const struct g_journal_desc *g_journal_filesystems[] = {
84 SYSCTL_DECL(_kern_geom);
86 int g_journal_debug = 0;
87 TUNABLE_INT("kern.geom.journal.debug", &g_journal_debug);
88 static u_int g_journal_switch_time = 10;
89 static u_int g_journal_force_switch = 70;
90 static u_int g_journal_parallel_flushes = 16;
91 static u_int g_journal_parallel_copies = 16;
92 static u_int g_journal_accept_immediately = 64;
93 static u_int g_journal_record_entries = GJ_RECORD_HEADER_NENTRIES;
94 static u_int g_journal_do_optimize = 1;
96 static SYSCTL_NODE(_kern_geom, OID_AUTO, journal, CTLFLAG_RW, 0,
97 "GEOM_JOURNAL stuff");
98 SYSCTL_INT(_kern_geom_journal, OID_AUTO, debug, CTLFLAG_RW, &g_journal_debug, 0,
100 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, switch_time, CTLFLAG_RW,
101 &g_journal_switch_time, 0, "Switch journals every N seconds");
102 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, force_switch, CTLFLAG_RW,
103 &g_journal_force_switch, 0, "Force switch when journal is N% full");
104 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_flushes, CTLFLAG_RW,
105 &g_journal_parallel_flushes, 0,
106 "Number of flush I/O requests to send in parallel");
107 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, accept_immediately, CTLFLAG_RW,
108 &g_journal_accept_immediately, 0,
109 "Number of I/O requests accepted immediately");
110 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_copies, CTLFLAG_RW,
111 &g_journal_parallel_copies, 0,
112 "Number of copy I/O requests to send in parallel");
114 g_journal_record_entries_sysctl(SYSCTL_HANDLER_ARGS)
119 entries = g_journal_record_entries;
120 error = sysctl_handle_int(oidp, &entries, 0, req);
121 if (error != 0 || req->newptr == NULL)
123 if (entries < 1 || entries > GJ_RECORD_HEADER_NENTRIES)
125 g_journal_record_entries = entries;
128 SYSCTL_PROC(_kern_geom_journal, OID_AUTO, record_entries,
129 CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_record_entries_sysctl, "I",
130 "Maximum number of entires in one journal record");
131 SYSCTL_UINT(_kern_geom_journal, OID_AUTO, optimize, CTLFLAG_RW,
132 &g_journal_do_optimize, 0, "Try to combine bios on flush and copy");
134 static u_int g_journal_cache_used = 0;
135 static u_int g_journal_cache_limit = 64 * 1024 * 1024;
136 TUNABLE_INT("kern.geom.journal.cache.limit", &g_journal_cache_limit);
137 static u_int g_journal_cache_divisor = 2;
138 TUNABLE_INT("kern.geom.journal.cache.divisor", &g_journal_cache_divisor);
139 static u_int g_journal_cache_switch = 90;
140 static u_int g_journal_cache_misses = 0;
141 static u_int g_journal_cache_alloc_failures = 0;
142 static u_int g_journal_cache_low = 0;
144 static SYSCTL_NODE(_kern_geom_journal, OID_AUTO, cache, CTLFLAG_RW, 0,
145 "GEOM_JOURNAL cache");
146 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, used, CTLFLAG_RD,
147 &g_journal_cache_used, 0, "Number of allocated bytes");
149 g_journal_cache_limit_sysctl(SYSCTL_HANDLER_ARGS)
154 limit = g_journal_cache_limit;
155 error = sysctl_handle_int(oidp, &limit, 0, req);
156 if (error != 0 || req->newptr == NULL)
158 g_journal_cache_limit = limit;
159 g_journal_cache_low = (limit / 100) * g_journal_cache_switch;
162 SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, limit,
163 CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_cache_limit_sysctl, "I",
164 "Maximum number of allocated bytes");
165 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, divisor, CTLFLAG_RDTUN,
166 &g_journal_cache_divisor, 0,
167 "(kmem_size / kern.geom.journal.cache.divisor) == cache size");
169 g_journal_cache_switch_sysctl(SYSCTL_HANDLER_ARGS)
174 cswitch = g_journal_cache_switch;
175 error = sysctl_handle_int(oidp, &cswitch, 0, req);
176 if (error != 0 || req->newptr == NULL)
178 if (cswitch < 0 || cswitch > 100)
180 g_journal_cache_switch = cswitch;
181 g_journal_cache_low = (g_journal_cache_limit / 100) * cswitch;
184 SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, switch,
185 CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_cache_switch_sysctl, "I",
186 "Force switch when we hit this percent of cache use");
187 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, misses, CTLFLAG_RW,
188 &g_journal_cache_misses, 0, "Number of cache misses");
189 SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, alloc_failures, CTLFLAG_RW,
190 &g_journal_cache_alloc_failures, 0, "Memory allocation failures");
192 static u_long g_journal_stats_bytes_skipped = 0;
193 static u_long g_journal_stats_combined_ios = 0;
194 static u_long g_journal_stats_switches = 0;
195 static u_long g_journal_stats_wait_for_copy = 0;
196 static u_long g_journal_stats_journal_full = 0;
197 static u_long g_journal_stats_low_mem = 0;
199 static SYSCTL_NODE(_kern_geom_journal, OID_AUTO, stats, CTLFLAG_RW, 0,
200 "GEOM_JOURNAL statistics");
201 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, skipped_bytes, CTLFLAG_RW,
202 &g_journal_stats_bytes_skipped, 0, "Number of skipped bytes");
203 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, combined_ios, CTLFLAG_RW,
204 &g_journal_stats_combined_ios, 0, "Number of combined I/O requests");
205 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, switches, CTLFLAG_RW,
206 &g_journal_stats_switches, 0, "Number of journal switches");
207 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, wait_for_copy, CTLFLAG_RW,
208 &g_journal_stats_wait_for_copy, 0, "Wait for journal copy on switch");
209 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, journal_full, CTLFLAG_RW,
210 &g_journal_stats_journal_full, 0,
211 "Number of times journal was almost full.");
212 SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, low_mem, CTLFLAG_RW,
213 &g_journal_stats_low_mem, 0, "Number of times low_mem hook was called.");
215 static g_taste_t g_journal_taste;
216 static g_ctl_req_t g_journal_config;
217 static g_dumpconf_t g_journal_dumpconf;
218 static g_init_t g_journal_init;
219 static g_fini_t g_journal_fini;
221 struct g_class g_journal_class = {
222 .name = G_JOURNAL_CLASS_NAME,
223 .version = G_VERSION,
224 .taste = g_journal_taste,
225 .ctlreq = g_journal_config,
226 .dumpconf = g_journal_dumpconf,
227 .init = g_journal_init,
228 .fini = g_journal_fini
231 static int g_journal_destroy(struct g_journal_softc *sc);
232 static void g_journal_metadata_update(struct g_journal_softc *sc);
233 static void g_journal_switch_wait(struct g_journal_softc *sc);
235 #define GJ_SWITCHER_WORKING 0
236 #define GJ_SWITCHER_DIE 1
237 #define GJ_SWITCHER_DIED 2
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;
245 struct stack mi_stack;
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.
254 gj_malloc(size_t size, int flags)
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);
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++;
274 g_journal_cache_used += size;
275 mtx_unlock(&g_journal_cache_mtx);
278 p = malloc(size, M_JOURNAL, flags | M_WAITOK);
280 mi = malloc(sizeof(*mi) + size, M_JOURNAL, flags | M_WAITOK);
281 p = (u_char *)mi + sizeof(*mi);
283 stack_save(&mi->mi_stack);
289 gj_free(void *p, size_t size)
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);
302 mi = p = (void *)((u_char *)p - sizeof(*mi));
303 if (mi->mi_size != size) {
304 printf("GJOURNAL: Size mismatch! %zu != %zu\n", size,
306 printf("GJOURNAL: Alloc backtrace:\n");
307 stack_print(&mi->mi_stack);
308 printf("GJOURNAL: Free backtrace:\n");
316 gj_realloc(void *p, size_t size, size_t oldsize)
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);
327 np = gj_malloc(size, M_WAITOK);
328 bcopy(p, np, MIN(oldsize, size));
335 g_journal_check_overflow(struct g_journal_softc *sc)
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 (joffset=%jd active=%jd inactive=%jd)",
345 (intmax_t)sc->sc_journal_offset,
346 (intmax_t)sc->sc_active.jj_offset,
347 (intmax_t)sc->sc_inactive.jj_offset);
349 if (sc->sc_active.jj_offset < sc->sc_inactive.jj_offset) {
350 length = sc->sc_inactive.jj_offset - sc->sc_active.jj_offset;
351 used = sc->sc_journal_offset - sc->sc_active.jj_offset;
353 length = sc->sc_jend - sc->sc_active.jj_offset;
354 length += sc->sc_inactive.jj_offset - sc->sc_jstart;
355 if (sc->sc_journal_offset >= sc->sc_active.jj_offset)
356 used = sc->sc_journal_offset - sc->sc_active.jj_offset;
358 used = sc->sc_jend - sc->sc_active.jj_offset;
359 used += sc->sc_journal_offset - sc->sc_jstart;
362 /* Already woken up? */
363 if (g_journal_switcher_wokenup)
366 * If the active journal takes more than g_journal_force_switch precent
367 * of free journal space, we force journal switch.
370 ("length=%jd used=%jd active=%jd inactive=%jd joffset=%jd",
371 (intmax_t)length, (intmax_t)used,
372 (intmax_t)sc->sc_active.jj_offset,
373 (intmax_t)sc->sc_inactive.jj_offset,
374 (intmax_t)sc->sc_journal_offset));
375 if ((used * 100) / length > g_journal_force_switch) {
376 g_journal_stats_journal_full++;
377 GJ_DEBUG(1, "Journal %s %jd%% full, forcing journal switch.",
378 sc->sc_name, (used * 100) / length);
379 mtx_lock(&g_journal_cache_mtx);
380 g_journal_switcher_wokenup = 1;
381 wakeup(&g_journal_switcher_state);
382 mtx_unlock(&g_journal_cache_mtx);
387 g_journal_orphan(struct g_consumer *cp)
389 struct g_journal_softc *sc;
394 sc = cp->geom->softc;
395 strlcpy(name, cp->provider->name, sizeof(name));
396 GJ_DEBUG(0, "Lost provider %s.", name);
399 error = g_journal_destroy(sc);
401 GJ_DEBUG(0, "Journal %s destroyed.", name);
403 GJ_DEBUG(0, "Cannot destroy journal %s (error=%d). "
404 "Destroy it manually after last close.", sc->sc_name,
410 g_journal_access(struct g_provider *pp, int acr, int acw, int ace)
412 struct g_journal_softc *sc;
416 GJ_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name,
423 sc = pp->geom->softc;
424 if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY)) {
425 if (acr <= 0 && acw <= 0 && ace <= 0)
430 if (pp->acw == 0 && dcw > 0) {
431 GJ_DEBUG(1, "Marking %s as dirty.", sc->sc_name);
432 sc->sc_flags &= ~GJF_DEVICE_CLEAN;
434 g_journal_metadata_update(sc);
436 } /* else if (pp->acw == 0 && dcw > 0 && JEMPTY(sc)) {
437 GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
438 sc->sc_flags |= GJF_DEVICE_CLEAN;
440 g_journal_metadata_update(sc);
447 g_journal_header_encode(struct g_journal_header *hdr, u_char *data)
450 bcopy(GJ_HEADER_MAGIC, data, sizeof(GJ_HEADER_MAGIC));
451 data += sizeof(GJ_HEADER_MAGIC);
452 le32enc(data, hdr->jh_journal_id);
454 le32enc(data, hdr->jh_journal_next_id);
458 g_journal_header_decode(const u_char *data, struct g_journal_header *hdr)
461 bcopy(data, hdr->jh_magic, sizeof(hdr->jh_magic));
462 data += sizeof(hdr->jh_magic);
463 if (bcmp(hdr->jh_magic, GJ_HEADER_MAGIC, sizeof(GJ_HEADER_MAGIC)) != 0)
465 hdr->jh_journal_id = le32dec(data);
467 hdr->jh_journal_next_id = le32dec(data);
472 g_journal_flush_cache(struct g_journal_softc *sc)
477 if (sc->sc_bio_flush == 0)
479 GJ_TIMER_START(1, &bt);
480 if (sc->sc_bio_flush & GJ_FLUSH_JOURNAL) {
481 error = g_io_flush(sc->sc_jconsumer);
482 GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.",
483 sc->sc_jconsumer->provider->name, error);
485 if (sc->sc_bio_flush & GJ_FLUSH_DATA) {
487 * TODO: This could be called in parallel with the
490 error = g_io_flush(sc->sc_dconsumer);
491 GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.",
492 sc->sc_dconsumer->provider->name, error);
494 GJ_TIMER_STOP(1, &bt, "Cache flush time");
498 g_journal_write_header(struct g_journal_softc *sc)
500 struct g_journal_header hdr;
501 struct g_consumer *cp;
505 cp = sc->sc_jconsumer;
506 buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
508 strlcpy(hdr.jh_magic, GJ_HEADER_MAGIC, sizeof(hdr.jh_magic));
509 hdr.jh_journal_id = sc->sc_journal_id;
510 hdr.jh_journal_next_id = sc->sc_journal_next_id;
511 g_journal_header_encode(&hdr, buf);
512 error = g_write_data(cp, sc->sc_journal_offset, buf,
513 cp->provider->sectorsize);
514 /* if (error == 0) */
515 sc->sc_journal_offset += cp->provider->sectorsize;
517 gj_free(buf, cp->provider->sectorsize);
522 * Every journal record has a header and data following it.
523 * Functions below are used to decode the header before storing it to
524 * little endian and to encode it after reading to system endianess.
527 g_journal_record_header_encode(struct g_journal_record_header *hdr,
530 struct g_journal_entry *ent;
533 bcopy(GJ_RECORD_HEADER_MAGIC, data, sizeof(GJ_RECORD_HEADER_MAGIC));
534 data += sizeof(GJ_RECORD_HEADER_MAGIC);
535 le32enc(data, hdr->jrh_journal_id);
537 le16enc(data, hdr->jrh_nentries);
539 bcopy(hdr->jrh_sum, data, sizeof(hdr->jrh_sum));
541 for (i = 0; i < hdr->jrh_nentries; i++) {
542 ent = &hdr->jrh_entries[i];
543 le64enc(data, ent->je_joffset);
545 le64enc(data, ent->je_offset);
547 le64enc(data, ent->je_length);
553 g_journal_record_header_decode(const u_char *data,
554 struct g_journal_record_header *hdr)
556 struct g_journal_entry *ent;
559 bcopy(data, hdr->jrh_magic, sizeof(hdr->jrh_magic));
560 data += sizeof(hdr->jrh_magic);
561 if (strcmp(hdr->jrh_magic, GJ_RECORD_HEADER_MAGIC) != 0)
563 hdr->jrh_journal_id = le32dec(data);
565 hdr->jrh_nentries = le16dec(data);
567 if (hdr->jrh_nentries > GJ_RECORD_HEADER_NENTRIES)
569 bcopy(data, hdr->jrh_sum, sizeof(hdr->jrh_sum));
571 for (i = 0; i < hdr->jrh_nentries; i++) {
572 ent = &hdr->jrh_entries[i];
573 ent->je_joffset = le64dec(data);
575 ent->je_offset = le64dec(data);
577 ent->je_length = le64dec(data);
584 * Function reads metadata from a provider (via the given consumer), decodes
585 * it to system endianess and verifies its correctness.
588 g_journal_metadata_read(struct g_consumer *cp, struct g_journal_metadata *md)
590 struct g_provider *pp;
596 error = g_access(cp, 1, 0, 0);
601 /* Metadata is stored in last sector. */
602 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
605 g_access(cp, -1, 0, 0);
607 GJ_DEBUG(1, "Cannot read metadata from %s (error=%d).",
608 cp->provider->name, error);
612 /* Decode metadata. */
613 error = journal_metadata_decode(buf, md);
615 /* Is this is gjournal provider at all? */
616 if (strcmp(md->md_magic, G_JOURNAL_MAGIC) != 0)
619 * Are we able to handle this version of metadata?
620 * We only maintain backward compatibility.
622 if (md->md_version > G_JOURNAL_VERSION) {
624 "Kernel module is too old to handle metadata from %s.",
628 /* Is checksum correct? */
630 GJ_DEBUG(0, "MD5 metadata hash mismatch for provider %s.",
638 * Two functions below are responsible for updating metadata.
639 * Only metadata on the data provider is updated (we need to update
640 * information about active journal in there).
643 g_journal_metadata_done(struct bio *bp)
647 * There is not much we can do on error except informing about it.
649 if (bp->bio_error != 0) {
650 GJ_LOGREQ(0, bp, "Cannot update metadata (error=%d).",
653 GJ_LOGREQ(2, bp, "Metadata updated.");
655 gj_free(bp->bio_data, bp->bio_length);
660 g_journal_metadata_update(struct g_journal_softc *sc)
662 struct g_journal_metadata md;
663 struct g_consumer *cp;
667 cp = sc->sc_dconsumer;
668 sector = gj_malloc(cp->provider->sectorsize, M_WAITOK);
669 strlcpy(md.md_magic, G_JOURNAL_MAGIC, sizeof(md.md_magic));
670 md.md_version = G_JOURNAL_VERSION;
671 md.md_id = sc->sc_id;
672 md.md_type = sc->sc_orig_type;
673 md.md_jstart = sc->sc_jstart;
674 md.md_jend = sc->sc_jend;
675 md.md_joffset = sc->sc_inactive.jj_offset;
676 md.md_jid = sc->sc_journal_previous_id;
678 if (sc->sc_flags & GJF_DEVICE_CLEAN)
679 md.md_flags |= GJ_FLAG_CLEAN;
681 if (sc->sc_flags & GJF_DEVICE_HARDCODED)
682 strlcpy(md.md_provider, sc->sc_name, sizeof(md.md_provider));
684 bzero(md.md_provider, sizeof(md.md_provider));
685 md.md_provsize = cp->provider->mediasize;
686 journal_metadata_encode(&md, sector);
689 * Flush the cache, so we know all data are on disk.
690 * We write here informations like "journal is consistent", so we need
691 * to be sure it is. Without BIO_FLUSH here, we can end up in situation
692 * where metadata is stored on disk, but not all data.
694 g_journal_flush_cache(sc);
697 bp->bio_offset = cp->provider->mediasize - cp->provider->sectorsize;
698 bp->bio_length = cp->provider->sectorsize;
699 bp->bio_data = sector;
700 bp->bio_cmd = BIO_WRITE;
701 if (!(sc->sc_flags & GJF_DEVICE_DESTROY)) {
702 bp->bio_done = g_journal_metadata_done;
703 g_io_request(bp, cp);
706 g_io_request(bp, cp);
707 biowait(bp, "gjmdu");
708 g_journal_metadata_done(bp);
712 * Be sure metadata reached the disk.
714 g_journal_flush_cache(sc);
718 * This is where the I/O request comes from the GEOM.
721 g_journal_start(struct bio *bp)
723 struct g_journal_softc *sc;
725 sc = bp->bio_to->geom->softc;
726 GJ_LOGREQ(3, bp, "Request received.");
728 switch (bp->bio_cmd) {
731 mtx_lock(&sc->sc_mtx);
732 bioq_insert_tail(&sc->sc_regular_queue, bp);
734 mtx_unlock(&sc->sc_mtx);
737 if (strcmp(bp->bio_attribute, "GJOURNAL::provider") == 0) {
738 strlcpy(bp->bio_data, bp->bio_to->name, bp->bio_length);
739 bp->bio_completed = strlen(bp->bio_to->name) + 1;
746 g_io_deliver(bp, EOPNOTSUPP);
752 g_journal_std_done(struct bio *bp)
754 struct g_journal_softc *sc;
756 sc = bp->bio_from->geom->softc;
757 mtx_lock(&sc->sc_mtx);
758 bioq_insert_tail(&sc->sc_back_queue, bp);
760 mtx_unlock(&sc->sc_mtx);
764 g_journal_new_bio(off_t start, off_t end, off_t joffset, u_char *data,
770 bp->bio_offset = start;
771 bp->bio_joffset = joffset;
772 bp->bio_length = end - start;
773 bp->bio_cmd = BIO_WRITE;
774 bp->bio_done = g_journal_std_done;
778 bp->bio_data = gj_malloc(bp->bio_length, flags);
779 if (bp->bio_data != NULL)
780 bcopy(data, bp->bio_data, bp->bio_length);
785 #define g_journal_insert_bio(head, bp, flags) \
786 g_journal_insert((head), (bp)->bio_offset, \
787 (bp)->bio_offset + (bp)->bio_length, (bp)->bio_joffset, \
788 (bp)->bio_data, flags)
790 * The function below does a lot more than just inserting bio to the queue.
791 * It keeps the queue sorted by offset and ensures that there are no doubled
792 * data (it combines bios where ranges overlap).
794 * The function returns the number of bios inserted (as bio can be splitted).
797 g_journal_insert(struct bio **head, off_t nstart, off_t nend, off_t joffset,
798 u_char *data, int flags)
800 struct bio *nbp, *cbp, *pbp;
805 GJ_DEBUG(3, "INSERT(%p): (%jd, %jd, %jd)", *head, nstart, nend,
809 GJQ_FOREACH(*head, cbp) {
810 cstart = cbp->bio_offset;
811 cend = cbp->bio_offset + cbp->bio_length;
813 if (nstart >= cend) {
817 * | current | +-------------+
820 * +-------------+ | bio |
824 GJ_DEBUG(3, "INSERT(%p): 1", *head);
825 } else if (nend <= cstart) {
829 * +-------------+ | current |
832 * | bio | +-------------+
836 nbp = g_journal_new_bio(nstart, nend, joffset, data,
844 GJ_DEBUG(3, "INSERT(%p): 2 (nbp=%p pbp=%p)", *head, nbp,
847 } else if (nstart <= cstart && nend >= cend) {
849 * +-------------+ +-------------+
850 * | current bio | | current bio |
851 * +---+-------------+---+ +-------------+---+
854 * | +-------------+ | +-------------+ |
855 * | new bio | | new bio |
856 * +---------------------+ +-----------------+
858 * +-------------+ +-------------+
859 * | current bio | | current bio |
860 * +---+-------------+ +-------------+
863 * | +-------------+ +-------------+
864 * | new bio | | new bio |
865 * +-----------------+ +-------------+
867 g_journal_stats_bytes_skipped += cbp->bio_length;
868 cbp->bio_offset = nstart;
869 cbp->bio_joffset = joffset;
870 cbp->bio_length = cend - nstart;
871 if (cbp->bio_data != NULL) {
872 gj_free(cbp->bio_data, cend - cstart);
873 cbp->bio_data = NULL;
876 cbp->bio_data = gj_malloc(cbp->bio_length,
878 if (cbp->bio_data != NULL) {
879 bcopy(data, cbp->bio_data,
882 data += cend - nstart;
884 joffset += cend - nstart;
886 GJ_DEBUG(3, "INSERT(%p): 3 (cbp=%p)", *head, cbp);
887 } else if (nstart > cstart && nend >= cend) {
889 * +-----------------+ +-------------+
890 * | current bio | | current bio |
891 * | +-------------+ | +---------+---+
894 * +---+-------------+ +---+---------+ |
895 * | new bio | | new bio |
896 * +-------------+ +-------------+
898 g_journal_stats_bytes_skipped += cend - nstart;
899 nbp = g_journal_new_bio(nstart, cend, joffset, data,
901 nbp->bio_next = cbp->bio_next;
903 cbp->bio_length = nstart - cstart;
904 if (cbp->bio_data != NULL) {
905 cbp->bio_data = gj_realloc(cbp->bio_data,
906 cbp->bio_length, cend - cstart);
909 data += cend - nstart;
910 joffset += cend - nstart;
913 GJ_DEBUG(3, "INSERT(%p): 4 (cbp=%p)", *head, cbp);
914 } else if (nstart > cstart && nend < cend) {
916 * +---------------------+
918 * | +-------------+ |
921 * +---+-------------+---+
925 g_journal_stats_bytes_skipped += nend - nstart;
926 nbp = g_journal_new_bio(nstart, nend, joffset, data,
928 nbp->bio_next = cbp->bio_next;
930 if (cbp->bio_data == NULL)
933 tmpdata = cbp->bio_data + nend - cstart;
934 nbp = g_journal_new_bio(nend, cend,
935 cbp->bio_joffset + nend - cstart, tmpdata, flags);
936 nbp->bio_next = ((struct bio *)cbp->bio_next)->bio_next;
937 ((struct bio *)cbp->bio_next)->bio_next = nbp;
938 cbp->bio_length = nstart - cstart;
939 if (cbp->bio_data != NULL) {
940 cbp->bio_data = gj_realloc(cbp->bio_data,
941 cbp->bio_length, cend - cstart);
944 GJ_DEBUG(3, "INSERT(%p): 5 (cbp=%p)", *head, cbp);
946 } else if (nstart <= cstart && nend < cend) {
948 * +-----------------+ +-------------+
949 * | current bio | | current bio |
950 * +-------------+ | +---+---------+ |
953 * +-------------+---+ | +---------+---+
954 * | new bio | | new bio |
955 * +-------------+ +-------------+
957 g_journal_stats_bytes_skipped += nend - nstart;
958 nbp = g_journal_new_bio(nstart, nend, joffset, data,
965 cbp->bio_offset = nend;
966 cbp->bio_length = cend - nend;
967 cbp->bio_joffset += nend - cstart;
968 tmpdata = cbp->bio_data;
969 if (tmpdata != NULL) {
970 cbp->bio_data = gj_malloc(cbp->bio_length,
972 if (cbp->bio_data != NULL) {
973 bcopy(tmpdata + nend - cstart,
974 cbp->bio_data, cbp->bio_length);
976 gj_free(tmpdata, cend - cstart);
979 GJ_DEBUG(3, "INSERT(%p): 6 (cbp=%p)", *head, cbp);
986 nbp = g_journal_new_bio(nstart, nend, joffset, data, flags);
991 nbp->bio_next = NULL;
993 GJ_DEBUG(3, "INSERT(%p): 8 (nbp=%p pbp=%p)", *head, nbp, pbp);
995 if (g_journal_debug >= 3) {
996 GJQ_FOREACH(*head, cbp) {
997 GJ_DEBUG(3, "ELEMENT: %p (%jd, %jd, %jd, %p)", cbp,
998 (intmax_t)cbp->bio_offset,
999 (intmax_t)cbp->bio_length,
1000 (intmax_t)cbp->bio_joffset, cbp->bio_data);
1002 GJ_DEBUG(3, "INSERT(%p): DONE %d", *head, n);
1008 * The function combines neighbour bios trying to squeeze as much data as
1009 * possible into one bio.
1011 * The function returns the number of bios combined (negative value).
1014 g_journal_optimize(struct bio *head)
1016 struct bio *cbp, *pbp;
1021 GJQ_FOREACH(head, cbp) {
1022 /* Skip bios which has to be read first. */
1023 if (cbp->bio_data == NULL) {
1027 /* There is no previous bio yet. */
1032 /* Is this a neighbour bio? */
1033 if (pbp->bio_offset + pbp->bio_length != cbp->bio_offset) {
1034 /* Be sure that bios queue is sorted. */
1035 KASSERT(pbp->bio_offset + pbp->bio_length < cbp->bio_offset,
1036 ("poffset=%jd plength=%jd coffset=%jd",
1037 (intmax_t)pbp->bio_offset,
1038 (intmax_t)pbp->bio_length,
1039 (intmax_t)cbp->bio_offset));
1043 /* Be sure we don't end up with too big bio. */
1044 if (pbp->bio_length + cbp->bio_length > MAXPHYS) {
1048 /* Ok, we can join bios. */
1049 GJ_LOGREQ(4, pbp, "Join: ");
1050 GJ_LOGREQ(4, cbp, "and: ");
1051 pbp->bio_data = gj_realloc(pbp->bio_data,
1052 pbp->bio_length + cbp->bio_length, pbp->bio_length);
1053 bcopy(cbp->bio_data, pbp->bio_data + pbp->bio_length,
1055 gj_free(cbp->bio_data, cbp->bio_length);
1056 pbp->bio_length += cbp->bio_length;
1057 pbp->bio_next = cbp->bio_next;
1060 g_journal_stats_combined_ios++;
1062 GJ_LOGREQ(4, pbp, "Got: ");
1068 * TODO: Update comment.
1069 * These are functions responsible for copying one portion of data from journal
1070 * to the destination provider.
1071 * The order goes like this:
1072 * 1. Read the header, which contains informations about data blocks
1074 * 2. Read the data blocks from the journal.
1075 * 3. Write the data blocks on the data provider.
1077 * g_journal_copy_start()
1078 * g_journal_copy_done() - got finished write request, logs potential errors.
1082 * When there is no data in cache, this function is used to read it.
1085 g_journal_read_first(struct g_journal_softc *sc, struct bio *bp)
1090 * We were short in memory, so data was freed.
1091 * In that case we need to read it back from journal.
1093 cbp = g_alloc_bio();
1094 cbp->bio_cflags = bp->bio_cflags;
1095 cbp->bio_parent = bp;
1096 cbp->bio_offset = bp->bio_joffset;
1097 cbp->bio_length = bp->bio_length;
1098 cbp->bio_data = gj_malloc(bp->bio_length, M_WAITOK);
1099 cbp->bio_cmd = BIO_READ;
1100 cbp->bio_done = g_journal_std_done;
1101 GJ_LOGREQ(4, cbp, "READ FIRST");
1102 g_io_request(cbp, sc->sc_jconsumer);
1103 g_journal_cache_misses++;
1107 g_journal_copy_send(struct g_journal_softc *sc)
1109 struct bio *bioq, *bp, *lbp;
1112 mtx_lock(&sc->sc_mtx);
1113 for (; sc->sc_copy_in_progress < g_journal_parallel_copies;) {
1114 bp = GJQ_FIRST(sc->sc_inactive.jj_queue);
1117 GJQ_REMOVE(sc->sc_inactive.jj_queue, bp);
1118 sc->sc_copy_in_progress++;
1119 GJQ_INSERT_AFTER(bioq, bp, lbp);
1122 mtx_unlock(&sc->sc_mtx);
1123 if (g_journal_do_optimize)
1124 sc->sc_copy_in_progress += g_journal_optimize(bioq);
1125 while ((bp = GJQ_FIRST(bioq)) != NULL) {
1126 GJQ_REMOVE(bioq, bp);
1127 GJQ_INSERT_HEAD(sc->sc_copy_queue, bp);
1128 bp->bio_cflags = GJ_BIO_COPY;
1129 if (bp->bio_data == NULL)
1130 g_journal_read_first(sc, bp);
1132 bp->bio_joffset = 0;
1133 GJ_LOGREQ(4, bp, "SEND");
1134 g_io_request(bp, sc->sc_dconsumer);
1140 g_journal_copy_start(struct g_journal_softc *sc)
1144 * Remember in metadata that we're starting to copy journaled data
1145 * to the data provider.
1146 * In case of power failure, we will copy these data once again on boot.
1148 if (!sc->sc_journal_copying) {
1149 sc->sc_journal_copying = 1;
1150 GJ_DEBUG(1, "Starting copy of journal.");
1151 g_journal_metadata_update(sc);
1153 g_journal_copy_send(sc);
1157 * Data block has been read from the journal provider.
1160 g_journal_copy_read_done(struct bio *bp)
1162 struct g_journal_softc *sc;
1163 struct g_consumer *cp;
1166 KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1167 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1169 sc = bp->bio_from->geom->softc;
1170 pbp = bp->bio_parent;
1172 if (bp->bio_error != 0) {
1173 GJ_DEBUG(0, "Error while reading data from %s (error=%d).",
1174 bp->bio_to->name, bp->bio_error);
1176 * We will not be able to deliver WRITE request as well.
1178 gj_free(bp->bio_data, bp->bio_length);
1181 sc->sc_copy_in_progress--;
1184 pbp->bio_data = bp->bio_data;
1185 cp = sc->sc_dconsumer;
1186 g_io_request(pbp, cp);
1187 GJ_LOGREQ(4, bp, "READ DONE");
1193 * Data block has been written to the data provider.
1196 g_journal_copy_write_done(struct bio *bp)
1198 struct g_journal_softc *sc;
1200 KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1201 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1203 sc = bp->bio_from->geom->softc;
1204 sc->sc_copy_in_progress--;
1206 if (bp->bio_error != 0) {
1207 GJ_LOGREQ(0, bp, "[copy] Error while writing data (error=%d)",
1210 GJQ_REMOVE(sc->sc_copy_queue, bp);
1211 gj_free(bp->bio_data, bp->bio_length);
1212 GJ_LOGREQ(4, bp, "DONE");
1215 if (sc->sc_copy_in_progress == 0) {
1217 * This was the last write request for this journal.
1219 GJ_DEBUG(1, "Data has been copied.");
1220 sc->sc_journal_copying = 0;
1224 static void g_journal_flush_done(struct bio *bp);
1227 * Flush one record onto active journal provider.
1230 g_journal_flush(struct g_journal_softc *sc)
1232 struct g_journal_record_header hdr;
1233 struct g_journal_entry *ent;
1234 struct g_provider *pp;
1236 struct bio *bp, *fbp, *pbp;
1237 off_t joffset, size;
1238 u_char *data, hash[16];
1242 if (sc->sc_current_count == 0)
1246 pp = sc->sc_jprovider;
1247 GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1248 joffset = sc->sc_journal_offset;
1250 GJ_DEBUG(2, "Storing %d journal entries on %s at %jd.",
1251 sc->sc_current_count, pp->name, (intmax_t)joffset);
1254 * Store 'journal id', so we know to which journal this record belongs.
1256 hdr.jrh_journal_id = sc->sc_journal_id;
1257 /* Could be less than g_journal_record_entries if called due timeout. */
1258 hdr.jrh_nentries = MIN(sc->sc_current_count, g_journal_record_entries);
1259 strlcpy(hdr.jrh_magic, GJ_RECORD_HEADER_MAGIC, sizeof(hdr.jrh_magic));
1261 bioq = &sc->sc_active.jj_queue;
1262 pbp = sc->sc_flush_queue;
1264 fbp = g_alloc_bio();
1265 fbp->bio_parent = NULL;
1266 fbp->bio_cflags = GJ_BIO_JOURNAL;
1267 fbp->bio_offset = -1;
1268 fbp->bio_joffset = joffset;
1269 fbp->bio_length = pp->sectorsize;
1270 fbp->bio_cmd = BIO_WRITE;
1271 fbp->bio_done = g_journal_std_done;
1272 GJQ_INSERT_AFTER(sc->sc_flush_queue, fbp, pbp);
1275 GJ_LOGREQ(4, fbp, "FLUSH_OUT");
1276 joffset += pp->sectorsize;
1277 sc->sc_flush_count++;
1278 if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1281 for (i = 0; i < hdr.jrh_nentries; i++) {
1282 bp = sc->sc_current_queue;
1283 KASSERT(bp != NULL, ("NULL bp"));
1285 GJ_LOGREQ(4, bp, "FLUSHED");
1286 sc->sc_current_queue = bp->bio_next;
1287 bp->bio_next = NULL;
1288 sc->sc_current_count--;
1290 /* Add to the header. */
1291 ent = &hdr.jrh_entries[i];
1292 ent->je_offset = bp->bio_offset;
1293 ent->je_joffset = joffset;
1294 ent->je_length = bp->bio_length;
1295 size += ent->je_length;
1297 data = bp->bio_data;
1298 if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1299 MD5Update(&ctx, data, ent->je_length);
1300 bzero(bp, sizeof(*bp));
1301 bp->bio_cflags = GJ_BIO_JOURNAL;
1302 bp->bio_offset = ent->je_offset;
1303 bp->bio_joffset = ent->je_joffset;
1304 bp->bio_length = ent->je_length;
1305 bp->bio_data = data;
1306 bp->bio_cmd = BIO_WRITE;
1307 bp->bio_done = g_journal_std_done;
1308 GJQ_INSERT_AFTER(sc->sc_flush_queue, bp, pbp);
1311 GJ_LOGREQ(4, bp, "FLUSH_OUT");
1312 joffset += bp->bio_length;
1313 sc->sc_flush_count++;
1316 * Add request to the active sc_journal_queue queue.
1317 * This is our cache. After journal switch we don't have to
1318 * read the data from the inactive journal, because we keep
1321 g_journal_insert(bioq, ent->je_offset,
1322 ent->je_offset + ent->je_length, ent->je_joffset, data,
1327 * After all requests, store valid header.
1329 data = gj_malloc(pp->sectorsize, M_WAITOK);
1330 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1331 MD5Final(hash, &ctx);
1332 bcopy(hash, hdr.jrh_sum, sizeof(hdr.jrh_sum));
1334 g_journal_record_header_encode(&hdr, data);
1335 fbp->bio_data = data;
1337 sc->sc_journal_offset = joffset;
1339 g_journal_check_overflow(sc);
1343 * Flush request finished.
1346 g_journal_flush_done(struct bio *bp)
1348 struct g_journal_softc *sc;
1349 struct g_consumer *cp;
1351 KASSERT((bp->bio_cflags & GJ_BIO_MASK) == GJ_BIO_JOURNAL,
1352 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_JOURNAL));
1355 sc = cp->geom->softc;
1356 sc->sc_flush_in_progress--;
1358 if (bp->bio_error != 0) {
1359 GJ_LOGREQ(0, bp, "[flush] Error while writing data (error=%d)",
1362 gj_free(bp->bio_data, bp->bio_length);
1363 GJ_LOGREQ(4, bp, "DONE");
1367 static void g_journal_release_delayed(struct g_journal_softc *sc);
1370 g_journal_flush_send(struct g_journal_softc *sc)
1372 struct g_consumer *cp;
1373 struct bio *bioq, *bp, *lbp;
1375 cp = sc->sc_jconsumer;
1377 while (sc->sc_flush_in_progress < g_journal_parallel_flushes) {
1378 /* Send one flush requests to the active journal. */
1379 bp = GJQ_FIRST(sc->sc_flush_queue);
1381 GJQ_REMOVE(sc->sc_flush_queue, bp);
1382 sc->sc_flush_count--;
1383 bp->bio_offset = bp->bio_joffset;
1384 bp->bio_joffset = 0;
1385 sc->sc_flush_in_progress++;
1386 GJQ_INSERT_AFTER(bioq, bp, lbp);
1389 /* Try to release delayed requests. */
1390 g_journal_release_delayed(sc);
1391 /* If there are no requests to flush, leave. */
1392 if (GJQ_FIRST(sc->sc_flush_queue) == NULL)
1395 if (g_journal_do_optimize)
1396 sc->sc_flush_in_progress += g_journal_optimize(bioq);
1397 while ((bp = GJQ_FIRST(bioq)) != NULL) {
1398 GJQ_REMOVE(bioq, bp);
1399 GJ_LOGREQ(3, bp, "Flush request send");
1400 g_io_request(bp, cp);
1405 g_journal_add_current(struct g_journal_softc *sc, struct bio *bp)
1409 GJ_LOGREQ(4, bp, "CURRENT %d", sc->sc_current_count);
1410 n = g_journal_insert_bio(&sc->sc_current_queue, bp, M_WAITOK);
1411 sc->sc_current_count += n;
1412 n = g_journal_optimize(sc->sc_current_queue);
1413 sc->sc_current_count += n;
1415 * For requests which are added to the current queue we deliver
1416 * response immediately.
1418 bp->bio_completed = bp->bio_length;
1419 g_io_deliver(bp, 0);
1420 if (sc->sc_current_count >= g_journal_record_entries) {
1422 * Let's flush one record onto active journal provider.
1424 g_journal_flush(sc);
1429 g_journal_release_delayed(struct g_journal_softc *sc)
1434 /* The flush queue is full, exit. */
1435 if (sc->sc_flush_count >= g_journal_accept_immediately)
1437 bp = bioq_takefirst(&sc->sc_delayed_queue);
1440 sc->sc_delayed_count--;
1441 g_journal_add_current(sc, bp);
1446 * Add I/O request to the current queue. If we have enough requests for one
1447 * journal record we flush them onto active journal provider.
1450 g_journal_add_request(struct g_journal_softc *sc, struct bio *bp)
1454 * The flush queue is full, we need to delay the request.
1456 if (sc->sc_delayed_count > 0 ||
1457 sc->sc_flush_count >= g_journal_accept_immediately) {
1458 GJ_LOGREQ(4, bp, "DELAYED");
1459 bioq_insert_tail(&sc->sc_delayed_queue, bp);
1460 sc->sc_delayed_count++;
1464 KASSERT(TAILQ_EMPTY(&sc->sc_delayed_queue.queue),
1465 ("DELAYED queue not empty."));
1466 g_journal_add_current(sc, bp);
1469 static void g_journal_read_done(struct bio *bp);
1472 * Try to find requested data in cache.
1475 g_journal_read_find(struct bio *head, int sorted, struct bio *pbp, off_t ostart,
1481 GJQ_FOREACH(head, bp) {
1482 if (bp->bio_offset == -1)
1484 cstart = MAX(ostart, bp->bio_offset);
1485 cend = MIN(oend, bp->bio_offset + bp->bio_length);
1488 else if (cstart >= oend) {
1496 if (bp->bio_data == NULL)
1498 GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend,
1500 bcopy(bp->bio_data + cstart - bp->bio_offset,
1501 pbp->bio_data + cstart - pbp->bio_offset, cend - cstart);
1502 pbp->bio_completed += cend - cstart;
1503 if (pbp->bio_completed == pbp->bio_length) {
1505 * Cool, the whole request was in cache, deliver happy
1508 g_io_deliver(pbp, 0);
1517 * Try to find requested data in cache.
1520 g_journal_read_queue_find(struct bio_queue *head, struct bio *pbp, off_t ostart,
1526 TAILQ_FOREACH(bp, head, bio_queue) {
1527 cstart = MAX(ostart, bp->bio_offset);
1528 cend = MIN(oend, bp->bio_offset + bp->bio_length);
1531 else if (cstart >= oend)
1533 KASSERT(bp->bio_data != NULL,
1534 ("%s: bio_data == NULL", __func__));
1535 GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend,
1537 bcopy(bp->bio_data + cstart - bp->bio_offset,
1538 pbp->bio_data + cstart - pbp->bio_offset, cend - cstart);
1539 pbp->bio_completed += cend - cstart;
1540 if (pbp->bio_completed == pbp->bio_length) {
1542 * Cool, the whole request was in cache, deliver happy
1545 g_io_deliver(pbp, 0);
1554 * This function is used for colecting data on read.
1555 * The complexity is because parts of the data can be stored in four different
1557 * - in delayed requests
1558 * - in memory - the data not yet send to the active journal provider
1559 * - in requests which are going to be sent to the active journal
1560 * - in the active journal
1561 * - in the inactive journal
1562 * - in the data provider
1565 g_journal_read(struct g_journal_softc *sc, struct bio *pbp, off_t ostart,
1568 struct bio *bp, *nbp, *head;
1570 u_int i, sorted = 0;
1572 GJ_DEBUG(3, "READ: (%jd, %jd)", ostart, oend);
1577 for (i = 0; i <= 5; i++) {
1579 case 0: /* Delayed requests. */
1583 case 1: /* Not-yet-send data. */
1584 head = sc->sc_current_queue;
1587 case 2: /* In-flight to the active journal. */
1588 head = sc->sc_flush_queue;
1591 case 3: /* Active journal. */
1592 head = sc->sc_active.jj_queue;
1595 case 4: /* Inactive journal. */
1597 * XXX: Here could be a race with g_journal_lowmem().
1599 head = sc->sc_inactive.jj_queue;
1602 case 5: /* In-flight to the data provider. */
1603 head = sc->sc_copy_queue;
1607 panic("gjournal %s: i=%d", __func__, i);
1610 bp = g_journal_read_queue_find(&sc->sc_delayed_queue.queue, pbp, ostart, oend);
1612 bp = g_journal_read_find(head, sorted, pbp, ostart, oend);
1613 if (bp == pbp) { /* Got the whole request. */
1614 GJ_DEBUG(2, "Got the whole request from %u.", i);
1616 } else if (bp != NULL) {
1617 cstart = MAX(ostart, bp->bio_offset);
1618 cend = MIN(oend, bp->bio_offset + bp->bio_length);
1619 GJ_DEBUG(2, "Got part of the request from %u (%jd-%jd).",
1620 i, (intmax_t)cstart, (intmax_t)cend);
1625 if (bp->bio_data == NULL) {
1626 nbp = g_duplicate_bio(pbp);
1627 nbp->bio_cflags = GJ_BIO_READ;
1629 pbp->bio_data + cstart - pbp->bio_offset;
1631 bp->bio_joffset + cstart - bp->bio_offset;
1632 nbp->bio_length = cend - cstart;
1633 nbp->bio_done = g_journal_read_done;
1634 g_io_request(nbp, sc->sc_jconsumer);
1637 * If we don't have the whole request yet, call g_journal_read()
1640 if (ostart < cstart)
1641 g_journal_read(sc, pbp, ostart, cstart);
1643 g_journal_read(sc, pbp, cend, oend);
1646 * No data in memory, no data in journal.
1647 * Its time for asking data provider.
1649 GJ_DEBUG(3, "READ(data): (%jd, %jd)", ostart, oend);
1650 nbp = g_duplicate_bio(pbp);
1651 nbp->bio_cflags = GJ_BIO_READ;
1652 nbp->bio_data = pbp->bio_data + ostart - pbp->bio_offset;
1653 nbp->bio_offset = ostart;
1654 nbp->bio_length = oend - ostart;
1655 nbp->bio_done = g_journal_read_done;
1656 g_io_request(nbp, sc->sc_dconsumer);
1657 /* We have the whole request, return here. */
1663 * Function responsible for handling finished READ requests.
1664 * Actually, g_std_done() could be used here, the only difference is that we
1668 g_journal_read_done(struct bio *bp)
1672 KASSERT(bp->bio_cflags == GJ_BIO_READ,
1673 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_READ));
1675 pbp = bp->bio_parent;
1677 pbp->bio_completed += bp->bio_length;
1679 if (bp->bio_error != 0) {
1680 if (pbp->bio_error == 0)
1681 pbp->bio_error = bp->bio_error;
1682 GJ_DEBUG(0, "Error while reading data from %s (error=%d).",
1683 bp->bio_to->name, bp->bio_error);
1686 if (pbp->bio_children == pbp->bio_inbed &&
1687 pbp->bio_completed == pbp->bio_length) {
1689 g_io_deliver(pbp, 0);
1694 * Deactive current journal and active next one.
1697 g_journal_switch(struct g_journal_softc *sc)
1699 struct g_provider *pp;
1702 GJ_DEBUG(3, "No need for %s switch.", sc->sc_name);
1703 pp = LIST_FIRST(&sc->sc_geom->provider);
1704 if (!(sc->sc_flags & GJF_DEVICE_CLEAN) && pp->acw == 0) {
1705 sc->sc_flags |= GJF_DEVICE_CLEAN;
1706 GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
1707 g_journal_metadata_update(sc);
1710 GJ_DEBUG(3, "Switching journal %s.", sc->sc_geom->name);
1712 pp = sc->sc_jprovider;
1714 sc->sc_journal_previous_id = sc->sc_journal_id;
1716 sc->sc_journal_id = sc->sc_journal_next_id;
1717 sc->sc_journal_next_id = arc4random();
1719 GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1721 g_journal_write_header(sc);
1723 sc->sc_inactive.jj_offset = sc->sc_active.jj_offset;
1724 sc->sc_inactive.jj_queue = sc->sc_active.jj_queue;
1726 sc->sc_active.jj_offset =
1727 sc->sc_journal_offset - pp->sectorsize;
1728 sc->sc_active.jj_queue = NULL;
1731 * Switch is done, start copying data from the (now) inactive
1732 * journal to the data provider.
1734 g_journal_copy_start(sc);
1736 mtx_lock(&sc->sc_mtx);
1737 sc->sc_flags &= ~GJF_DEVICE_SWITCH;
1738 mtx_unlock(&sc->sc_mtx);
1742 g_journal_initialize(struct g_journal_softc *sc)
1745 sc->sc_journal_id = arc4random();
1746 sc->sc_journal_next_id = arc4random();
1747 sc->sc_journal_previous_id = sc->sc_journal_id;
1748 sc->sc_journal_offset = sc->sc_jstart;
1749 sc->sc_inactive.jj_offset = sc->sc_jstart;
1750 g_journal_write_header(sc);
1751 sc->sc_active.jj_offset = sc->sc_jstart;
1755 g_journal_mark_as_dirty(struct g_journal_softc *sc)
1757 const struct g_journal_desc *desc;
1760 GJ_DEBUG(1, "Marking file system %s as dirty.", sc->sc_name);
1761 for (i = 0; (desc = g_journal_filesystems[i]) != NULL; i++)
1762 desc->jd_dirty(sc->sc_dconsumer);
1766 * Function read record header from the given journal.
1767 * It is very simlar to g_read_data(9), but it doesn't allocate memory for bio
1768 * and data on every call.
1771 g_journal_sync_read(struct g_consumer *cp, struct bio *bp, off_t offset,
1776 bzero(bp, sizeof(*bp));
1777 bp->bio_cmd = BIO_READ;
1778 bp->bio_done = NULL;
1779 bp->bio_offset = offset;
1780 bp->bio_length = cp->provider->sectorsize;
1781 bp->bio_data = data;
1782 g_io_request(bp, cp);
1783 error = biowait(bp, "gjs_read");
1789 * Function is called when we start the journal device and we detect that
1790 * one of the journals was not fully copied.
1791 * The purpose of this function is to read all records headers from journal
1792 * and placed them in the inactive queue, so we can start journal
1793 * synchronization process and the journal provider itself.
1794 * Design decision was taken to not synchronize the whole journal here as it
1795 * can take too much time. Reading headers only and delaying synchronization
1796 * process until after journal provider is started should be the best choice.
1801 g_journal_sync(struct g_journal_softc *sc)
1803 struct g_journal_record_header rhdr;
1804 struct g_journal_entry *ent;
1805 struct g_journal_header jhdr;
1806 struct g_consumer *cp;
1807 struct bio *bp, *fbp, *tbp;
1808 off_t joffset, offset;
1809 u_char *buf, sum[16];
1812 int error, found, i;
1816 cp = sc->sc_jconsumer;
1818 buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
1819 offset = joffset = sc->sc_inactive.jj_offset = sc->sc_journal_offset;
1821 GJ_DEBUG(2, "Looking for termination at %jd.", (intmax_t)joffset);
1824 * Read and decode first journal header.
1826 error = g_journal_sync_read(cp, bp, offset, buf);
1828 GJ_DEBUG(0, "Error while reading journal header from %s.",
1829 cp->provider->name);
1832 error = g_journal_header_decode(buf, &jhdr);
1834 GJ_DEBUG(0, "Cannot decode journal header from %s.",
1835 cp->provider->name);
1838 id = sc->sc_journal_id;
1839 if (jhdr.jh_journal_id != sc->sc_journal_id) {
1840 GJ_DEBUG(1, "Journal ID mismatch at %jd (0x%08x != 0x%08x).",
1841 (intmax_t)offset, (u_int)jhdr.jh_journal_id, (u_int)id);
1844 offset += cp->provider->sectorsize;
1845 id = sc->sc_journal_next_id = jhdr.jh_journal_next_id;
1849 * If the biggest record won't fit, look for a record header or
1850 * journal header from the begining.
1852 GJ_VALIDATE_OFFSET(offset, sc);
1853 error = g_journal_sync_read(cp, bp, offset, buf);
1856 * Not good. Having an error while reading header
1857 * means, that we cannot read next headers and in
1858 * consequence we cannot find termination.
1861 "Error while reading record header from %s.",
1862 cp->provider->name);
1866 error = g_journal_record_header_decode(buf, &rhdr);
1868 GJ_DEBUG(2, "Not a record header at %jd (error=%d).",
1869 (intmax_t)offset, error);
1871 * This is not a record header.
1872 * If we are lucky, this is next journal header.
1874 error = g_journal_header_decode(buf, &jhdr);
1876 GJ_DEBUG(1, "Not a journal header at %jd (error=%d).",
1877 (intmax_t)offset, error);
1879 * Nope, this is not journal header, which
1880 * bascially means that journal is not
1881 * terminated properly.
1887 * Ok. This is header of _some_ journal. Now we need to
1888 * verify if this is header of the _next_ journal.
1890 if (jhdr.jh_journal_id != id) {
1891 GJ_DEBUG(1, "Journal ID mismatch at %jd "
1892 "(0x%08x != 0x%08x).", (intmax_t)offset,
1893 (u_int)jhdr.jh_journal_id, (u_int)id);
1898 /* Found termination. */
1900 GJ_DEBUG(1, "Found termination at %jd (id=0x%08x).",
1901 (intmax_t)offset, (u_int)id);
1902 sc->sc_active.jj_offset = offset;
1903 sc->sc_journal_offset =
1904 offset + cp->provider->sectorsize;
1905 sc->sc_journal_id = id;
1906 id = sc->sc_journal_next_id = jhdr.jh_journal_next_id;
1908 while ((tbp = fbp) != NULL) {
1909 fbp = tbp->bio_next;
1910 GJ_LOGREQ(3, tbp, "Adding request.");
1911 g_journal_insert_bio(&sc->sc_inactive.jj_queue,
1915 /* Skip journal's header. */
1916 offset += cp->provider->sectorsize;
1920 /* Skip record's header. */
1921 offset += cp->provider->sectorsize;
1924 * Add information about every record entry to the inactive
1927 if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1929 for (i = 0; i < rhdr.jrh_nentries; i++) {
1930 ent = &rhdr.jrh_entries[i];
1931 GJ_DEBUG(3, "Insert entry: %jd %jd.",
1932 (intmax_t)ent->je_offset, (intmax_t)ent->je_length);
1933 g_journal_insert(&fbp, ent->je_offset,
1934 ent->je_offset + ent->je_length, ent->je_joffset,
1936 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1940 * TODO: Should use faster function (like
1941 * g_journal_sync_read()).
1943 buf2 = g_read_data(cp, offset, ent->je_length,
1946 GJ_DEBUG(0, "Cannot read data at %jd.",
1949 MD5Update(&ctx, buf2, ent->je_length);
1953 /* Skip entry's data. */
1954 offset += ent->je_length;
1956 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1957 MD5Final(sum, &ctx);
1958 if (bcmp(sum, rhdr.jrh_sum, sizeof(rhdr.jrh_sum)) != 0) {
1959 GJ_DEBUG(0, "MD5 hash mismatch at %jd!",
1965 gj_free(bp->bio_data, cp->provider->sectorsize);
1968 /* Remove bios from unterminated journal. */
1969 while ((tbp = fbp) != NULL) {
1970 fbp = tbp->bio_next;
1974 if (found < 1 && joffset > 0) {
1975 GJ_DEBUG(0, "Journal on %s is broken/corrupted. Initializing.",
1977 while ((tbp = sc->sc_inactive.jj_queue) != NULL) {
1978 sc->sc_inactive.jj_queue = tbp->bio_next;
1981 g_journal_initialize(sc);
1982 g_journal_mark_as_dirty(sc);
1984 GJ_DEBUG(0, "Journal %s consistent.", sc->sc_name);
1985 g_journal_copy_start(sc);
1990 * Wait for requests.
1991 * If we have requests in the current queue, flush them after 3 seconds from the
1992 * last flush. In this way we don't wait forever (or for journal switch) with
1993 * storing not full records on journal.
1996 g_journal_wait(struct g_journal_softc *sc, time_t last_write)
2000 GJ_DEBUG(3, "%s: enter", __func__);
2001 if (sc->sc_current_count == 0) {
2002 if (g_journal_debug < 2)
2003 msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", 0);
2006 * If we have debug turned on, show number of elements
2007 * in various queues.
2010 error = msleep(sc, &sc->sc_mtx, PRIBIO,
2013 mtx_unlock(&sc->sc_mtx);
2016 GJ_DEBUG(3, "Report: current count=%d",
2017 sc->sc_current_count);
2018 GJ_DEBUG(3, "Report: flush count=%d",
2019 sc->sc_flush_count);
2020 GJ_DEBUG(3, "Report: flush in progress=%d",
2021 sc->sc_flush_in_progress);
2022 GJ_DEBUG(3, "Report: copy in progress=%d",
2023 sc->sc_copy_in_progress);
2024 GJ_DEBUG(3, "Report: delayed=%d",
2025 sc->sc_delayed_count);
2028 GJ_DEBUG(3, "%s: exit 1", __func__);
2033 * Flush even not full records every 3 seconds.
2035 timeout = (last_write + 3 - time_second) * hz;
2037 mtx_unlock(&sc->sc_mtx);
2038 g_journal_flush(sc);
2039 g_journal_flush_send(sc);
2040 GJ_DEBUG(3, "%s: exit 2", __func__);
2043 error = msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", timeout);
2044 if (error == EWOULDBLOCK)
2045 g_journal_flush_send(sc);
2046 GJ_DEBUG(3, "%s: exit 3", __func__);
2053 g_journal_worker(void *arg)
2055 struct g_journal_softc *sc;
2057 struct g_provider *pp;
2062 thread_lock(curthread);
2063 sched_prio(curthread, PRIBIO);
2064 thread_unlock(curthread);
2069 if (sc->sc_flags & GJF_DEVICE_CLEAN) {
2070 GJ_DEBUG(0, "Journal %s clean.", sc->sc_name);
2071 g_journal_initialize(sc);
2076 * Check if we can use BIO_FLUSH.
2078 sc->sc_bio_flush = 0;
2079 if (g_io_flush(sc->sc_jconsumer) == 0) {
2080 sc->sc_bio_flush |= GJ_FLUSH_JOURNAL;
2081 GJ_DEBUG(1, "BIO_FLUSH supported by %s.",
2082 sc->sc_jconsumer->provider->name);
2084 GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2085 sc->sc_jconsumer->provider->name);
2087 if (sc->sc_jconsumer != sc->sc_dconsumer) {
2088 if (g_io_flush(sc->sc_dconsumer) == 0) {
2089 sc->sc_bio_flush |= GJ_FLUSH_DATA;
2090 GJ_DEBUG(1, "BIO_FLUSH supported by %s.",
2091 sc->sc_dconsumer->provider->name);
2093 GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2094 sc->sc_dconsumer->provider->name);
2100 pp = g_new_providerf(gp, "%s.journal", sc->sc_name);
2101 pp->mediasize = sc->sc_mediasize;
2103 * There could be a problem when data provider and journal providers
2104 * have different sectorsize, but such scenario is prevented on journal
2107 pp->sectorsize = sc->sc_sectorsize;
2108 g_error_provider(pp, 0);
2109 g_topology_unlock();
2110 last_write = time_second;
2112 if (sc->sc_rootmount != NULL) {
2113 GJ_DEBUG(1, "root_mount_rel %p", sc->sc_rootmount);
2114 root_mount_rel(sc->sc_rootmount);
2115 sc->sc_rootmount = NULL;
2119 /* Get first request from the queue. */
2120 mtx_lock(&sc->sc_mtx);
2121 bp = bioq_first(&sc->sc_back_queue);
2123 type = (bp->bio_cflags & GJ_BIO_MASK);
2125 bp = bioq_first(&sc->sc_regular_queue);
2127 type = GJ_BIO_REGULAR;
2131 if ((sc->sc_flags & GJF_DEVICE_SWITCH) ||
2132 (sc->sc_flags & GJF_DEVICE_DESTROY)) {
2133 if (sc->sc_current_count > 0) {
2134 mtx_unlock(&sc->sc_mtx);
2135 g_journal_flush(sc);
2136 g_journal_flush_send(sc);
2139 if (sc->sc_flush_in_progress > 0)
2141 if (sc->sc_copy_in_progress > 0)
2144 if (sc->sc_flags & GJF_DEVICE_SWITCH) {
2145 mtx_unlock(&sc->sc_mtx);
2146 g_journal_switch(sc);
2147 wakeup(&sc->sc_journal_copying);
2150 if (sc->sc_flags & GJF_DEVICE_DESTROY) {
2151 GJ_DEBUG(1, "Shutting down worker "
2152 "thread for %s.", gp->name);
2153 sc->sc_worker = NULL;
2154 wakeup(&sc->sc_worker);
2155 mtx_unlock(&sc->sc_mtx);
2159 g_journal_wait(sc, last_write);
2163 * If we're in switch process, we need to delay all new
2164 * write requests until its done.
2166 if ((sc->sc_flags & GJF_DEVICE_SWITCH) &&
2167 type == GJ_BIO_REGULAR && bp->bio_cmd == BIO_WRITE) {
2168 GJ_LOGREQ(2, bp, "WRITE on SWITCH");
2171 if (type == GJ_BIO_REGULAR)
2172 bioq_remove(&sc->sc_regular_queue, bp);
2174 bioq_remove(&sc->sc_back_queue, bp);
2175 mtx_unlock(&sc->sc_mtx);
2177 case GJ_BIO_REGULAR:
2178 /* Regular request. */
2179 switch (bp->bio_cmd) {
2181 g_journal_read(sc, bp, bp->bio_offset,
2182 bp->bio_offset + bp->bio_length);
2185 last_write = time_second;
2186 g_journal_add_request(sc, bp);
2187 g_journal_flush_send(sc);
2190 panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2194 switch (bp->bio_cmd) {
2196 if (g_journal_copy_read_done(bp))
2197 g_journal_copy_send(sc);
2200 g_journal_copy_write_done(bp);
2201 g_journal_copy_send(sc);
2204 panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2207 case GJ_BIO_JOURNAL:
2208 g_journal_flush_done(bp);
2209 g_journal_flush_send(sc);
2213 panic("Invalid bio (%d).", type);
2219 g_journal_destroy_event(void *arg, int flags __unused)
2221 struct g_journal_softc *sc;
2223 g_topology_assert();
2225 g_journal_destroy(sc);
2229 g_journal_timeout(void *arg)
2231 struct g_journal_softc *sc;
2234 GJ_DEBUG(0, "Timeout. Journal %s cannot be completed.",
2236 g_post_event(g_journal_destroy_event, sc, M_NOWAIT, NULL);
2239 static struct g_geom *
2240 g_journal_create(struct g_class *mp, struct g_provider *pp,
2241 const struct g_journal_metadata *md)
2243 struct g_journal_softc *sc;
2245 struct g_consumer *cp;
2248 sc = NULL; /* gcc */
2250 g_topology_assert();
2252 * There are two possibilities:
2253 * 1. Data and both journals are on the same provider.
2254 * 2. Data and journals are all on separated providers.
2256 /* Look for journal device with the same ID. */
2257 LIST_FOREACH(gp, &mp->geom, geom) {
2261 if (sc->sc_id == md->md_id)
2266 else if (sc != NULL && (sc->sc_type & md->md_type) != 0) {
2267 GJ_DEBUG(1, "Journal device %u already configured.", sc->sc_id);
2270 if (md->md_type == 0 || (md->md_type & ~GJ_TYPE_COMPLETE) != 0) {
2271 GJ_DEBUG(0, "Invalid type on %s.", pp->name);
2274 if (md->md_type & GJ_TYPE_DATA) {
2275 GJ_DEBUG(0, "Journal %u: %s contains data.", md->md_id,
2278 if (md->md_type & GJ_TYPE_JOURNAL) {
2279 GJ_DEBUG(0, "Journal %u: %s contains journal.", md->md_id,
2285 sc = malloc(sizeof(*sc), M_JOURNAL, M_WAITOK | M_ZERO);
2286 sc->sc_id = md->md_id;
2289 sc->sc_worker = NULL;
2291 gp = g_new_geomf(mp, "gjournal %u", sc->sc_id);
2292 gp->start = g_journal_start;
2293 gp->orphan = g_journal_orphan;
2294 gp->access = g_journal_access;
2296 gp->flags |= G_GEOM_VOLATILE_BIO;
2299 mtx_init(&sc->sc_mtx, "gjournal", NULL, MTX_DEF);
2301 bioq_init(&sc->sc_back_queue);
2302 bioq_init(&sc->sc_regular_queue);
2303 bioq_init(&sc->sc_delayed_queue);
2304 sc->sc_delayed_count = 0;
2305 sc->sc_current_queue = NULL;
2306 sc->sc_current_count = 0;
2307 sc->sc_flush_queue = NULL;
2308 sc->sc_flush_count = 0;
2309 sc->sc_flush_in_progress = 0;
2310 sc->sc_copy_queue = NULL;
2311 sc->sc_copy_in_progress = 0;
2312 sc->sc_inactive.jj_queue = NULL;
2313 sc->sc_active.jj_queue = NULL;
2315 sc->sc_rootmount = root_mount_hold("GJOURNAL");
2316 GJ_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2318 callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
2319 if (md->md_type != GJ_TYPE_COMPLETE) {
2321 * Journal and data are on separate providers.
2322 * At this point we have only one of them.
2323 * We setup a timeout in case the other part will not
2324 * appear, so we won't wait forever.
2326 callout_reset(&sc->sc_callout, 5 * hz,
2327 g_journal_timeout, sc);
2331 /* Remember type of the data provider. */
2332 if (md->md_type & GJ_TYPE_DATA)
2333 sc->sc_orig_type = md->md_type;
2334 sc->sc_type |= md->md_type;
2337 if (md->md_type & GJ_TYPE_DATA) {
2338 if (md->md_flags & GJ_FLAG_CLEAN)
2339 sc->sc_flags |= GJF_DEVICE_CLEAN;
2340 if (md->md_flags & GJ_FLAG_CHECKSUM)
2341 sc->sc_flags |= GJF_DEVICE_CHECKSUM;
2342 cp = g_new_consumer(gp);
2343 error = g_attach(cp, pp);
2344 KASSERT(error == 0, ("Cannot attach to %s (error=%d).",
2346 error = g_access(cp, 1, 1, 1);
2348 GJ_DEBUG(0, "Cannot access %s (error=%d).", pp->name,
2350 g_journal_destroy(sc);
2353 sc->sc_dconsumer = cp;
2354 sc->sc_mediasize = pp->mediasize - pp->sectorsize;
2355 sc->sc_sectorsize = pp->sectorsize;
2356 sc->sc_jstart = md->md_jstart;
2357 sc->sc_jend = md->md_jend;
2358 if (md->md_provider[0] != '\0')
2359 sc->sc_flags |= GJF_DEVICE_HARDCODED;
2360 sc->sc_journal_offset = md->md_joffset;
2361 sc->sc_journal_id = md->md_jid;
2362 sc->sc_journal_previous_id = md->md_jid;
2364 if (md->md_type & GJ_TYPE_JOURNAL) {
2366 cp = g_new_consumer(gp);
2367 error = g_attach(cp, pp);
2368 KASSERT(error == 0, ("Cannot attach to %s (error=%d).",
2370 error = g_access(cp, 1, 1, 1);
2372 GJ_DEBUG(0, "Cannot access %s (error=%d).",
2374 g_journal_destroy(sc);
2379 * Journal is on the same provider as data, which means
2380 * that data provider ends where journal starts.
2382 sc->sc_mediasize = md->md_jstart;
2384 sc->sc_jconsumer = cp;
2387 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE) {
2388 /* Journal is not complete yet. */
2391 /* Journal complete, cancel timeout. */
2392 callout_drain(&sc->sc_callout);
2395 error = kproc_create(g_journal_worker, sc, &sc->sc_worker, 0, 0,
2396 "g_journal %s", sc->sc_name);
2398 GJ_DEBUG(0, "Cannot create worker thread for %s.journal.",
2400 g_journal_destroy(sc);
2408 g_journal_destroy_consumer(void *arg, int flags __unused)
2410 struct g_consumer *cp;
2412 g_topology_assert();
2415 g_destroy_consumer(cp);
2419 g_journal_destroy(struct g_journal_softc *sc)
2422 struct g_provider *pp;
2423 struct g_consumer *cp;
2425 g_topology_assert();
2431 pp = LIST_FIRST(&gp->provider);
2433 if (pp->acr != 0 || pp->acw != 0 || pp->ace != 0) {
2434 GJ_DEBUG(1, "Device %s is still open (r%dw%de%d).",
2435 pp->name, pp->acr, pp->acw, pp->ace);
2438 g_error_provider(pp, ENXIO);
2440 g_journal_flush(sc);
2441 g_journal_flush_send(sc);
2442 g_journal_switch(sc);
2445 sc->sc_flags |= (GJF_DEVICE_DESTROY | GJF_DEVICE_CLEAN);
2447 g_topology_unlock();
2449 if (sc->sc_rootmount != NULL) {
2450 GJ_DEBUG(1, "root_mount_rel %p", sc->sc_rootmount);
2451 root_mount_rel(sc->sc_rootmount);
2452 sc->sc_rootmount = NULL;
2455 callout_drain(&sc->sc_callout);
2456 mtx_lock(&sc->sc_mtx);
2458 while (sc->sc_worker != NULL)
2459 msleep(&sc->sc_worker, &sc->sc_mtx, PRIBIO, "gj:destroy", 0);
2460 mtx_unlock(&sc->sc_mtx);
2463 GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
2464 g_journal_metadata_update(sc);
2466 pp->flags |= G_PF_WITHER;
2467 g_orphan_provider(pp, ENXIO);
2471 mtx_destroy(&sc->sc_mtx);
2473 if (sc->sc_current_count != 0) {
2474 GJ_DEBUG(0, "Warning! Number of current requests %d.",
2475 sc->sc_current_count);
2478 LIST_FOREACH(cp, &gp->consumer, consumer) {
2479 if (cp->acr + cp->acw + cp->ace > 0)
2480 g_access(cp, -1, -1, -1);
2482 * We keep all consumers open for writting, so if I'll detach
2483 * and destroy consumer here, I'll get providers for taste, so
2484 * journal will be started again.
2485 * Sending an event here, prevents this from happening.
2487 g_post_event(g_journal_destroy_consumer, cp, M_WAITOK, NULL);
2490 g_wither_geom(gp, ENXIO);
2491 free(sc, M_JOURNAL);
2496 g_journal_taste_orphan(struct g_consumer *cp)
2499 KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2500 cp->provider->name));
2503 static struct g_geom *
2504 g_journal_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2506 struct g_journal_metadata md;
2507 struct g_consumer *cp;
2511 g_topology_assert();
2512 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2513 GJ_DEBUG(2, "Tasting %s.", pp->name);
2514 if (pp->geom->class == mp)
2517 gp = g_new_geomf(mp, "journal:taste");
2518 /* This orphan function should be never called. */
2519 gp->orphan = g_journal_taste_orphan;
2520 cp = g_new_consumer(gp);
2522 error = g_journal_metadata_read(cp, &md);
2524 g_destroy_consumer(cp);
2530 if (md.md_provider[0] != '\0' &&
2531 !g_compare_names(md.md_provider, pp->name))
2533 if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
2535 if (g_journal_debug >= 2)
2536 journal_metadata_dump(&md);
2538 gp = g_journal_create(mp, pp, &md);
2542 static struct g_journal_softc *
2543 g_journal_find_device(struct g_class *mp, const char *name)
2545 struct g_journal_softc *sc;
2547 struct g_provider *pp;
2549 if (strncmp(name, "/dev/", 5) == 0)
2551 LIST_FOREACH(gp, &mp->geom, geom) {
2555 if (sc->sc_flags & GJF_DEVICE_DESTROY)
2557 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2559 pp = LIST_FIRST(&gp->provider);
2560 if (strcmp(sc->sc_name, name) == 0)
2562 if (pp != NULL && strcmp(pp->name, name) == 0)
2569 g_journal_ctl_destroy(struct gctl_req *req, struct g_class *mp)
2571 struct g_journal_softc *sc;
2577 g_topology_assert();
2579 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
2580 if (nargs == NULL) {
2581 gctl_error(req, "No '%s' argument.", "nargs");
2585 gctl_error(req, "Missing device(s).");
2589 for (i = 0; i < *nargs; i++) {
2590 snprintf(param, sizeof(param), "arg%d", i);
2591 name = gctl_get_asciiparam(req, param);
2593 gctl_error(req, "No 'arg%d' argument.", i);
2596 sc = g_journal_find_device(mp, name);
2598 gctl_error(req, "No such device: %s.", name);
2601 error = g_journal_destroy(sc);
2603 gctl_error(req, "Cannot destroy device %s (error=%d).",
2604 LIST_FIRST(&sc->sc_geom->provider)->name, error);
2611 g_journal_ctl_sync(struct gctl_req *req __unused, struct g_class *mp __unused)
2614 g_topology_assert();
2615 g_topology_unlock();
2616 g_journal_sync_requested++;
2617 wakeup(&g_journal_switcher_state);
2618 while (g_journal_sync_requested > 0)
2619 tsleep(&g_journal_sync_requested, PRIBIO, "j:sreq", hz / 2);
2624 g_journal_config(struct gctl_req *req, struct g_class *mp, const char *verb)
2628 g_topology_assert();
2630 version = gctl_get_paraml(req, "version", sizeof(*version));
2631 if (version == NULL) {
2632 gctl_error(req, "No '%s' argument.", "version");
2635 if (*version != G_JOURNAL_VERSION) {
2636 gctl_error(req, "Userland and kernel parts are out of sync.");
2640 if (strcmp(verb, "destroy") == 0 || strcmp(verb, "stop") == 0) {
2641 g_journal_ctl_destroy(req, mp);
2643 } else if (strcmp(verb, "sync") == 0) {
2644 g_journal_ctl_sync(req, mp);
2648 gctl_error(req, "Unknown verb.");
2652 g_journal_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
2653 struct g_consumer *cp, struct g_provider *pp)
2655 struct g_journal_softc *sc;
2657 g_topology_assert();
2664 } else if (cp != NULL) {
2667 sbuf_printf(sb, "%s<Role>", indent);
2668 if (cp == sc->sc_dconsumer) {
2669 sbuf_printf(sb, "Data");
2672 if (cp == sc->sc_jconsumer) {
2674 sbuf_printf(sb, ",");
2675 sbuf_printf(sb, "Journal");
2677 sbuf_printf(sb, "</Role>\n");
2678 if (cp == sc->sc_jconsumer) {
2679 sbuf_printf(sb, "<Jstart>%jd</Jstart>\n",
2680 (intmax_t)sc->sc_jstart);
2681 sbuf_printf(sb, "<Jend>%jd</Jend>\n",
2682 (intmax_t)sc->sc_jend);
2685 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
2689 static eventhandler_tag g_journal_event_shutdown = NULL;
2690 static eventhandler_tag g_journal_event_lowmem = NULL;
2693 g_journal_shutdown(void *arg, int howto __unused)
2696 struct g_geom *gp, *gp2;
2698 if (panicstr != NULL)
2703 LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
2704 if (gp->softc == NULL)
2706 GJ_DEBUG(0, "Shutting down geom %s.", gp->name);
2707 g_journal_destroy(gp->softc);
2709 g_topology_unlock();
2714 * Free cached requests from inactive queue in case of low memory.
2715 * We free GJ_FREE_AT_ONCE elements at once.
2717 #define GJ_FREE_AT_ONCE 4
2719 g_journal_lowmem(void *arg, int howto __unused)
2721 struct g_journal_softc *sc;
2725 u_int nfree = GJ_FREE_AT_ONCE;
2727 g_journal_stats_low_mem++;
2731 LIST_FOREACH(gp, &mp->geom, geom) {
2733 if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY))
2735 mtx_lock(&sc->sc_mtx);
2736 for (bp = sc->sc_inactive.jj_queue; nfree > 0 && bp != NULL;
2737 nfree--, bp = bp->bio_next) {
2739 * This is safe to free the bio_data, because:
2740 * 1. If bio_data is NULL it will be read from the
2742 * 2. If bp is sent down, it is first removed from the
2743 * inactive queue, so it's impossible to free the
2744 * data from under in-flight bio.
2745 * On the other hand, freeing elements from the active
2746 * queue, is not safe.
2748 if (bp->bio_data != NULL) {
2749 GJ_DEBUG(2, "Freeing data from %s.",
2751 gj_free(bp->bio_data, bp->bio_length);
2752 bp->bio_data = NULL;
2755 mtx_unlock(&sc->sc_mtx);
2759 g_topology_unlock();
2763 static void g_journal_switcher(void *arg);
2766 g_journal_init(struct g_class *mp)
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;
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;
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 error = kproc_create(g_journal_switcher, mp, NULL, 0, 0,
2789 "g_journal switcher");
2790 KASSERT(error == 0, ("Cannot create switcher thread."));
2794 g_journal_fini(struct g_class *mp)
2797 if (g_journal_event_shutdown != NULL) {
2798 EVENTHANDLER_DEREGISTER(shutdown_post_sync,
2799 g_journal_event_shutdown);
2801 if (g_journal_event_lowmem != NULL)
2802 EVENTHANDLER_DEREGISTER(vm_lowmem, g_journal_event_lowmem);
2803 g_journal_switcher_state = GJ_SWITCHER_DIE;
2804 wakeup(&g_journal_switcher_state);
2805 while (g_journal_switcher_state != GJ_SWITCHER_DIED)
2806 tsleep(&g_journal_switcher_state, PRIBIO, "jfini:wait", hz / 5);
2807 GJ_DEBUG(1, "Switcher died.");
2810 DECLARE_GEOM_CLASS(g_journal_class, g_journal);
2812 static const struct g_journal_desc *
2813 g_journal_find_desc(const char *fstype)
2815 const struct g_journal_desc *desc;
2818 for (desc = g_journal_filesystems[i = 0]; desc != NULL;
2819 desc = g_journal_filesystems[++i]) {
2820 if (strcmp(desc->jd_fstype, fstype) == 0)
2827 g_journal_switch_wait(struct g_journal_softc *sc)
2831 mtx_assert(&sc->sc_mtx, MA_OWNED);
2832 if (g_journal_debug >= 2) {
2833 if (sc->sc_flush_in_progress > 0) {
2834 GJ_DEBUG(2, "%d requests flushing.",
2835 sc->sc_flush_in_progress);
2837 if (sc->sc_copy_in_progress > 0) {
2838 GJ_DEBUG(2, "%d requests copying.",
2839 sc->sc_copy_in_progress);
2841 if (sc->sc_flush_count > 0) {
2842 GJ_DEBUG(2, "%d requests to flush.",
2843 sc->sc_flush_count);
2845 if (sc->sc_delayed_count > 0) {
2846 GJ_DEBUG(2, "%d requests delayed.",
2847 sc->sc_delayed_count);
2850 g_journal_stats_switches++;
2851 if (sc->sc_copy_in_progress > 0)
2852 g_journal_stats_wait_for_copy++;
2853 GJ_TIMER_START(1, &bt);
2854 sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2855 sc->sc_flags |= GJF_DEVICE_SWITCH;
2857 while (sc->sc_flags & GJF_DEVICE_SWITCH) {
2858 msleep(&sc->sc_journal_copying, &sc->sc_mtx, PRIBIO,
2861 GJ_TIMER_STOP(1, &bt, "Switch time of %s", sc->sc_name);
2865 g_journal_do_switch(struct g_class *classp)
2867 struct g_journal_softc *sc;
2868 const struct g_journal_desc *desc;
2873 int error, save, vfslocked;
2877 LIST_FOREACH(gp, &classp->geom, geom) {
2881 if (sc->sc_flags & GJF_DEVICE_DESTROY)
2883 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2885 mtx_lock(&sc->sc_mtx);
2886 sc->sc_flags |= GJF_DEVICE_BEFORE_SWITCH;
2887 mtx_unlock(&sc->sc_mtx);
2889 g_topology_unlock();
2892 mtx_lock(&mountlist_mtx);
2893 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2894 if (mp->mnt_gjprovider == NULL)
2896 if (mp->mnt_flag & MNT_RDONLY)
2898 desc = g_journal_find_desc(mp->mnt_stat.f_fstypename);
2901 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
2903 /* mtx_unlock(&mountlist_mtx) was done inside vfs_busy() */
2907 sc = g_journal_find_device(classp, mp->mnt_gjprovider);
2908 g_topology_unlock();
2912 GJ_DEBUG(0, "Cannot find journal geom for %s.",
2913 mp->mnt_gjprovider);
2915 } else if (JEMPTY(sc)) {
2916 mtx_lock(&sc->sc_mtx);
2917 sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2918 mtx_unlock(&sc->sc_mtx);
2919 GJ_DEBUG(3, "No need for %s switch.", sc->sc_name);
2923 mountpoint = mp->mnt_stat.f_mntonname;
2925 vfslocked = VFS_LOCK_GIANT(mp);
2927 error = vn_start_write(NULL, &mp, V_WAIT);
2929 VFS_UNLOCK_GIANT(vfslocked);
2930 GJ_DEBUG(0, "vn_start_write(%s) failed (error=%d).",
2935 save = curthread_pflags_set(TDP_SYNCIO);
2937 GJ_TIMER_START(1, &bt);
2938 vfs_msync(mp, MNT_NOWAIT);
2939 GJ_TIMER_STOP(1, &bt, "Msync time of %s", mountpoint);
2941 GJ_TIMER_START(1, &bt);
2942 error = VFS_SYNC(mp, MNT_NOWAIT);
2944 GJ_TIMER_STOP(1, &bt, "Sync time of %s", mountpoint);
2946 GJ_DEBUG(0, "Cannot sync file system %s (error=%d).",
2950 curthread_pflags_restore(save);
2952 vn_finished_write(mp);
2955 VFS_UNLOCK_GIANT(vfslocked);
2960 * Send BIO_FLUSH before freezing the file system, so it can be
2961 * faster after the freeze.
2963 GJ_TIMER_START(1, &bt);
2964 g_journal_flush_cache(sc);
2965 GJ_TIMER_STOP(1, &bt, "BIO_FLUSH time of %s", sc->sc_name);
2967 GJ_TIMER_START(1, &bt);
2968 error = vfs_write_suspend(mp);
2969 VFS_UNLOCK_GIANT(vfslocked);
2970 GJ_TIMER_STOP(1, &bt, "Suspend time of %s", mountpoint);
2972 GJ_DEBUG(0, "Cannot suspend file system %s (error=%d).",
2977 error = desc->jd_clean(mp);
2981 mtx_lock(&sc->sc_mtx);
2982 g_journal_switch_wait(sc);
2983 mtx_unlock(&sc->sc_mtx);
2985 vfs_write_resume(mp);
2987 mtx_lock(&mountlist_mtx);
2990 mtx_unlock(&mountlist_mtx);
2996 LIST_FOREACH(gp, &g_journal_class.geom, geom) {
3000 mtx_lock(&sc->sc_mtx);
3001 if ((sc->sc_type & GJ_TYPE_COMPLETE) == GJ_TYPE_COMPLETE &&
3002 !(sc->sc_flags & GJF_DEVICE_DESTROY) &&
3003 (sc->sc_flags & GJF_DEVICE_BEFORE_SWITCH)) {
3006 mtx_unlock(&sc->sc_mtx);
3009 g_topology_unlock();
3013 mtx_assert(&sc->sc_mtx, MA_OWNED);
3014 g_journal_switch_wait(sc);
3015 mtx_unlock(&sc->sc_mtx);
3020 * TODO: Switcher thread should be started on first geom creation and killed on
3021 * last geom destruction.
3024 g_journal_switcher(void *arg)
3031 curthread->td_pflags |= TDP_NORUNNINGBUF;
3033 g_journal_switcher_wokenup = 0;
3034 error = tsleep(&g_journal_switcher_state, PRIBIO, "jsw:wait",
3035 g_journal_switch_time * hz);
3036 if (g_journal_switcher_state == GJ_SWITCHER_DIE) {
3037 g_journal_switcher_state = GJ_SWITCHER_DIED;
3038 GJ_DEBUG(1, "Switcher exiting.");
3039 wakeup(&g_journal_switcher_state);
3042 if (error == 0 && g_journal_sync_requested == 0) {
3043 GJ_DEBUG(1, "Out of cache, force switch (used=%u "
3044 "limit=%u).", g_journal_cache_used,
3045 g_journal_cache_limit);
3047 GJ_TIMER_START(1, &bt);
3048 g_journal_do_switch(mp);
3049 GJ_TIMER_STOP(1, &bt, "Entire switch time");
3050 if (g_journal_sync_requested > 0) {
3051 g_journal_sync_requested = 0;
3052 wakeup(&g_journal_sync_requested);