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 "
345 "(id = %u joffset=%jd active=%jd inactive=%jd)",
347 (intmax_t)sc->sc_journal_offset,
348 (intmax_t)sc->sc_active.jj_offset,
349 (intmax_t)sc->sc_inactive.jj_offset);
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;
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;
360 used = sc->sc_jend - sc->sc_active.jj_offset;
361 used += sc->sc_journal_offset - sc->sc_jstart;
364 /* Already woken up? */
365 if (g_journal_switcher_wokenup)
368 * If the active journal takes more than g_journal_force_switch precent
369 * of free journal space, we force journal switch.
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);
389 g_journal_orphan(struct g_consumer *cp)
391 struct g_journal_softc *sc;
396 sc = cp->geom->softc;
397 strlcpy(name, cp->provider->name, sizeof(name));
398 GJ_DEBUG(0, "Lost provider %s.", name);
401 error = g_journal_destroy(sc);
403 GJ_DEBUG(0, "Journal %s destroyed.", name);
405 GJ_DEBUG(0, "Cannot destroy journal %s (error=%d). "
406 "Destroy it manually after last close.", sc->sc_name,
412 g_journal_access(struct g_provider *pp, int acr, int acw, int ace)
414 struct g_journal_softc *sc;
418 GJ_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name,
425 sc = pp->geom->softc;
426 if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY)) {
427 if (acr <= 0 && acw <= 0 && ace <= 0)
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;
436 g_journal_metadata_update(sc);
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;
442 g_journal_metadata_update(sc);
449 g_journal_header_encode(struct g_journal_header *hdr, u_char *data)
452 bcopy(GJ_HEADER_MAGIC, data, sizeof(GJ_HEADER_MAGIC));
453 data += sizeof(GJ_HEADER_MAGIC);
454 le32enc(data, hdr->jh_journal_id);
456 le32enc(data, hdr->jh_journal_next_id);
460 g_journal_header_decode(const u_char *data, struct g_journal_header *hdr)
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)
467 hdr->jh_journal_id = le32dec(data);
469 hdr->jh_journal_next_id = le32dec(data);
474 g_journal_flush_cache(struct g_journal_softc *sc)
479 if (sc->sc_bio_flush == 0)
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);
487 if (sc->sc_bio_flush & GJ_FLUSH_DATA) {
489 * TODO: This could be called in parallel with the
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);
496 GJ_TIMER_STOP(1, &bt, "Cache flush time");
500 g_journal_write_header(struct g_journal_softc *sc)
502 struct g_journal_header hdr;
503 struct g_consumer *cp;
507 cp = sc->sc_jconsumer;
508 buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
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;
519 gj_free(buf, cp->provider->sectorsize);
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 endianess.
529 g_journal_record_header_encode(struct g_journal_record_header *hdr,
532 struct g_journal_entry *ent;
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);
539 le16enc(data, hdr->jrh_nentries);
541 bcopy(hdr->jrh_sum, data, sizeof(hdr->jrh_sum));
543 for (i = 0; i < hdr->jrh_nentries; i++) {
544 ent = &hdr->jrh_entries[i];
545 le64enc(data, ent->je_joffset);
547 le64enc(data, ent->je_offset);
549 le64enc(data, ent->je_length);
555 g_journal_record_header_decode(const u_char *data,
556 struct g_journal_record_header *hdr)
558 struct g_journal_entry *ent;
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)
565 hdr->jrh_journal_id = le32dec(data);
567 hdr->jrh_nentries = le16dec(data);
569 if (hdr->jrh_nentries > GJ_RECORD_HEADER_NENTRIES)
571 bcopy(data, hdr->jrh_sum, sizeof(hdr->jrh_sum));
573 for (i = 0; i < hdr->jrh_nentries; i++) {
574 ent = &hdr->jrh_entries[i];
575 ent->je_joffset = le64dec(data);
577 ent->je_offset = le64dec(data);
579 ent->je_length = le64dec(data);
586 * Function reads metadata from a provider (via the given consumer), decodes
587 * it to system endianess and verifies its correctness.
590 g_journal_metadata_read(struct g_consumer *cp, struct g_journal_metadata *md)
592 struct g_provider *pp;
598 error = g_access(cp, 1, 0, 0);
603 /* Metadata is stored in last sector. */
604 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
607 g_access(cp, -1, 0, 0);
609 GJ_DEBUG(1, "Cannot read metadata from %s (error=%d).",
610 cp->provider->name, error);
614 /* Decode metadata. */
615 error = journal_metadata_decode(buf, md);
617 /* Is this is gjournal provider at all? */
618 if (strcmp(md->md_magic, G_JOURNAL_MAGIC) != 0)
621 * Are we able to handle this version of metadata?
622 * We only maintain backward compatibility.
624 if (md->md_version > G_JOURNAL_VERSION) {
626 "Kernel module is too old to handle metadata from %s.",
630 /* Is checksum correct? */
632 GJ_DEBUG(0, "MD5 metadata hash mismatch for provider %s.",
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).
645 g_journal_metadata_done(struct bio *bp)
649 * There is not much we can do on error except informing about it.
651 if (bp->bio_error != 0) {
652 GJ_LOGREQ(0, bp, "Cannot update metadata (error=%d).",
655 GJ_LOGREQ(2, bp, "Metadata updated.");
657 gj_free(bp->bio_data, bp->bio_length);
662 g_journal_metadata_update(struct g_journal_softc *sc)
664 struct g_journal_metadata md;
665 struct g_consumer *cp;
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;
680 if (sc->sc_flags & GJF_DEVICE_CLEAN)
681 md.md_flags |= GJ_FLAG_CLEAN;
683 if (sc->sc_flags & GJF_DEVICE_HARDCODED)
684 strlcpy(md.md_provider, sc->sc_name, sizeof(md.md_provider));
686 bzero(md.md_provider, sizeof(md.md_provider));
687 md.md_provsize = cp->provider->mediasize;
688 journal_metadata_encode(&md, sector);
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.
696 g_journal_flush_cache(sc);
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);
708 g_io_request(bp, cp);
709 biowait(bp, "gjmdu");
710 g_journal_metadata_done(bp);
714 * Be sure metadata reached the disk.
716 g_journal_flush_cache(sc);
720 * This is where the I/O request comes from the GEOM.
723 g_journal_start(struct bio *bp)
725 struct g_journal_softc *sc;
727 sc = bp->bio_to->geom->softc;
728 GJ_LOGREQ(3, bp, "Request received.");
730 switch (bp->bio_cmd) {
733 mtx_lock(&sc->sc_mtx);
734 bioq_insert_tail(&sc->sc_regular_queue, bp);
736 mtx_unlock(&sc->sc_mtx);
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;
748 g_io_deliver(bp, EOPNOTSUPP);
754 g_journal_std_done(struct bio *bp)
756 struct g_journal_softc *sc;
758 sc = bp->bio_from->geom->softc;
759 mtx_lock(&sc->sc_mtx);
760 bioq_insert_tail(&sc->sc_back_queue, bp);
762 mtx_unlock(&sc->sc_mtx);
766 g_journal_new_bio(off_t start, off_t end, off_t joffset, u_char *data,
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;
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);
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)
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).
796 * The function returns the number of bios inserted (as bio can be splitted).
799 g_journal_insert(struct bio **head, off_t nstart, off_t nend, off_t joffset,
800 u_char *data, int flags)
802 struct bio *nbp, *cbp, *pbp;
807 GJ_DEBUG(3, "INSERT(%p): (%jd, %jd, %jd)", *head, nstart, nend,
811 GJQ_FOREACH(*head, cbp) {
812 cstart = cbp->bio_offset;
813 cend = cbp->bio_offset + cbp->bio_length;
815 if (nstart >= cend) {
819 * | current | +-------------+
822 * +-------------+ | bio |
826 GJ_DEBUG(3, "INSERT(%p): 1", *head);
827 } else if (nend <= cstart) {
831 * +-------------+ | current |
834 * | bio | +-------------+
838 nbp = g_journal_new_bio(nstart, nend, joffset, data,
846 GJ_DEBUG(3, "INSERT(%p): 2 (nbp=%p pbp=%p)", *head, nbp,
849 } else if (nstart <= cstart && nend >= cend) {
851 * +-------------+ +-------------+
852 * | current bio | | current bio |
853 * +---+-------------+---+ +-------------+---+
856 * | +-------------+ | +-------------+ |
857 * | new bio | | new bio |
858 * +---------------------+ +-----------------+
860 * +-------------+ +-------------+
861 * | current bio | | current bio |
862 * +---+-------------+ +-------------+
865 * | +-------------+ +-------------+
866 * | new bio | | new bio |
867 * +-----------------+ +-------------+
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;
878 cbp->bio_data = gj_malloc(cbp->bio_length,
880 if (cbp->bio_data != NULL) {
881 bcopy(data, cbp->bio_data,
884 data += cend - nstart;
886 joffset += cend - nstart;
888 GJ_DEBUG(3, "INSERT(%p): 3 (cbp=%p)", *head, cbp);
889 } else if (nstart > cstart && nend >= cend) {
891 * +-----------------+ +-------------+
892 * | current bio | | current bio |
893 * | +-------------+ | +---------+---+
896 * +---+-------------+ +---+---------+ |
897 * | new bio | | new bio |
898 * +-------------+ +-------------+
900 g_journal_stats_bytes_skipped += cend - nstart;
901 nbp = g_journal_new_bio(nstart, cend, joffset, data,
903 nbp->bio_next = cbp->bio_next;
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);
911 data += cend - nstart;
912 joffset += cend - nstart;
915 GJ_DEBUG(3, "INSERT(%p): 4 (cbp=%p)", *head, cbp);
916 } else if (nstart > cstart && nend < cend) {
918 * +---------------------+
920 * | +-------------+ |
923 * +---+-------------+---+
927 g_journal_stats_bytes_skipped += nend - nstart;
928 nbp = g_journal_new_bio(nstart, nend, joffset, data,
930 nbp->bio_next = cbp->bio_next;
932 if (cbp->bio_data == NULL)
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);
946 GJ_DEBUG(3, "INSERT(%p): 5 (cbp=%p)", *head, cbp);
948 } else if (nstart <= cstart && nend < cend) {
950 * +-----------------+ +-------------+
951 * | current bio | | current bio |
952 * +-------------+ | +---+---------+ |
955 * +-------------+---+ | +---------+---+
956 * | new bio | | new bio |
957 * +-------------+ +-------------+
959 g_journal_stats_bytes_skipped += nend - nstart;
960 nbp = g_journal_new_bio(nstart, nend, joffset, data,
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,
974 if (cbp->bio_data != NULL) {
975 bcopy(tmpdata + nend - cstart,
976 cbp->bio_data, cbp->bio_length);
978 gj_free(tmpdata, cend - cstart);
981 GJ_DEBUG(3, "INSERT(%p): 6 (cbp=%p)", *head, cbp);
988 nbp = g_journal_new_bio(nstart, nend, joffset, data, flags);
993 nbp->bio_next = NULL;
995 GJ_DEBUG(3, "INSERT(%p): 8 (nbp=%p pbp=%p)", *head, nbp, pbp);
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);
1004 GJ_DEBUG(3, "INSERT(%p): DONE %d", *head, n);
1010 * The function combines neighbour bios trying to squeeze as much data as
1011 * possible into one bio.
1013 * The function returns the number of bios combined (negative value).
1016 g_journal_optimize(struct bio *head)
1018 struct bio *cbp, *pbp;
1023 GJQ_FOREACH(head, cbp) {
1024 /* Skip bios which has to be read first. */
1025 if (cbp->bio_data == NULL) {
1029 /* There is no previous bio yet. */
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));
1045 /* Be sure we don't end up with too big bio. */
1046 if (pbp->bio_length + cbp->bio_length > MAXPHYS) {
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,
1057 gj_free(cbp->bio_data, cbp->bio_length);
1058 pbp->bio_length += cbp->bio_length;
1059 pbp->bio_next = cbp->bio_next;
1062 g_journal_stats_combined_ios++;
1064 GJ_LOGREQ(4, pbp, "Got: ");
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
1076 * 2. Read the data blocks from the journal.
1077 * 3. Write the data blocks on the data provider.
1079 * g_journal_copy_start()
1080 * g_journal_copy_done() - got finished write request, logs potential errors.
1084 * When there is no data in cache, this function is used to read it.
1087 g_journal_read_first(struct g_journal_softc *sc, struct bio *bp)
1092 * We were short in memory, so data was freed.
1093 * In that case we need to read it back from journal.
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++;
1109 g_journal_copy_send(struct g_journal_softc *sc)
1111 struct bio *bioq, *bp, *lbp;
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);
1119 GJQ_REMOVE(sc->sc_inactive.jj_queue, bp);
1120 sc->sc_copy_in_progress++;
1121 GJQ_INSERT_AFTER(bioq, bp, lbp);
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);
1134 bp->bio_joffset = 0;
1135 GJ_LOGREQ(4, bp, "SEND");
1136 g_io_request(bp, sc->sc_dconsumer);
1142 g_journal_copy_start(struct g_journal_softc *sc)
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.
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);
1155 g_journal_copy_send(sc);
1159 * Data block has been read from the journal provider.
1162 g_journal_copy_read_done(struct bio *bp)
1164 struct g_journal_softc *sc;
1165 struct g_consumer *cp;
1168 KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1169 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1171 sc = bp->bio_from->geom->softc;
1172 pbp = bp->bio_parent;
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);
1178 * We will not be able to deliver WRITE request as well.
1180 gj_free(bp->bio_data, bp->bio_length);
1183 sc->sc_copy_in_progress--;
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");
1195 * Data block has been written to the data provider.
1198 g_journal_copy_write_done(struct bio *bp)
1200 struct g_journal_softc *sc;
1202 KASSERT(bp->bio_cflags == GJ_BIO_COPY,
1203 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY));
1205 sc = bp->bio_from->geom->softc;
1206 sc->sc_copy_in_progress--;
1208 if (bp->bio_error != 0) {
1209 GJ_LOGREQ(0, bp, "[copy] Error while writing data (error=%d)",
1212 GJQ_REMOVE(sc->sc_copy_queue, bp);
1213 gj_free(bp->bio_data, bp->bio_length);
1214 GJ_LOGREQ(4, bp, "DONE");
1217 if (sc->sc_copy_in_progress == 0) {
1219 * This was the last write request for this journal.
1221 GJ_DEBUG(1, "Data has been copied.");
1222 sc->sc_journal_copying = 0;
1226 static void g_journal_flush_done(struct bio *bp);
1229 * Flush one record onto active journal provider.
1232 g_journal_flush(struct g_journal_softc *sc)
1234 struct g_journal_record_header hdr;
1235 struct g_journal_entry *ent;
1236 struct g_provider *pp;
1238 struct bio *bp, *fbp, *pbp;
1239 off_t joffset, size;
1240 u_char *data, hash[16];
1244 if (sc->sc_current_count == 0)
1248 pp = sc->sc_jprovider;
1249 GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1250 joffset = sc->sc_journal_offset;
1252 GJ_DEBUG(2, "Storing %d journal entries on %s at %jd.",
1253 sc->sc_current_count, pp->name, (intmax_t)joffset);
1256 * Store 'journal id', so we know to which journal this record belongs.
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));
1263 bioq = &sc->sc_active.jj_queue;
1264 pbp = sc->sc_flush_queue;
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);
1277 GJ_LOGREQ(4, fbp, "FLUSH_OUT");
1278 joffset += pp->sectorsize;
1279 sc->sc_flush_count++;
1280 if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1283 for (i = 0; i < hdr.jrh_nentries; i++) {
1284 bp = sc->sc_current_queue;
1285 KASSERT(bp != NULL, ("NULL bp"));
1287 GJ_LOGREQ(4, bp, "FLUSHED");
1288 sc->sc_current_queue = bp->bio_next;
1289 bp->bio_next = NULL;
1290 sc->sc_current_count--;
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;
1299 data = bp->bio_data;
1300 if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
1301 MD5Update(&ctx, data, ent->je_length);
1302 bzero(bp, sizeof(*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);
1313 GJ_LOGREQ(4, bp, "FLUSH_OUT");
1314 joffset += bp->bio_length;
1315 sc->sc_flush_count++;
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
1323 g_journal_insert(bioq, ent->je_offset,
1324 ent->je_offset + ent->je_length, ent->je_joffset, data,
1329 * After all requests, store valid header.
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));
1336 g_journal_record_header_encode(&hdr, data);
1337 fbp->bio_data = data;
1339 sc->sc_journal_offset = joffset;
1341 g_journal_check_overflow(sc);
1345 * Flush request finished.
1348 g_journal_flush_done(struct bio *bp)
1350 struct g_journal_softc *sc;
1351 struct g_consumer *cp;
1353 KASSERT((bp->bio_cflags & GJ_BIO_MASK) == GJ_BIO_JOURNAL,
1354 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_JOURNAL));
1357 sc = cp->geom->softc;
1358 sc->sc_flush_in_progress--;
1360 if (bp->bio_error != 0) {
1361 GJ_LOGREQ(0, bp, "[flush] Error while writing data (error=%d)",
1364 gj_free(bp->bio_data, bp->bio_length);
1365 GJ_LOGREQ(4, bp, "DONE");
1369 static void g_journal_release_delayed(struct g_journal_softc *sc);
1372 g_journal_flush_send(struct g_journal_softc *sc)
1374 struct g_consumer *cp;
1375 struct bio *bioq, *bp, *lbp;
1377 cp = sc->sc_jconsumer;
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);
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);
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)
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);
1407 g_journal_add_current(struct g_journal_softc *sc, struct bio *bp)
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;
1417 * For requests which are added to the current queue we deliver
1418 * response immediately.
1420 bp->bio_completed = bp->bio_length;
1421 g_io_deliver(bp, 0);
1422 if (sc->sc_current_count >= g_journal_record_entries) {
1424 * Let's flush one record onto active journal provider.
1426 g_journal_flush(sc);
1431 g_journal_release_delayed(struct g_journal_softc *sc)
1436 /* The flush queue is full, exit. */
1437 if (sc->sc_flush_count >= g_journal_accept_immediately)
1439 bp = bioq_takefirst(&sc->sc_delayed_queue);
1442 sc->sc_delayed_count--;
1443 g_journal_add_current(sc, bp);
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.
1452 g_journal_add_request(struct g_journal_softc *sc, struct bio *bp)
1456 * The flush queue is full, we need to delay the request.
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++;
1466 KASSERT(TAILQ_EMPTY(&sc->sc_delayed_queue.queue),
1467 ("DELAYED queue not empty."));
1468 g_journal_add_current(sc, bp);
1471 static void g_journal_read_done(struct bio *bp);
1474 * Try to find requested data in cache.
1477 g_journal_read_find(struct bio *head, int sorted, struct bio *pbp, off_t ostart,
1483 GJQ_FOREACH(head, bp) {
1484 if (bp->bio_offset == -1)
1486 cstart = MAX(ostart, bp->bio_offset);
1487 cend = MIN(oend, bp->bio_offset + bp->bio_length);
1490 else if (cstart >= oend) {
1498 if (bp->bio_data == NULL)
1500 GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend,
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) {
1507 * Cool, the whole request was in cache, deliver happy
1510 g_io_deliver(pbp, 0);
1519 * Try to find requested data in cache.
1522 g_journal_read_queue_find(struct bio_queue *head, struct bio *pbp, off_t ostart,
1528 TAILQ_FOREACH(bp, head, bio_queue) {
1529 cstart = MAX(ostart, bp->bio_offset);
1530 cend = MIN(oend, bp->bio_offset + bp->bio_length);
1533 else if (cstart >= oend)
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,
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) {
1544 * Cool, the whole request was in cache, deliver happy
1547 g_io_deliver(pbp, 0);
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
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
1567 g_journal_read(struct g_journal_softc *sc, struct bio *pbp, off_t ostart,
1570 struct bio *bp, *nbp, *head;
1572 u_int i, sorted = 0;
1574 GJ_DEBUG(3, "READ: (%jd, %jd)", ostart, oend);
1579 for (i = 0; i <= 5; i++) {
1581 case 0: /* Delayed requests. */
1585 case 1: /* Not-yet-send data. */
1586 head = sc->sc_current_queue;
1589 case 2: /* In-flight to the active journal. */
1590 head = sc->sc_flush_queue;
1593 case 3: /* Active journal. */
1594 head = sc->sc_active.jj_queue;
1597 case 4: /* Inactive journal. */
1599 * XXX: Here could be a race with g_journal_lowmem().
1601 head = sc->sc_inactive.jj_queue;
1604 case 5: /* In-flight to the data provider. */
1605 head = sc->sc_copy_queue;
1609 panic("gjournal %s: i=%d", __func__, i);
1612 bp = g_journal_read_queue_find(&sc->sc_delayed_queue.queue, pbp, ostart, oend);
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);
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);
1627 if (bp->bio_data == NULL) {
1628 nbp = g_duplicate_bio(pbp);
1629 nbp->bio_cflags = GJ_BIO_READ;
1631 pbp->bio_data + cstart - pbp->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);
1639 * If we don't have the whole request yet, call g_journal_read()
1642 if (ostart < cstart)
1643 g_journal_read(sc, pbp, ostart, cstart);
1645 g_journal_read(sc, pbp, cend, oend);
1648 * No data in memory, no data in journal.
1649 * Its time for asking data provider.
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. */
1665 * Function responsible for handling finished READ requests.
1666 * Actually, g_std_done() could be used here, the only difference is that we
1670 g_journal_read_done(struct bio *bp)
1674 KASSERT(bp->bio_cflags == GJ_BIO_READ,
1675 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_READ));
1677 pbp = bp->bio_parent;
1679 pbp->bio_completed += bp->bio_length;
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);
1688 if (pbp->bio_children == pbp->bio_inbed &&
1689 pbp->bio_completed == pbp->bio_length) {
1691 g_io_deliver(pbp, 0);
1696 * Deactive current journal and active next one.
1699 g_journal_switch(struct g_journal_softc *sc)
1701 struct g_provider *pp;
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);
1712 GJ_DEBUG(3, "Switching journal %s.", sc->sc_geom->name);
1714 pp = sc->sc_jprovider;
1716 sc->sc_journal_previous_id = sc->sc_journal_id;
1718 sc->sc_journal_id = sc->sc_journal_next_id;
1719 sc->sc_journal_next_id = arc4random();
1721 GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc);
1723 g_journal_write_header(sc);
1725 sc->sc_inactive.jj_offset = sc->sc_active.jj_offset;
1726 sc->sc_inactive.jj_queue = sc->sc_active.jj_queue;
1728 sc->sc_active.jj_offset =
1729 sc->sc_journal_offset - pp->sectorsize;
1730 sc->sc_active.jj_queue = NULL;
1733 * Switch is done, start copying data from the (now) inactive
1734 * journal to the data provider.
1736 g_journal_copy_start(sc);
1738 mtx_lock(&sc->sc_mtx);
1739 sc->sc_flags &= ~GJF_DEVICE_SWITCH;
1740 mtx_unlock(&sc->sc_mtx);
1744 g_journal_initialize(struct g_journal_softc *sc)
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;
1757 g_journal_mark_as_dirty(struct g_journal_softc *sc)
1759 const struct g_journal_desc *desc;
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);
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.
1773 g_journal_sync_read(struct g_consumer *cp, struct bio *bp, off_t offset,
1778 bzero(bp, sizeof(*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");
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.
1803 g_journal_sync(struct g_journal_softc *sc)
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];
1814 int error, found, i;
1818 cp = sc->sc_jconsumer;
1820 buf = gj_malloc(cp->provider->sectorsize, M_WAITOK);
1821 offset = joffset = sc->sc_inactive.jj_offset = sc->sc_journal_offset;
1823 GJ_DEBUG(2, "Looking for termination at %jd.", (intmax_t)joffset);
1826 * Read and decode first journal header.
1828 error = g_journal_sync_read(cp, bp, offset, buf);
1830 GJ_DEBUG(0, "Error while reading journal header from %s.",
1831 cp->provider->name);
1834 error = g_journal_header_decode(buf, &jhdr);
1836 GJ_DEBUG(0, "Cannot decode journal header from %s.",
1837 cp->provider->name);
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);
1846 offset += cp->provider->sectorsize;
1847 id = sc->sc_journal_next_id = jhdr.jh_journal_next_id;
1851 * If the biggest record won't fit, look for a record header or
1852 * journal header from the begining.
1854 GJ_VALIDATE_OFFSET(offset, sc);
1855 error = g_journal_sync_read(cp, bp, offset, buf);
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.
1863 "Error while reading record header from %s.",
1864 cp->provider->name);
1868 error = g_journal_record_header_decode(buf, &rhdr);
1870 GJ_DEBUG(2, "Not a record header at %jd (error=%d).",
1871 (intmax_t)offset, error);
1873 * This is not a record header.
1874 * If we are lucky, this is next journal header.
1876 error = g_journal_header_decode(buf, &jhdr);
1878 GJ_DEBUG(1, "Not a journal header at %jd (error=%d).",
1879 (intmax_t)offset, error);
1881 * Nope, this is not journal header, which
1882 * bascially means that journal is not
1883 * terminated properly.
1889 * Ok. This is header of _some_ journal. Now we need to
1890 * verify if this is header of the _next_ journal.
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);
1900 /* Found termination. */
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;
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,
1917 /* Skip journal's header. */
1918 offset += cp->provider->sectorsize;
1922 /* Skip record's header. */
1923 offset += cp->provider->sectorsize;
1926 * Add information about every record entry to the inactive
1929 if (sc->sc_flags & GJF_DEVICE_CHECKSUM)
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,
1938 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) {
1942 * TODO: Should use faster function (like
1943 * g_journal_sync_read()).
1945 buf2 = g_read_data(cp, offset, ent->je_length,
1948 GJ_DEBUG(0, "Cannot read data at %jd.",
1951 MD5Update(&ctx, buf2, ent->je_length);
1955 /* Skip entry's data. */
1956 offset += ent->je_length;
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!",
1967 gj_free(bp->bio_data, cp->provider->sectorsize);
1970 /* Remove bios from unterminated journal. */
1971 while ((tbp = fbp) != NULL) {
1972 fbp = tbp->bio_next;
1976 if (found < 1 && joffset > 0) {
1977 GJ_DEBUG(0, "Journal on %s is broken/corrupted. Initializing.",
1979 while ((tbp = sc->sc_inactive.jj_queue) != NULL) {
1980 sc->sc_inactive.jj_queue = tbp->bio_next;
1983 g_journal_initialize(sc);
1984 g_journal_mark_as_dirty(sc);
1986 GJ_DEBUG(0, "Journal %s consistent.", sc->sc_name);
1987 g_journal_copy_start(sc);
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.
1998 g_journal_wait(struct g_journal_softc *sc, time_t last_write)
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);
2008 * If we have debug turned on, show number of elements
2009 * in various queues.
2012 error = msleep(sc, &sc->sc_mtx, PRIBIO,
2015 mtx_unlock(&sc->sc_mtx);
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);
2030 GJ_DEBUG(3, "%s: exit 1", __func__);
2035 * Flush even not full records every 3 seconds.
2037 timeout = (last_write + 3 - time_second) * hz;
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__);
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__);
2055 g_journal_worker(void *arg)
2057 struct g_journal_softc *sc;
2059 struct g_provider *pp;
2064 thread_lock(curthread);
2065 sched_prio(curthread, PRIBIO);
2066 thread_unlock(curthread);
2071 if (sc->sc_flags & GJF_DEVICE_CLEAN) {
2072 GJ_DEBUG(0, "Journal %s clean.", sc->sc_name);
2073 g_journal_initialize(sc);
2078 * Check if we can use BIO_FLUSH.
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);
2086 GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2087 sc->sc_jconsumer->provider->name);
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);
2095 GJ_DEBUG(0, "BIO_FLUSH not supported by %s.",
2096 sc->sc_dconsumer->provider->name);
2102 pp = g_new_providerf(gp, "%s.journal", sc->sc_name);
2103 pp->mediasize = sc->sc_mediasize;
2105 * There could be a problem when data provider and journal providers
2106 * have different sectorsize, but such scenario is prevented on journal
2109 pp->sectorsize = sc->sc_sectorsize;
2110 g_error_provider(pp, 0);
2111 g_topology_unlock();
2112 last_write = time_second;
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;
2121 /* Get first request from the queue. */
2122 mtx_lock(&sc->sc_mtx);
2123 bp = bioq_first(&sc->sc_back_queue);
2125 type = (bp->bio_cflags & GJ_BIO_MASK);
2127 bp = bioq_first(&sc->sc_regular_queue);
2129 type = GJ_BIO_REGULAR;
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);
2141 if (sc->sc_flush_in_progress > 0)
2143 if (sc->sc_copy_in_progress > 0)
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);
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);
2161 g_journal_wait(sc, last_write);
2165 * If we're in switch process, we need to delay all new
2166 * write requests until its done.
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");
2173 if (type == GJ_BIO_REGULAR)
2174 bioq_remove(&sc->sc_regular_queue, bp);
2176 bioq_remove(&sc->sc_back_queue, bp);
2177 mtx_unlock(&sc->sc_mtx);
2179 case GJ_BIO_REGULAR:
2180 /* Regular request. */
2181 switch (bp->bio_cmd) {
2183 g_journal_read(sc, bp, bp->bio_offset,
2184 bp->bio_offset + bp->bio_length);
2187 last_write = time_second;
2188 g_journal_add_request(sc, bp);
2189 g_journal_flush_send(sc);
2192 panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2196 switch (bp->bio_cmd) {
2198 if (g_journal_copy_read_done(bp))
2199 g_journal_copy_send(sc);
2202 g_journal_copy_write_done(bp);
2203 g_journal_copy_send(sc);
2206 panic("Invalid bio_cmd (%d).", bp->bio_cmd);
2209 case GJ_BIO_JOURNAL:
2210 g_journal_flush_done(bp);
2211 g_journal_flush_send(sc);
2215 panic("Invalid bio (%d).", type);
2221 g_journal_destroy_event(void *arg, int flags __unused)
2223 struct g_journal_softc *sc;
2225 g_topology_assert();
2227 g_journal_destroy(sc);
2231 g_journal_timeout(void *arg)
2233 struct g_journal_softc *sc;
2236 GJ_DEBUG(0, "Timeout. Journal %s cannot be completed.",
2238 g_post_event(g_journal_destroy_event, sc, M_NOWAIT, NULL);
2241 static struct g_geom *
2242 g_journal_create(struct g_class *mp, struct g_provider *pp,
2243 const struct g_journal_metadata *md)
2245 struct g_journal_softc *sc;
2247 struct g_consumer *cp;
2250 sc = NULL; /* gcc */
2252 g_topology_assert();
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.
2258 /* Look for journal device with the same ID. */
2259 LIST_FOREACH(gp, &mp->geom, geom) {
2263 if (sc->sc_id == md->md_id)
2268 else if (sc != NULL && (sc->sc_type & md->md_type) != 0) {
2269 GJ_DEBUG(1, "Journal device %u already configured.", sc->sc_id);
2272 if (md->md_type == 0 || (md->md_type & ~GJ_TYPE_COMPLETE) != 0) {
2273 GJ_DEBUG(0, "Invalid type on %s.", pp->name);
2276 if (md->md_type & GJ_TYPE_DATA) {
2277 GJ_DEBUG(0, "Journal %u: %s contains data.", md->md_id,
2280 if (md->md_type & GJ_TYPE_JOURNAL) {
2281 GJ_DEBUG(0, "Journal %u: %s contains journal.", md->md_id,
2287 sc = malloc(sizeof(*sc), M_JOURNAL, M_WAITOK | M_ZERO);
2288 sc->sc_id = md->md_id;
2291 sc->sc_worker = NULL;
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;
2298 gp->flags |= G_GEOM_VOLATILE_BIO;
2301 mtx_init(&sc->sc_mtx, "gjournal", NULL, MTX_DEF);
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;
2317 sc->sc_rootmount = root_mount_hold("GJOURNAL");
2318 GJ_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2320 callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
2321 if (md->md_type != GJ_TYPE_COMPLETE) {
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.
2328 callout_reset(&sc->sc_callout, 5 * hz,
2329 g_journal_timeout, sc);
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;
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).",
2348 error = g_access(cp, 1, 1, 1);
2350 GJ_DEBUG(0, "Cannot access %s (error=%d).", pp->name,
2352 g_journal_destroy(sc);
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;
2366 if (md->md_type & GJ_TYPE_JOURNAL) {
2368 cp = g_new_consumer(gp);
2369 error = g_attach(cp, pp);
2370 KASSERT(error == 0, ("Cannot attach to %s (error=%d).",
2372 error = g_access(cp, 1, 1, 1);
2374 GJ_DEBUG(0, "Cannot access %s (error=%d).",
2376 g_journal_destroy(sc);
2381 * Journal is on the same provider as data, which means
2382 * that data provider ends where journal starts.
2384 sc->sc_mediasize = md->md_jstart;
2386 sc->sc_jconsumer = cp;
2389 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE) {
2390 /* Journal is not complete yet. */
2393 /* Journal complete, cancel timeout. */
2394 callout_drain(&sc->sc_callout);
2397 error = kproc_create(g_journal_worker, sc, &sc->sc_worker, 0, 0,
2398 "g_journal %s", sc->sc_name);
2400 GJ_DEBUG(0, "Cannot create worker thread for %s.journal.",
2402 g_journal_destroy(sc);
2410 g_journal_destroy_consumer(void *arg, int flags __unused)
2412 struct g_consumer *cp;
2414 g_topology_assert();
2417 g_destroy_consumer(cp);
2421 g_journal_destroy(struct g_journal_softc *sc)
2424 struct g_provider *pp;
2425 struct g_consumer *cp;
2427 g_topology_assert();
2433 pp = LIST_FIRST(&gp->provider);
2435 if (pp->acr != 0 || pp->acw != 0 || pp->ace != 0) {
2436 GJ_DEBUG(1, "Device %s is still open (r%dw%de%d).",
2437 pp->name, pp->acr, pp->acw, pp->ace);
2440 g_error_provider(pp, ENXIO);
2442 g_journal_flush(sc);
2443 g_journal_flush_send(sc);
2444 g_journal_switch(sc);
2447 sc->sc_flags |= (GJF_DEVICE_DESTROY | GJF_DEVICE_CLEAN);
2449 g_topology_unlock();
2451 if (sc->sc_rootmount != NULL) {
2452 GJ_DEBUG(1, "root_mount_rel %p", sc->sc_rootmount);
2453 root_mount_rel(sc->sc_rootmount);
2454 sc->sc_rootmount = NULL;
2457 callout_drain(&sc->sc_callout);
2458 mtx_lock(&sc->sc_mtx);
2460 while (sc->sc_worker != NULL)
2461 msleep(&sc->sc_worker, &sc->sc_mtx, PRIBIO, "gj:destroy", 0);
2462 mtx_unlock(&sc->sc_mtx);
2465 GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name);
2466 g_journal_metadata_update(sc);
2468 pp->flags |= G_PF_WITHER;
2469 g_orphan_provider(pp, ENXIO);
2473 mtx_destroy(&sc->sc_mtx);
2475 if (sc->sc_current_count != 0) {
2476 GJ_DEBUG(0, "Warning! Number of current requests %d.",
2477 sc->sc_current_count);
2480 LIST_FOREACH(cp, &gp->consumer, consumer) {
2481 if (cp->acr + cp->acw + cp->ace > 0)
2482 g_access(cp, -1, -1, -1);
2484 * We keep all consumers open for writting, so if I'll detach
2485 * and destroy consumer here, I'll get providers for taste, so
2486 * journal will be started again.
2487 * Sending an event here, prevents this from happening.
2489 g_post_event(g_journal_destroy_consumer, cp, M_WAITOK, NULL);
2492 g_wither_geom(gp, ENXIO);
2493 free(sc, M_JOURNAL);
2498 g_journal_taste_orphan(struct g_consumer *cp)
2501 KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
2502 cp->provider->name));
2505 static struct g_geom *
2506 g_journal_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
2508 struct g_journal_metadata md;
2509 struct g_consumer *cp;
2513 g_topology_assert();
2514 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
2515 GJ_DEBUG(2, "Tasting %s.", pp->name);
2516 if (pp->geom->class == mp)
2519 gp = g_new_geomf(mp, "journal:taste");
2520 /* This orphan function should be never called. */
2521 gp->orphan = g_journal_taste_orphan;
2522 cp = g_new_consumer(gp);
2524 error = g_journal_metadata_read(cp, &md);
2526 g_destroy_consumer(cp);
2532 if (md.md_provider[0] != '\0' &&
2533 !g_compare_names(md.md_provider, pp->name))
2535 if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
2537 if (g_journal_debug >= 2)
2538 journal_metadata_dump(&md);
2540 gp = g_journal_create(mp, pp, &md);
2544 static struct g_journal_softc *
2545 g_journal_find_device(struct g_class *mp, const char *name)
2547 struct g_journal_softc *sc;
2549 struct g_provider *pp;
2551 if (strncmp(name, "/dev/", 5) == 0)
2553 LIST_FOREACH(gp, &mp->geom, geom) {
2557 if (sc->sc_flags & GJF_DEVICE_DESTROY)
2559 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2561 pp = LIST_FIRST(&gp->provider);
2562 if (strcmp(sc->sc_name, name) == 0)
2564 if (pp != NULL && strcmp(pp->name, name) == 0)
2571 g_journal_ctl_destroy(struct gctl_req *req, struct g_class *mp)
2573 struct g_journal_softc *sc;
2579 g_topology_assert();
2581 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
2582 if (nargs == NULL) {
2583 gctl_error(req, "No '%s' argument.", "nargs");
2587 gctl_error(req, "Missing device(s).");
2591 for (i = 0; i < *nargs; i++) {
2592 snprintf(param, sizeof(param), "arg%d", i);
2593 name = gctl_get_asciiparam(req, param);
2595 gctl_error(req, "No 'arg%d' argument.", i);
2598 sc = g_journal_find_device(mp, name);
2600 gctl_error(req, "No such device: %s.", name);
2603 error = g_journal_destroy(sc);
2605 gctl_error(req, "Cannot destroy device %s (error=%d).",
2606 LIST_FIRST(&sc->sc_geom->provider)->name, error);
2613 g_journal_ctl_sync(struct gctl_req *req __unused, struct g_class *mp __unused)
2616 g_topology_assert();
2617 g_topology_unlock();
2618 g_journal_sync_requested++;
2619 wakeup(&g_journal_switcher_state);
2620 while (g_journal_sync_requested > 0)
2621 tsleep(&g_journal_sync_requested, PRIBIO, "j:sreq", hz / 2);
2626 g_journal_config(struct gctl_req *req, struct g_class *mp, const char *verb)
2630 g_topology_assert();
2632 version = gctl_get_paraml(req, "version", sizeof(*version));
2633 if (version == NULL) {
2634 gctl_error(req, "No '%s' argument.", "version");
2637 if (*version != G_JOURNAL_VERSION) {
2638 gctl_error(req, "Userland and kernel parts are out of sync.");
2642 if (strcmp(verb, "destroy") == 0 || strcmp(verb, "stop") == 0) {
2643 g_journal_ctl_destroy(req, mp);
2645 } else if (strcmp(verb, "sync") == 0) {
2646 g_journal_ctl_sync(req, mp);
2650 gctl_error(req, "Unknown verb.");
2654 g_journal_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
2655 struct g_consumer *cp, struct g_provider *pp)
2657 struct g_journal_softc *sc;
2659 g_topology_assert();
2666 } else if (cp != NULL) {
2669 sbuf_printf(sb, "%s<Role>", indent);
2670 if (cp == sc->sc_dconsumer) {
2671 sbuf_printf(sb, "Data");
2674 if (cp == sc->sc_jconsumer) {
2676 sbuf_printf(sb, ",");
2677 sbuf_printf(sb, "Journal");
2679 sbuf_printf(sb, "</Role>\n");
2680 if (cp == sc->sc_jconsumer) {
2681 sbuf_printf(sb, "<Jstart>%jd</Jstart>\n",
2682 (intmax_t)sc->sc_jstart);
2683 sbuf_printf(sb, "<Jend>%jd</Jend>\n",
2684 (intmax_t)sc->sc_jend);
2687 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
2691 static eventhandler_tag g_journal_event_shutdown = NULL;
2692 static eventhandler_tag g_journal_event_lowmem = NULL;
2695 g_journal_shutdown(void *arg, int howto __unused)
2698 struct g_geom *gp, *gp2;
2700 if (panicstr != NULL)
2705 LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
2706 if (gp->softc == NULL)
2708 GJ_DEBUG(0, "Shutting down geom %s.", gp->name);
2709 g_journal_destroy(gp->softc);
2711 g_topology_unlock();
2716 * Free cached requests from inactive queue in case of low memory.
2717 * We free GJ_FREE_AT_ONCE elements at once.
2719 #define GJ_FREE_AT_ONCE 4
2721 g_journal_lowmem(void *arg, int howto __unused)
2723 struct g_journal_softc *sc;
2727 u_int nfree = GJ_FREE_AT_ONCE;
2729 g_journal_stats_low_mem++;
2733 LIST_FOREACH(gp, &mp->geom, geom) {
2735 if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY))
2737 mtx_lock(&sc->sc_mtx);
2738 for (bp = sc->sc_inactive.jj_queue; nfree > 0 && bp != NULL;
2739 nfree--, bp = bp->bio_next) {
2741 * This is safe to free the bio_data, because:
2742 * 1. If bio_data is NULL it will be read from the
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.
2750 if (bp->bio_data != NULL) {
2751 GJ_DEBUG(2, "Freeing data from %s.",
2753 gj_free(bp->bio_data, bp->bio_length);
2754 bp->bio_data = NULL;
2757 mtx_unlock(&sc->sc_mtx);
2761 g_topology_unlock();
2765 static void g_journal_switcher(void *arg);
2768 g_journal_init(struct g_class *mp)
2772 /* Pick a conservative value if provided value sucks. */
2773 if (g_journal_cache_divisor <= 0 ||
2774 (vm_kmem_size / g_journal_cache_divisor == 0)) {
2775 g_journal_cache_divisor = 5;
2777 if (g_journal_cache_limit > 0) {
2778 g_journal_cache_limit = vm_kmem_size / g_journal_cache_divisor;
2779 g_journal_cache_low =
2780 (g_journal_cache_limit / 100) * g_journal_cache_switch;
2782 g_journal_event_shutdown = EVENTHANDLER_REGISTER(shutdown_post_sync,
2783 g_journal_shutdown, mp, EVENTHANDLER_PRI_FIRST);
2784 if (g_journal_event_shutdown == NULL)
2785 GJ_DEBUG(0, "Warning! Cannot register shutdown event.");
2786 g_journal_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem,
2787 g_journal_lowmem, mp, EVENTHANDLER_PRI_FIRST);
2788 if (g_journal_event_lowmem == NULL)
2789 GJ_DEBUG(0, "Warning! Cannot register lowmem event.");
2790 error = kproc_create(g_journal_switcher, mp, NULL, 0, 0,
2791 "g_journal switcher");
2792 KASSERT(error == 0, ("Cannot create switcher thread."));
2796 g_journal_fini(struct g_class *mp)
2799 if (g_journal_event_shutdown != NULL) {
2800 EVENTHANDLER_DEREGISTER(shutdown_post_sync,
2801 g_journal_event_shutdown);
2803 if (g_journal_event_lowmem != NULL)
2804 EVENTHANDLER_DEREGISTER(vm_lowmem, g_journal_event_lowmem);
2805 g_journal_switcher_state = GJ_SWITCHER_DIE;
2806 wakeup(&g_journal_switcher_state);
2807 while (g_journal_switcher_state != GJ_SWITCHER_DIED)
2808 tsleep(&g_journal_switcher_state, PRIBIO, "jfini:wait", hz / 5);
2809 GJ_DEBUG(1, "Switcher died.");
2812 DECLARE_GEOM_CLASS(g_journal_class, g_journal);
2814 static const struct g_journal_desc *
2815 g_journal_find_desc(const char *fstype)
2817 const struct g_journal_desc *desc;
2820 for (desc = g_journal_filesystems[i = 0]; desc != NULL;
2821 desc = g_journal_filesystems[++i]) {
2822 if (strcmp(desc->jd_fstype, fstype) == 0)
2829 g_journal_switch_wait(struct g_journal_softc *sc)
2833 mtx_assert(&sc->sc_mtx, MA_OWNED);
2834 if (g_journal_debug >= 2) {
2835 if (sc->sc_flush_in_progress > 0) {
2836 GJ_DEBUG(2, "%d requests flushing.",
2837 sc->sc_flush_in_progress);
2839 if (sc->sc_copy_in_progress > 0) {
2840 GJ_DEBUG(2, "%d requests copying.",
2841 sc->sc_copy_in_progress);
2843 if (sc->sc_flush_count > 0) {
2844 GJ_DEBUG(2, "%d requests to flush.",
2845 sc->sc_flush_count);
2847 if (sc->sc_delayed_count > 0) {
2848 GJ_DEBUG(2, "%d requests delayed.",
2849 sc->sc_delayed_count);
2852 g_journal_stats_switches++;
2853 if (sc->sc_copy_in_progress > 0)
2854 g_journal_stats_wait_for_copy++;
2855 GJ_TIMER_START(1, &bt);
2856 sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2857 sc->sc_flags |= GJF_DEVICE_SWITCH;
2859 while (sc->sc_flags & GJF_DEVICE_SWITCH) {
2860 msleep(&sc->sc_journal_copying, &sc->sc_mtx, PRIBIO,
2863 GJ_TIMER_STOP(1, &bt, "Switch time of %s", sc->sc_name);
2867 g_journal_do_switch(struct g_class *classp)
2869 struct g_journal_softc *sc;
2870 const struct g_journal_desc *desc;
2875 int error, save, vfslocked;
2879 LIST_FOREACH(gp, &classp->geom, geom) {
2883 if (sc->sc_flags & GJF_DEVICE_DESTROY)
2885 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE)
2887 mtx_lock(&sc->sc_mtx);
2888 sc->sc_flags |= GJF_DEVICE_BEFORE_SWITCH;
2889 mtx_unlock(&sc->sc_mtx);
2891 g_topology_unlock();
2894 mtx_lock(&mountlist_mtx);
2895 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2896 if (mp->mnt_gjprovider == NULL)
2898 if (mp->mnt_flag & MNT_RDONLY)
2900 desc = g_journal_find_desc(mp->mnt_stat.f_fstypename);
2903 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
2905 /* mtx_unlock(&mountlist_mtx) was done inside vfs_busy() */
2909 sc = g_journal_find_device(classp, mp->mnt_gjprovider);
2910 g_topology_unlock();
2914 GJ_DEBUG(0, "Cannot find journal geom for %s.",
2915 mp->mnt_gjprovider);
2917 } else if (JEMPTY(sc)) {
2918 mtx_lock(&sc->sc_mtx);
2919 sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH;
2920 mtx_unlock(&sc->sc_mtx);
2921 GJ_DEBUG(3, "No need for %s switch.", sc->sc_name);
2925 mountpoint = mp->mnt_stat.f_mntonname;
2927 vfslocked = VFS_LOCK_GIANT(mp);
2929 error = vn_start_write(NULL, &mp, V_WAIT);
2931 VFS_UNLOCK_GIANT(vfslocked);
2932 GJ_DEBUG(0, "vn_start_write(%s) failed (error=%d).",
2937 save = curthread_pflags_set(TDP_SYNCIO);
2939 GJ_TIMER_START(1, &bt);
2940 vfs_msync(mp, MNT_NOWAIT);
2941 GJ_TIMER_STOP(1, &bt, "Msync time of %s", mountpoint);
2943 GJ_TIMER_START(1, &bt);
2944 error = VFS_SYNC(mp, MNT_NOWAIT);
2946 GJ_TIMER_STOP(1, &bt, "Sync time of %s", mountpoint);
2948 GJ_DEBUG(0, "Cannot sync file system %s (error=%d).",
2952 curthread_pflags_restore(save);
2954 vn_finished_write(mp);
2957 VFS_UNLOCK_GIANT(vfslocked);
2962 * Send BIO_FLUSH before freezing the file system, so it can be
2963 * faster after the freeze.
2965 GJ_TIMER_START(1, &bt);
2966 g_journal_flush_cache(sc);
2967 GJ_TIMER_STOP(1, &bt, "BIO_FLUSH time of %s", sc->sc_name);
2969 GJ_TIMER_START(1, &bt);
2970 error = vfs_write_suspend(mp);
2971 VFS_UNLOCK_GIANT(vfslocked);
2972 GJ_TIMER_STOP(1, &bt, "Suspend time of %s", mountpoint);
2974 GJ_DEBUG(0, "Cannot suspend file system %s (error=%d).",
2979 error = desc->jd_clean(mp);
2983 mtx_lock(&sc->sc_mtx);
2984 g_journal_switch_wait(sc);
2985 mtx_unlock(&sc->sc_mtx);
2987 vfs_write_resume(mp);
2989 mtx_lock(&mountlist_mtx);
2992 mtx_unlock(&mountlist_mtx);
2998 LIST_FOREACH(gp, &g_journal_class.geom, geom) {
3002 mtx_lock(&sc->sc_mtx);
3003 if ((sc->sc_type & GJ_TYPE_COMPLETE) == GJ_TYPE_COMPLETE &&
3004 !(sc->sc_flags & GJF_DEVICE_DESTROY) &&
3005 (sc->sc_flags & GJF_DEVICE_BEFORE_SWITCH)) {
3008 mtx_unlock(&sc->sc_mtx);
3011 g_topology_unlock();
3015 mtx_assert(&sc->sc_mtx, MA_OWNED);
3016 g_journal_switch_wait(sc);
3017 mtx_unlock(&sc->sc_mtx);
3022 * TODO: Switcher thread should be started on first geom creation and killed on
3023 * last geom destruction.
3026 g_journal_switcher(void *arg)
3033 curthread->td_pflags |= TDP_NORUNNINGBUF;
3035 g_journal_switcher_wokenup = 0;
3036 error = tsleep(&g_journal_switcher_state, PRIBIO, "jsw:wait",
3037 g_journal_switch_time * hz);
3038 if (g_journal_switcher_state == GJ_SWITCHER_DIE) {
3039 g_journal_switcher_state = GJ_SWITCHER_DIED;
3040 GJ_DEBUG(1, "Switcher exiting.");
3041 wakeup(&g_journal_switcher_state);
3044 if (error == 0 && g_journal_sync_requested == 0) {
3045 GJ_DEBUG(1, "Out of cache, force switch (used=%u "
3046 "limit=%u).", g_journal_cache_used,
3047 g_journal_cache_limit);
3049 GJ_TIMER_START(1, &bt);
3050 g_journal_do_switch(mp);
3051 GJ_TIMER_STOP(1, &bt, "Entire switch time");
3052 if (g_journal_sync_requested > 0) {
3053 g_journal_sync_requested = 0;
3054 wakeup(&g_journal_sync_requested);