2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 #include <afsconfig.h>
11 #include <afs/param.h>
16 #include <sys/types.h>
22 #include <netinet/in.h>
37 #include <afs/afsutil.h>
39 #define UBIK_INTERNALS
43 /* This module is responsible for determining when the system has
44 * recovered to the point that it can handle new transactions. It
45 * replays logs, polls to determine the current dbase after a crash,
46 * and distributes the new database to the others.
49 /* The sync site associates a version number with each database. It
50 * broadcasts the version associated with its current dbase in every
51 * one of its beacon messages. When the sync site send a dbase to a
52 * server, it also sends the db's version. A non-sync site server can
53 * tell if it has the right dbase version by simply comparing the
54 * version from the beacon message (uvote_dbVersion) with the version
55 * associated with the database (ubik_dbase->version). The sync site
56 * itself simply has one counter to keep track of all of this (again
57 * ubik_dbase->version).
60 /* sync site: routine called when the sync site loses its quorum; this
61 * procedure is called "up" from the beacon package. It resyncs the
62 * dbase and nudges the recovery daemon to try to propagate out the
63 * changes. It also resets the recovery daemon's state, since
64 * recovery must potentially find a new dbase to propagate out. This
65 * routine should not do anything with variables used by non-sync site
69 /* if this flag is set, then ubik will use only the primary address
70 ** ( the address specified in the CellServDB) to contact other
71 ** ubik servers. Ubik recovery will not try opening connections
72 ** to the alternate interface addresses.
74 int ubikPrimaryAddrOnly;
77 urecovery_ResetState(void)
80 LWP_NoYieldSignal(&urecovery_state);
84 /* sync site: routine called when a non-sync site server goes down; restarts recovery
85 * process to send missing server the new db when it comes back up.
86 * This routine should not do anything with variables used by non-sync site servers.
89 urecovery_LostServer(void)
91 LWP_NoYieldSignal(&urecovery_state);
95 /* return true iff we have a current database (called by both sync
96 * sites and non-sync sites) How do we determine this? If we're the
97 * sync site, we wait until recovery has finished fetching and
98 * re-labelling its dbase (it may still be trying to propagate it out
99 * to everyone else; that's THEIR problem). If we're not the sync
100 * site, then we must have a dbase labelled with the right version,
101 * and we must have a currently-good sync site.
104 urecovery_AllBetter(register struct ubik_dbase *adbase, int areadAny)
106 register afs_int32 rcode;
108 ubik_dprint("allbetter checking\n");
113 if (ubik_dbase->version.epoch > 1)
114 rcode = 1; /* Happy with any good version of database */
117 /* Check if we're sync site and we've got the right data */
118 else if (ubeacon_AmSyncSite() && (urecovery_state & UBIK_RECHAVEDB)) {
122 /* next, check if we're aux site, and we've ever been sent the
123 * right data (note that if a dbase update fails, we won't think
124 * that the sync site is still the sync site, 'cause it won't talk
125 * to us until a timeout period has gone by. When we recover, we
126 * leave this clear until we get a new dbase */
127 else if ((uvote_GetSyncSite() && (vcmp(ubik_dbVersion, ubik_dbase->version) == 0))) { /* && order is important */
131 ubik_dprint("allbetter: returning %d\n", rcode);
135 /* abort all transactions on this database */
137 urecovery_AbortAll(struct ubik_dbase *adbase)
139 register struct ubik_trans *tt;
140 for (tt = adbase->activeTrans; tt; tt = tt->next) {
146 /* this routine aborts the current remote transaction, if any, if the tid is wrong */
148 urecovery_CheckTid(register struct ubik_tid *atid)
150 if (ubik_currentTrans) {
151 /* there is remote write trans, see if we match, see if this
152 * is a new transaction */
153 if (atid->epoch != ubik_currentTrans->tid.epoch
154 || atid->counter > ubik_currentTrans->tid.counter) {
155 /* don't match, abort it */
156 /* If the thread is not waiting for lock - ok to end it */
157 #if !defined(UBIK_PAUSE)
158 if (ubik_currentTrans->locktype != LOCKWAIT) {
159 #endif /* UBIK_PAUSE */
160 udisk_end(ubik_currentTrans);
161 #if !defined(UBIK_PAUSE)
163 #endif /* UBIK_PAUSE */
164 ubik_currentTrans = (struct ubik_trans *)0;
169 /* log format is defined here, and implicitly in disk.c
171 * 4 byte opcode, followed by parameters, each 4 bytes long. All integers
172 * are in logged in network standard byte order, in case we want to move logs
173 * from machine-to-machine someday.
175 * Begin transaction: opcode
176 * Commit transaction: opcode, version (8 bytes)
177 * Truncate file: opcode, file number, length
178 * Abort transaction: opcode
179 * Write data: opcode, file, position, length, <length> data bytes
181 * A very simple routine, it just replays the log. Note that this is a new-value only log, which
182 * implies that no uncommitted data is written to the dbase: one writes data to the log, including
183 * the commit record, then we allow data to be written through to the dbase. In our particular
184 * implementation, once a transaction is done, we write out the pages to the database, so that
185 * our buffer package doesn't have to know about stable and uncommitted data in the memory buffers:
186 * any changed data while there is an uncommitted write transaction can be zapped during an
187 * abort and the remaining dbase on the disk is exactly the right dbase, without having to read
193 ReplayLog(register struct ubik_dbase *adbase)
196 register afs_int32 code, tpos;
198 afs_int32 len, thisSize, tfile, filePos;
200 afs_int32 syncFile = -1;
201 afs_int32 data[1024];
203 /* read the lock twice, once to see whether we have a transaction to deal
204 * with that committed, (theoretically, we should support more than one
205 * trans in the log at once, but not yet), and once replaying the
209 /* for now, assume that all ops in log pertain to one transaction; see if there's a commit */
212 (*adbase->read) (adbase, LOGFILE, &opcode, tpos,
214 if (code != sizeof(afs_int32))
216 if (opcode == LOGNEW) {
217 /* handle begin trans */
218 tpos += sizeof(afs_int32);
219 } else if (opcode == LOGABORT)
221 else if (opcode == LOGEND) {
224 } else if (opcode == LOGTRUNCATE) {
227 (*adbase->read) (adbase, LOGFILE, buffer, tpos,
228 2 * sizeof(afs_int32));
229 if (code != 2 * sizeof(afs_int32))
230 break; /* premature eof or io error */
231 tpos += 2 * sizeof(afs_int32);
232 } else if (opcode == LOGDATA) {
235 (*adbase->read) (adbase, LOGFILE, buffer, tpos,
236 3 * sizeof(afs_int32));
237 if (code != 3 * sizeof(afs_int32))
239 /* otherwise, skip over the data bytes, too */
240 tpos += buffer[2] + 3 * sizeof(afs_int32);
242 ubik_dprint("corrupt log opcode (%d) at position %d\n", opcode,
244 break; /* corrupt log! */
248 /* actually do the replay; log should go all the way through the commit record, since
249 * we just read it above. */
255 (*adbase->read) (adbase, LOGFILE, &opcode, tpos,
257 if (code != sizeof(afs_int32))
259 if (opcode == LOGNEW) {
260 /* handle begin trans */
261 tpos += sizeof(afs_int32);
262 } else if (opcode == LOGABORT)
263 panic("log abort\n");
264 else if (opcode == LOGEND) {
267 (*adbase->read) (adbase, LOGFILE, buffer, tpos,
268 2 * sizeof(afs_int32));
269 if (code != 2 * sizeof(afs_int32))
271 code = (*adbase->setlabel) (adbase, 0, buffer);
275 break; /* all done now */
276 } else if (opcode == LOGTRUNCATE) {
279 (*adbase->read) (adbase, LOGFILE, buffer, tpos,
280 2 * sizeof(afs_int32));
281 if (code != 2 * sizeof(afs_int32))
282 break; /* premature eof or io error */
283 tpos += 2 * sizeof(afs_int32);
285 (*adbase->truncate) (adbase, ntohl(buffer[0]),
289 } else if (opcode == LOGDATA) {
292 (*adbase->read) (adbase, LOGFILE, buffer, tpos,
293 3 * sizeof(afs_int32));
294 if (code != 3 * sizeof(afs_int32))
296 tpos += 3 * sizeof(afs_int32);
297 /* otherwise, skip over the data bytes, too */
298 len = ntohl(buffer[2]); /* total number of bytes to copy */
299 filePos = ntohl(buffer[1]);
300 tfile = ntohl(buffer[0]);
301 /* try to minimize file syncs */
302 if (syncFile != tfile) {
304 code = (*adbase->sync) (adbase, syncFile);
312 thisSize = (len > sizeof(data) ? sizeof(data) : len);
313 /* copy sizeof(data) buffer bytes at a time */
315 (*adbase->read) (adbase, LOGFILE, data, tpos,
317 if (code != thisSize)
320 (*adbase->write) (adbase, tfile, data, filePos,
322 if (code != thisSize)
329 ubik_dprint("corrupt log opcode (%d) at position %d\n",
331 break; /* corrupt log! */
336 code = (*adbase->sync) (adbase, syncFile);
340 ubik_dprint("Log read error on pass 2\n");
345 /* now truncate the log, we're done with it */
346 code = (*adbase->truncate) (adbase, LOGFILE, 0);
350 /* Called at initialization to figure out version of the dbase we really have.
351 * This routine is called after replaying the log; it reads the restored labels.
354 InitializeDB(register struct ubik_dbase *adbase)
356 register afs_int32 code;
358 code = (*adbase->getlabel) (adbase, 0, &adbase->version);
360 /* try setting the label to a new value */
361 adbase->version.epoch = 1; /* value for newly-initialized db */
362 adbase->version.counter = 1;
363 code = (*adbase->setlabel) (adbase, 0, &adbase->version);
365 /* failed, try to set it back */
366 adbase->version.epoch = 0;
367 adbase->version.counter = 0;
368 (*adbase->setlabel) (adbase, 0, &adbase->version);
370 LWP_NoYieldSignal(&adbase->version);
375 /* initialize the local dbase
376 * We replay the logs and then read the resulting file to figure out what version we've really got.
379 urecovery_Initialize(register struct ubik_dbase *adbase)
381 register afs_int32 code;
383 code = ReplayLog(adbase);
386 code = InitializeDB(adbase);
390 /* Main interaction loop for the recovery manager
391 * The recovery light-weight process only runs when you're the
392 * synchronization site. It performs the following tasks, if and only
393 * if the prerequisite tasks have been performed successfully (it
394 * keeps track of which ones have been performed in its bit map,
397 * First, it is responsible for probing that all servers are up. This
398 * is the only operation that must be performed even if this is not
399 * yet the sync site, since otherwise this site may not notice that
400 * enough other machines are running to even elect this guy to be the
403 * After that, the recovery process does nothing until the beacon and
404 * voting modules manage to get this site elected sync site.
406 * After becoming sync site, recovery first attempts to find the best
407 * database available in the network (it must do this in order to
408 * ensure finding the latest committed data). After finding the right
409 * database, it must fetch this dbase to the sync site.
411 * After fetching the dbase, it relabels it with a new version number,
412 * to ensure that everyone recognizes this dbase as the most recent
415 * One the dbase has been relabelled, this machine can start handling
416 * requests. However, the recovery module still has one more task:
417 * propagating the dbase out to everyone who is up in the network.
420 urecovery_Interact(void)
422 afs_int32 code, tcode;
423 struct ubik_server *bestServer = NULL;
424 struct ubik_server *ts;
425 int dbok, doingRPC, now;
426 afs_int32 lastProbeTime, lastDBVCheck;
427 /* if we're the sync site, the best db version we've found yet */
428 static struct ubik_version bestDBVersion;
429 struct ubik_version tversion;
431 int length, tlen, offset, file, nbytes;
432 struct rx_call *rxcall;
434 struct ubik_stat ubikstat;
435 struct in_addr inAddr;
437 /* otherwise, begin interaction */
442 /* Run through this loop every 4 seconds */
445 IOMGR_Select(0, 0, 0, 0, &tv);
447 ubik_dprint("recovery running in state %x\n", urecovery_state);
449 /* Every 30 seconds, check all the down servers and mark them
450 * as up if they respond. When a server comes up or found to
451 * not be current, then re-find the the best database and
454 if ((now = FT_ApproxTime()) > 30 + lastProbeTime) {
455 for (ts = ubik_servers, doingRPC = 0; ts; ts = ts->next) {
461 urecovery_state &= ~UBIK_RECFOUNDDB;
463 } else if (!ts->currentDB) {
464 urecovery_state &= ~UBIK_RECFOUNDDB;
468 now = FT_ApproxTime();
472 /* Mark whether we are the sync site */
473 if (!ubeacon_AmSyncSite()) {
474 urecovery_state &= ~UBIK_RECSYNCSITE;
475 continue; /* nothing to do */
477 urecovery_state |= UBIK_RECSYNCSITE;
479 /* If a server has just come up or if we have not found the
480 * most current database, then go find the most current db.
482 if (!(urecovery_state & UBIK_RECFOUNDDB)) {
483 bestServer = (struct ubik_server *)0;
484 bestDBVersion.epoch = 0;
485 bestDBVersion.counter = 0;
486 for (ts = ubik_servers; ts; ts = ts->next) {
488 continue; /* don't bother with these guys */
491 code = DISK_GetVersion(ts->disk_rxcid, &ts->version);
493 /* perhaps this is the best version */
494 if (vcmp(ts->version, bestDBVersion) > 0) {
495 /* new best version */
496 bestDBVersion = ts->version;
501 /* take into consideration our version. Remember if we,
502 * the sync site, have the best version. Also note that
503 * we may need to send the best version out.
505 if (vcmp(ubik_dbase->version, bestDBVersion) >= 0) {
506 bestDBVersion = ubik_dbase->version;
507 bestServer = (struct ubik_server *)0;
508 urecovery_state |= UBIK_RECHAVEDB;
510 /* Clear the flag only when we know we have to retrieve
511 * the db. Because urecovery_AllBetter() looks at it.
513 urecovery_state &= ~UBIK_RECHAVEDB;
515 lastDBVCheck = FT_ApproxTime();
516 urecovery_state |= UBIK_RECFOUNDDB;
517 urecovery_state &= ~UBIK_RECSENTDB;
519 #if defined(UBIK_PAUSE)
520 /* it's not possible for UBIK_RECFOUNDDB not to be set here.
521 * However, we might have lost UBIK_RECSYNCSITE, and that
524 if (!(urecovery_state & UBIK_RECSYNCSITE))
525 continue; /* lost sync */
527 if (!(urecovery_state & UBIK_RECFOUNDDB))
528 continue; /* not ready */
529 #endif /* UBIK_PAUSE */
531 /* If we, the sync site, do not have the best db version, then
532 * go and get it from the server that does.
534 if ((urecovery_state & UBIK_RECHAVEDB) || !bestServer) {
535 urecovery_state |= UBIK_RECHAVEDB;
537 /* we don't have the best version; we should fetch it. */
538 #if defined(UBIK_PAUSE)
541 ObtainWriteLock(&ubik_dbase->versionLock);
542 #endif /* UBIK_PAUSE */
543 urecovery_AbortAll(ubik_dbase);
545 /* Rx code to do the Bulk fetch */
548 rxcall = rx_NewCall(bestServer->disk_rxcid);
550 ubik_print("Ubik: Synchronize database with server %s\n",
551 afs_inet_ntoa(bestServer->addr[0]));
553 code = StartDISK_GetFile(rxcall, file);
555 ubik_dprint("StartDiskGetFile failed=%d\n", code);
558 nbytes = rx_Read(rxcall, &length, sizeof(afs_int32));
559 length = ntohl(length);
560 if (nbytes != sizeof(afs_int32)) {
561 ubik_dprint("Rx-read length error=%d\n", code = BULK_ERROR);
566 /* Truncate the file firest */
567 code = (*ubik_dbase->truncate) (ubik_dbase, file, 0);
569 ubik_dprint("truncate io error=%d\n", code);
573 /* give invalid label during file transit */
575 tversion.counter = 0;
576 code = (*ubik_dbase->setlabel) (ubik_dbase, file, &tversion);
578 ubik_dprint("setlabel io error=%d\n", code);
583 tlen = (length > sizeof(tbuffer) ? sizeof(tbuffer) : length);
584 nbytes = rx_Read(rxcall, tbuffer, tlen);
585 if (nbytes != tlen) {
586 ubik_dprint("Rx-read bulk error=%d\n", code = BULK_ERROR);
591 (*ubik_dbase->write) (ubik_dbase, file, tbuffer, offset,
593 if (nbytes != tlen) {
600 code = EndDISK_GetFile(rxcall, &tversion);
602 tcode = rx_EndCall(rxcall, code);
606 /* we got a new file, set up its header */
607 urecovery_state |= UBIK_RECHAVEDB;
608 memcpy(&ubik_dbase->version, &tversion,
609 sizeof(struct ubik_version));
610 (*ubik_dbase->sync) (ubik_dbase, 0); /* get data out first */
611 /* after data is good, sync disk with correct label */
613 (*ubik_dbase->setlabel) (ubik_dbase, 0,
614 &ubik_dbase->version);
617 ubik_dbase->version.epoch = 0;
618 ubik_dbase->version.counter = 0;
619 ubik_print("Ubik: Synchronize database failed (error = %d)\n",
622 ubik_print("Ubik: Synchronize database completed\n");
624 udisk_Invalidate(ubik_dbase, 0); /* data has changed */
625 LWP_NoYieldSignal(&ubik_dbase->version);
626 #if defined(UBIK_PAUSE)
629 ReleaseWriteLock(&ubik_dbase->versionLock);
630 #endif /* UBIK_PAUSE */
632 #if defined(UBIK_PAUSE)
633 if (!(urecovery_state & UBIK_RECSYNCSITE))
634 continue; /* lost sync */
635 #endif /* UBIK_PAUSE */
636 if (!(urecovery_state & UBIK_RECHAVEDB))
637 continue; /* not ready */
639 /* If the database was newly initialized, then when we establish quorum, write
640 * a new label. This allows urecovery_AllBetter() to allow access for reads.
641 * Setting it to 2 also allows another site to come along with a newer
642 * database and overwrite this one.
644 if (ubik_dbase->version.epoch == 1) {
645 #if defined(UBIK_PAUSE)
648 ObtainWriteLock(&ubik_dbase->versionLock);
649 #endif /* UBIK_PAUSE */
650 urecovery_AbortAll(ubik_dbase);
652 ubik_dbase->version.epoch = ubik_epochTime;
653 ubik_dbase->version.counter = 1;
655 (*ubik_dbase->setlabel) (ubik_dbase, 0, &ubik_dbase->version);
656 udisk_Invalidate(ubik_dbase, 0); /* data may have changed */
657 LWP_NoYieldSignal(&ubik_dbase->version);
658 #if defined(UBIK_PAUSE)
661 ReleaseWriteLock(&ubik_dbase->versionLock);
662 #endif /* UBIK_PAUSE */
665 /* Check the other sites and send the database to them if they
666 * do not have the current db.
668 if (!(urecovery_state & UBIK_RECSENTDB)) {
669 /* now propagate out new version to everyone else */
670 dbok = 1; /* start off assuming they all worked */
672 #if defined(UBIK_PAUSE)
675 ObtainWriteLock(&ubik_dbase->versionLock);
676 #endif /* UBIK_PAUSE */
678 * Check if a write transaction is in progress. We can't send the
679 * db when a write is in progress here because the db would be
680 * obsolete as soon as it goes there. Also, ops after the begin
681 * trans would reach the recepient and wouldn't find a transaction
682 * pending there. Frankly, I don't think it's possible to get past
683 * the write-lock above if there is a write transaction in progress,
684 * but then, it won't hurt to check, will it?
686 if (ubik_dbase->flags & DBWRITING) {
691 while ((ubik_dbase->flags & DBWRITING) && (safety < 500)) {
692 #if defined(UBIK_PAUSE)
695 ReleaseWriteLock(&ubik_dbase->versionLock);
696 #endif /* UBIK_PAUSE */
697 /* sleep for a little while */
698 IOMGR_Select(0, 0, 0, 0, &tv);
701 #if defined(UBIK_PAUSE)
704 ObtainWriteLock(&ubik_dbase->versionLock);
705 #endif /* UBIK_PAUSE */
709 for (ts = ubik_servers; ts; ts = ts->next) {
710 inAddr.s_addr = ts->addr[0];
712 ubik_dprint("recovery cannot send version to %s\n",
713 afs_inet_ntoa(inAddr.s_addr));
717 ubik_dprint("recovery sending version to %s\n",
718 afs_inet_ntoa(inAddr.s_addr));
719 if (vcmp(ts->version, ubik_dbase->version) != 0) {
720 ubik_dprint("recovery stating local database\n");
722 /* Rx code to do the Bulk Store */
723 code = (*ubik_dbase->stat) (ubik_dbase, 0, &ubikstat);
725 length = ubikstat.size;
727 rxcall = rx_NewCall(ts->disk_rxcid);
729 StartDISK_SendFile(rxcall, file, length,
730 &ubik_dbase->version);
732 ubik_dprint("StartDiskSendFile failed=%d\n",
739 sizeof(tbuffer) ? sizeof(tbuffer) : length);
741 (*ubik_dbase->read) (ubik_dbase, file,
742 tbuffer, offset, tlen);
743 if (nbytes != tlen) {
744 ubik_dprint("Local disk read error=%d\n",
748 nbytes = rx_Write(rxcall, tbuffer, tlen);
749 if (nbytes != tlen) {
750 ubik_dprint("Rx-write bulk error=%d\n", code =
757 code = EndDISK_SendFile(rxcall);
759 code = rx_EndCall(rxcall, code);
762 /* we set a new file, process its header */
763 ts->version = ubik_dbase->version;
768 /* mark file up to date */
772 #if defined(UBIK_PAUSE)
775 ReleaseWriteLock(&ubik_dbase->versionLock);
776 #endif /* UBIK_PAUSE */
778 urecovery_state |= UBIK_RECSENTDB;
784 ** send a Probe to all the network address of this server
785 ** Return 0 if success, else return 1
788 DoProbe(struct ubik_server *server)
790 struct rx_connection *conns[UBIK_MAX_INTERFACE_ADDR];
791 struct rx_connection *connSuccess = 0;
795 extern afs_int32 ubikSecIndex;
796 extern struct rx_securityClass *ubikSecClass;
798 for (i = 0; (addr = server->addr[i]) && (i < UBIK_MAX_INTERFACE_ADDR);
801 rx_NewConnection(addr, ubik_callPortal, DISK_SERVICE_ID,
802 ubikSecClass, ubikSecIndex);
804 /* user requirement to use only the primary interface */
805 if (ubikPrimaryAddrOnly) {
810 assert(i); /* at least one interface address for this server */
814 if (!multi_error) { /* first success */
815 addr = server->addr[multi_i]; /* successful interface addr */
817 if (server->disk_rxcid) /* destroy existing conn */
818 rx_DestroyConnection(server->disk_rxcid);
819 if (server->vote_rxcid)
820 rx_DestroyConnection(server->vote_rxcid);
822 /* make new connections */
823 server->disk_rxcid = conns[multi_i];
824 server->vote_rxcid = rx_NewConnection(addr, ubik_callPortal, VOTE_SERVICE_ID, ubikSecClass, ubikSecIndex); /* for vote reqs */
826 connSuccess = conns[multi_i];
827 strcpy(buffer, (char *)afs_inet_ntoa(server->addr[0]));
829 ("ubik:server %s is back up: will be contacted through %s\n",
830 buffer, afs_inet_ntoa(addr));
836 /* Destroy all connections except the one on which we succeeded */
837 for (j = 0; j < i; j++)
838 if (conns[j] != connSuccess)
839 rx_DestroyConnection(conns[j]);
842 ubik_dprint("ubik:server %s still down\n",
843 afs_inet_ntoa(server->addr[0]));
846 return 0; /* success */
848 return 1; /* failure */