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;
170 /* log format is defined here, and implicitly in disk.c
172 * 4 byte opcode, followed by parameters, each 4 bytes long. All integers
173 * are in logged in network standard byte order, in case we want to move logs
174 * from machine-to-machine someday.
176 * Begin transaction: opcode
177 * Commit transaction: opcode, version (8 bytes)
178 * Truncate file: opcode, file number, length
179 * Abort transaction: opcode
180 * Write data: opcode, file, position, length, <length> data bytes
182 * A very simple routine, it just replays the log. Note that this is a new-value only log, which
183 * implies that no uncommitted data is written to the dbase: one writes data to the log, including
184 * the commit record, then we allow data to be written through to the dbase. In our particular
185 * implementation, once a transaction is done, we write out the pages to the database, so that
186 * our buffer package doesn't have to know about stable and uncommitted data in the memory buffers:
187 * any changed data while there is an uncommitted write transaction can be zapped during an
188 * abort and the remaining dbase on the disk is exactly the right dbase, without having to read
194 ReplayLog(register struct ubik_dbase *adbase)
197 register afs_int32 code, tpos;
199 afs_int32 len, thisSize, tfile, filePos;
201 afs_int32 syncFile = -1;
202 afs_int32 data[1024];
204 /* read the lock twice, once to see whether we have a transaction to deal
205 * with that committed, (theoretically, we should support more than one
206 * trans in the log at once, but not yet), and once replaying the
210 /* for now, assume that all ops in log pertain to one transaction; see if there's a commit */
213 (*adbase->read) (adbase, LOGFILE, (char *)&opcode, tpos,
215 if (code != sizeof(afs_int32))
217 if (opcode == LOGNEW) {
218 /* handle begin trans */
219 tpos += sizeof(afs_int32);
220 } else if (opcode == LOGABORT)
222 else if (opcode == LOGEND) {
225 } else if (opcode == LOGTRUNCATE) {
228 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
229 2 * sizeof(afs_int32));
230 if (code != 2 * sizeof(afs_int32))
231 break; /* premature eof or io error */
232 tpos += 2 * sizeof(afs_int32);
233 } else if (opcode == LOGDATA) {
236 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
237 3 * sizeof(afs_int32));
238 if (code != 3 * sizeof(afs_int32))
240 /* otherwise, skip over the data bytes, too */
241 tpos += buffer[2] + 3 * sizeof(afs_int32);
243 ubik_dprint("corrupt log opcode (%d) at position %d\n", opcode,
245 break; /* corrupt log! */
249 /* actually do the replay; log should go all the way through the commit record, since
250 * we just read it above. */
256 (*adbase->read) (adbase, LOGFILE, (char *)&opcode, tpos,
258 if (code != sizeof(afs_int32))
260 if (opcode == LOGNEW) {
261 /* handle begin trans */
262 tpos += sizeof(afs_int32);
263 } else if (opcode == LOGABORT)
264 panic("log abort\n");
265 else if (opcode == LOGEND) {
268 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
269 2 * sizeof(afs_int32));
270 if (code != 2 * sizeof(afs_int32))
272 code = (*adbase->setlabel) (adbase, 0, (ubik_version *)buffer);
276 break; /* all done now */
277 } else if (opcode == LOGTRUNCATE) {
280 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
281 2 * sizeof(afs_int32));
282 if (code != 2 * sizeof(afs_int32))
283 break; /* premature eof or io error */
284 tpos += 2 * sizeof(afs_int32);
286 (*adbase->truncate) (adbase, ntohl(buffer[0]),
290 } else if (opcode == LOGDATA) {
293 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
294 3 * sizeof(afs_int32));
295 if (code != 3 * sizeof(afs_int32))
297 tpos += 3 * sizeof(afs_int32);
298 /* otherwise, skip over the data bytes, too */
299 len = ntohl(buffer[2]); /* total number of bytes to copy */
300 filePos = ntohl(buffer[1]);
301 tfile = ntohl(buffer[0]);
302 /* try to minimize file syncs */
303 if (syncFile != tfile) {
305 code = (*adbase->sync) (adbase, syncFile);
313 thisSize = (len > sizeof(data) ? sizeof(data) : len);
314 /* copy sizeof(data) buffer bytes at a time */
316 (*adbase->read) (adbase, LOGFILE, (char *)data, tpos,
318 if (code != thisSize)
321 (*adbase->write) (adbase, tfile, (char *)data, filePos,
323 if (code != thisSize)
330 ubik_dprint("corrupt log opcode (%d) at position %d\n",
332 break; /* corrupt log! */
337 code = (*adbase->sync) (adbase, syncFile);
341 ubik_dprint("Log read error on pass 2\n");
346 /* now truncate the log, we're done with it */
347 code = (*adbase->truncate) (adbase, LOGFILE, 0);
351 /* Called at initialization to figure out version of the dbase we really have.
352 * This routine is called after replaying the log; it reads the restored labels.
355 InitializeDB(register struct ubik_dbase *adbase)
357 register afs_int32 code;
359 code = (*adbase->getlabel) (adbase, 0, &adbase->version);
361 /* try setting the label to a new value */
362 adbase->version.epoch = 1; /* value for newly-initialized db */
363 adbase->version.counter = 1;
364 code = (*adbase->setlabel) (adbase, 0, &adbase->version);
366 /* failed, try to set it back */
367 adbase->version.epoch = 0;
368 adbase->version.counter = 0;
369 (*adbase->setlabel) (adbase, 0, &adbase->version);
371 LWP_NoYieldSignal(&adbase->version);
376 /* initialize the local dbase
377 * We replay the logs and then read the resulting file to figure out what version we've really got.
380 urecovery_Initialize(register struct ubik_dbase *adbase)
382 register afs_int32 code;
384 code = ReplayLog(adbase);
387 code = InitializeDB(adbase);
391 /* Main interaction loop for the recovery manager
392 * The recovery light-weight process only runs when you're the
393 * synchronization site. It performs the following tasks, if and only
394 * if the prerequisite tasks have been performed successfully (it
395 * keeps track of which ones have been performed in its bit map,
398 * First, it is responsible for probing that all servers are up. This
399 * is the only operation that must be performed even if this is not
400 * yet the sync site, since otherwise this site may not notice that
401 * enough other machines are running to even elect this guy to be the
404 * After that, the recovery process does nothing until the beacon and
405 * voting modules manage to get this site elected sync site.
407 * After becoming sync site, recovery first attempts to find the best
408 * database available in the network (it must do this in order to
409 * ensure finding the latest committed data). After finding the right
410 * database, it must fetch this dbase to the sync site.
412 * After fetching the dbase, it relabels it with a new version number,
413 * to ensure that everyone recognizes this dbase as the most recent
416 * One the dbase has been relabelled, this machine can start handling
417 * requests. However, the recovery module still has one more task:
418 * propagating the dbase out to everyone who is up in the network.
421 urecovery_Interact(void)
423 afs_int32 code, tcode;
424 struct ubik_server *bestServer = NULL;
425 struct ubik_server *ts;
426 int dbok, doingRPC, now;
427 afs_int32 lastProbeTime, lastDBVCheck;
428 /* if we're the sync site, the best db version we've found yet */
429 static struct ubik_version bestDBVersion;
430 struct ubik_version tversion;
432 int length, tlen, offset, file, nbytes;
433 struct rx_call *rxcall;
435 struct ubik_stat ubikstat;
436 struct in_addr inAddr;
438 /* otherwise, begin interaction */
443 /* Run through this loop every 4 seconds */
446 IOMGR_Select(0, 0, 0, 0, &tv);
448 ubik_dprint("recovery running in state %x\n", urecovery_state);
450 /* Every 30 seconds, check all the down servers and mark them
451 * as up if they respond. When a server comes up or found to
452 * not be current, then re-find the the best database and
455 if ((now = FT_ApproxTime()) > 30 + lastProbeTime) {
456 for (ts = ubik_servers, doingRPC = 0; ts; ts = ts->next) {
462 urecovery_state &= ~UBIK_RECFOUNDDB;
464 } else if (!ts->currentDB) {
465 urecovery_state &= ~UBIK_RECFOUNDDB;
469 now = FT_ApproxTime();
473 /* Mark whether we are the sync site */
474 if (!ubeacon_AmSyncSite()) {
475 urecovery_state &= ~UBIK_RECSYNCSITE;
476 continue; /* nothing to do */
478 urecovery_state |= UBIK_RECSYNCSITE;
480 /* If a server has just come up or if we have not found the
481 * most current database, then go find the most current db.
483 if (!(urecovery_state & UBIK_RECFOUNDDB)) {
484 bestServer = (struct ubik_server *)0;
485 bestDBVersion.epoch = 0;
486 bestDBVersion.counter = 0;
487 for (ts = ubik_servers; ts; ts = ts->next) {
489 continue; /* don't bother with these guys */
492 code = DISK_GetVersion(ts->disk_rxcid, &ts->version);
494 /* perhaps this is the best version */
495 if (vcmp(ts->version, bestDBVersion) > 0) {
496 /* new best version */
497 bestDBVersion = ts->version;
502 /* take into consideration our version. Remember if we,
503 * the sync site, have the best version. Also note that
504 * we may need to send the best version out.
506 if (vcmp(ubik_dbase->version, bestDBVersion) >= 0) {
507 bestDBVersion = ubik_dbase->version;
508 bestServer = (struct ubik_server *)0;
509 urecovery_state |= UBIK_RECHAVEDB;
511 /* Clear the flag only when we know we have to retrieve
512 * the db. Because urecovery_AllBetter() looks at it.
514 urecovery_state &= ~UBIK_RECHAVEDB;
516 lastDBVCheck = FT_ApproxTime();
517 urecovery_state |= UBIK_RECFOUNDDB;
518 urecovery_state &= ~UBIK_RECSENTDB;
520 #if defined(UBIK_PAUSE)
521 /* it's not possible for UBIK_RECFOUNDDB not to be set here.
522 * However, we might have lost UBIK_RECSYNCSITE, and that
525 if (!(urecovery_state & UBIK_RECSYNCSITE))
526 continue; /* lost sync */
528 if (!(urecovery_state & UBIK_RECFOUNDDB))
529 continue; /* not ready */
530 #endif /* UBIK_PAUSE */
532 /* If we, the sync site, do not have the best db version, then
533 * go and get it from the server that does.
535 if ((urecovery_state & UBIK_RECHAVEDB) || !bestServer) {
536 urecovery_state |= UBIK_RECHAVEDB;
538 /* we don't have the best version; we should fetch it. */
539 #if defined(UBIK_PAUSE)
542 ObtainWriteLock(&ubik_dbase->versionLock);
543 #endif /* UBIK_PAUSE */
544 urecovery_AbortAll(ubik_dbase);
546 /* Rx code to do the Bulk fetch */
549 rxcall = rx_NewCall(bestServer->disk_rxcid);
551 ubik_print("Ubik: Synchronize database with server %s\n",
552 afs_inet_ntoa(bestServer->addr[0]));
554 code = StartDISK_GetFile(rxcall, file);
556 ubik_dprint("StartDiskGetFile failed=%d\n", code);
559 nbytes = rx_Read(rxcall, (char *)&length, sizeof(afs_int32));
560 length = ntohl(length);
561 if (nbytes != sizeof(afs_int32)) {
562 ubik_dprint("Rx-read length error=%d\n", code = BULK_ERROR);
567 /* Truncate the file firest */
568 code = (*ubik_dbase->truncate) (ubik_dbase, file, 0);
570 ubik_dprint("truncate io error=%d\n", code);
574 /* give invalid label during file transit */
576 tversion.counter = 0;
577 code = (*ubik_dbase->setlabel) (ubik_dbase, file, &tversion);
579 ubik_dprint("setlabel io error=%d\n", code);
584 tlen = (length > sizeof(tbuffer) ? sizeof(tbuffer) : length);
585 nbytes = rx_Read(rxcall, tbuffer, tlen);
586 if (nbytes != tlen) {
587 ubik_dprint("Rx-read bulk error=%d\n", code = BULK_ERROR);
592 (*ubik_dbase->write) (ubik_dbase, file, tbuffer, offset,
594 if (nbytes != tlen) {
601 code = EndDISK_GetFile(rxcall, &tversion);
603 tcode = rx_EndCall(rxcall, code);
607 /* we got a new file, set up its header */
608 urecovery_state |= UBIK_RECHAVEDB;
609 memcpy(&ubik_dbase->version, &tversion,
610 sizeof(struct ubik_version));
611 (*ubik_dbase->sync) (ubik_dbase, 0); /* get data out first */
612 /* after data is good, sync disk with correct label */
614 (*ubik_dbase->setlabel) (ubik_dbase, 0,
615 &ubik_dbase->version);
618 ubik_dbase->version.epoch = 0;
619 ubik_dbase->version.counter = 0;
620 ubik_print("Ubik: Synchronize database failed (error = %d)\n",
623 ubik_print("Ubik: Synchronize database completed\n");
625 udisk_Invalidate(ubik_dbase, 0); /* data has changed */
626 LWP_NoYieldSignal(&ubik_dbase->version);
627 #if defined(UBIK_PAUSE)
630 ReleaseWriteLock(&ubik_dbase->versionLock);
631 #endif /* UBIK_PAUSE */
633 #if defined(UBIK_PAUSE)
634 if (!(urecovery_state & UBIK_RECSYNCSITE))
635 continue; /* lost sync */
636 #endif /* UBIK_PAUSE */
637 if (!(urecovery_state & UBIK_RECHAVEDB))
638 continue; /* not ready */
640 /* If the database was newly initialized, then when we establish quorum, write
641 * a new label. This allows urecovery_AllBetter() to allow access for reads.
642 * Setting it to 2 also allows another site to come along with a newer
643 * database and overwrite this one.
645 if (ubik_dbase->version.epoch == 1) {
646 #if defined(UBIK_PAUSE)
649 ObtainWriteLock(&ubik_dbase->versionLock);
650 #endif /* UBIK_PAUSE */
651 urecovery_AbortAll(ubik_dbase);
653 ubik_dbase->version.epoch = ubik_epochTime;
654 ubik_dbase->version.counter = 1;
656 (*ubik_dbase->setlabel) (ubik_dbase, 0, &ubik_dbase->version);
657 udisk_Invalidate(ubik_dbase, 0); /* data may have changed */
658 LWP_NoYieldSignal(&ubik_dbase->version);
659 #if defined(UBIK_PAUSE)
662 ReleaseWriteLock(&ubik_dbase->versionLock);
663 #endif /* UBIK_PAUSE */
666 /* Check the other sites and send the database to them if they
667 * do not have the current db.
669 if (!(urecovery_state & UBIK_RECSENTDB)) {
670 /* now propagate out new version to everyone else */
671 dbok = 1; /* start off assuming they all worked */
673 #if defined(UBIK_PAUSE)
676 ObtainWriteLock(&ubik_dbase->versionLock);
677 #endif /* UBIK_PAUSE */
679 * Check if a write transaction is in progress. We can't send the
680 * db when a write is in progress here because the db would be
681 * obsolete as soon as it goes there. Also, ops after the begin
682 * trans would reach the recepient and wouldn't find a transaction
683 * pending there. Frankly, I don't think it's possible to get past
684 * the write-lock above if there is a write transaction in progress,
685 * but then, it won't hurt to check, will it?
687 if (ubik_dbase->flags & DBWRITING) {
692 while ((ubik_dbase->flags & DBWRITING) && (safety < 500)) {
693 #if defined(UBIK_PAUSE)
696 ReleaseWriteLock(&ubik_dbase->versionLock);
697 #endif /* UBIK_PAUSE */
698 /* sleep for a little while */
699 IOMGR_Select(0, 0, 0, 0, &tv);
702 #if defined(UBIK_PAUSE)
705 ObtainWriteLock(&ubik_dbase->versionLock);
706 #endif /* UBIK_PAUSE */
710 for (ts = ubik_servers; ts; ts = ts->next) {
711 inAddr.s_addr = ts->addr[0];
713 ubik_dprint("recovery cannot send version to %s\n",
714 afs_inet_ntoa(inAddr.s_addr));
718 ubik_dprint("recovery sending version to %s\n",
719 afs_inet_ntoa(inAddr.s_addr));
720 if (vcmp(ts->version, ubik_dbase->version) != 0) {
721 ubik_dprint("recovery stating local database\n");
723 /* Rx code to do the Bulk Store */
724 code = (*ubik_dbase->stat) (ubik_dbase, 0, &ubikstat);
726 length = ubikstat.size;
728 rxcall = rx_NewCall(ts->disk_rxcid);
730 StartDISK_SendFile(rxcall, file, length,
731 &ubik_dbase->version);
733 ubik_dprint("StartDiskSendFile failed=%d\n",
740 sizeof(tbuffer) ? sizeof(tbuffer) : length);
742 (*ubik_dbase->read) (ubik_dbase, file,
743 tbuffer, offset, tlen);
744 if (nbytes != tlen) {
745 ubik_dprint("Local disk read error=%d\n",
749 nbytes = rx_Write(rxcall, tbuffer, tlen);
750 if (nbytes != tlen) {
751 ubik_dprint("Rx-write bulk error=%d\n", code =
758 code = EndDISK_SendFile(rxcall);
760 code = rx_EndCall(rxcall, code);
763 /* we set a new file, process its header */
764 ts->version = ubik_dbase->version;
769 /* mark file up to date */
773 #if defined(UBIK_PAUSE)
776 ReleaseWriteLock(&ubik_dbase->versionLock);
777 #endif /* UBIK_PAUSE */
779 urecovery_state |= UBIK_RECSENTDB;
785 ** send a Probe to all the network address of this server
786 ** Return 0 if success, else return 1
789 DoProbe(struct ubik_server *server)
791 struct rx_connection *conns[UBIK_MAX_INTERFACE_ADDR];
792 struct rx_connection *connSuccess = 0;
796 extern afs_int32 ubikSecIndex;
797 extern struct rx_securityClass *ubikSecClass;
799 for (i = 0; (addr = server->addr[i]) && (i < UBIK_MAX_INTERFACE_ADDR);
802 rx_NewConnection(addr, ubik_callPortal, DISK_SERVICE_ID,
803 ubikSecClass, ubikSecIndex);
805 /* user requirement to use only the primary interface */
806 if (ubikPrimaryAddrOnly) {
811 assert(i); /* at least one interface address for this server */
815 if (!multi_error) { /* first success */
816 addr = server->addr[multi_i]; /* successful interface addr */
818 if (server->disk_rxcid) /* destroy existing conn */
819 rx_DestroyConnection(server->disk_rxcid);
820 if (server->vote_rxcid)
821 rx_DestroyConnection(server->vote_rxcid);
823 /* make new connections */
824 server->disk_rxcid = conns[multi_i];
825 server->vote_rxcid = rx_NewConnection(addr, ubik_callPortal, VOTE_SERVICE_ID, ubikSecClass, ubikSecIndex); /* for vote reqs */
827 connSuccess = conns[multi_i];
828 strcpy(buffer, (char *)afs_inet_ntoa(server->addr[0]));
830 ("ubik:server %s is back up: will be contacted through %s\n",
831 buffer, afs_inet_ntoa(addr));
837 /* Destroy all connections except the one on which we succeeded */
838 for (j = 0; j < i; j++)
839 if (conns[j] != connSuccess)
840 rx_DestroyConnection(conns[j]);
843 ubik_dprint("ubik:server %s still down\n",
844 afs_inet_ntoa(server->addr[0]));
847 return 0; /* success */
849 return 1; /* failure */