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>
23 #include <netinet/in.h>
32 #include <afs/afsutil.h>
34 #define UBIK_INTERNALS
38 /* This module is responsible for determining when the system has
39 * recovered to the point that it can handle new transactions. It
40 * replays logs, polls to determine the current dbase after a crash,
41 * and distributes the new database to the others.
44 /* The sync site associates a version number with each database. It
45 * broadcasts the version associated with its current dbase in every
46 * one of its beacon messages. When the sync site send a dbase to a
47 * server, it also sends the db's version. A non-sync site server can
48 * tell if it has the right dbase version by simply comparing the
49 * version from the beacon message (uvote_dbVersion) with the version
50 * associated with the database (ubik_dbase->version). The sync site
51 * itself simply has one counter to keep track of all of this (again
52 * ubik_dbase->version).
55 /* sync site: routine called when the sync site loses its quorum; this
56 * procedure is called "up" from the beacon package. It resyncs the
57 * dbase and nudges the recovery daemon to try to propagate out the
58 * changes. It also resets the recovery daemon's state, since
59 * recovery must potentially find a new dbase to propagate out. This
60 * routine should not do anything with variables used by non-sync site
64 /* if this flag is set, then ubik will use only the primary address
65 ** ( the address specified in the CellServDB) to contact other
66 ** ubik servers. Ubik recovery will not try opening connections
67 ** to the alternate interface addresses.
69 int ubikPrimaryAddrOnly;
72 urecovery_ResetState(void)
75 LWP_NoYieldSignal(&urecovery_state);
79 /* sync site: routine called when a non-sync site server goes down; restarts recovery
80 * process to send missing server the new db when it comes back up.
81 * This routine should not do anything with variables used by non-sync site servers.
84 urecovery_LostServer(void)
86 LWP_NoYieldSignal(&urecovery_state);
90 /* return true iff we have a current database (called by both sync
91 * sites and non-sync sites) How do we determine this? If we're the
92 * sync site, we wait until recovery has finished fetching and
93 * re-labelling its dbase (it may still be trying to propagate it out
94 * to everyone else; that's THEIR problem). If we're not the sync
95 * site, then we must have a dbase labelled with the right version,
96 * and we must have a currently-good sync site.
99 urecovery_AllBetter(register struct ubik_dbase *adbase, int areadAny)
101 register afs_int32 rcode;
103 ubik_dprint("allbetter checking\n");
108 if (ubik_dbase->version.epoch > 1)
109 rcode = 1; /* Happy with any good version of database */
112 /* Check if we're sync site and we've got the right data */
113 else if (ubeacon_AmSyncSite() && (urecovery_state & UBIK_RECHAVEDB)) {
117 /* next, check if we're aux site, and we've ever been sent the
118 * right data (note that if a dbase update fails, we won't think
119 * that the sync site is still the sync site, 'cause it won't talk
120 * to us until a timeout period has gone by. When we recover, we
121 * leave this clear until we get a new dbase */
122 else if ((uvote_GetSyncSite() && (vcmp(ubik_dbVersion, ubik_dbase->version) == 0))) { /* && order is important */
126 ubik_dprint("allbetter: returning %d\n", rcode);
130 /* abort all transactions on this database */
132 urecovery_AbortAll(struct ubik_dbase *adbase)
134 register struct ubik_trans *tt;
135 for (tt = adbase->activeTrans; tt; tt = tt->next) {
141 /* this routine aborts the current remote transaction, if any, if the tid is wrong */
143 urecovery_CheckTid(register struct ubik_tid *atid)
145 if (ubik_currentTrans) {
146 /* there is remote write trans, see if we match, see if this
147 * is a new transaction */
148 if (atid->epoch != ubik_currentTrans->tid.epoch
149 || atid->counter > ubik_currentTrans->tid.counter) {
150 /* don't match, abort it */
151 /* If the thread is not waiting for lock - ok to end it */
152 #if !defined(UBIK_PAUSE)
153 if (ubik_currentTrans->locktype != LOCKWAIT) {
154 #endif /* UBIK_PAUSE */
155 udisk_end(ubik_currentTrans);
156 #if !defined(UBIK_PAUSE)
158 #endif /* UBIK_PAUSE */
159 ubik_currentTrans = (struct ubik_trans *)0;
165 /* log format is defined here, and implicitly in disk.c
167 * 4 byte opcode, followed by parameters, each 4 bytes long. All integers
168 * are in logged in network standard byte order, in case we want to move logs
169 * from machine-to-machine someday.
171 * Begin transaction: opcode
172 * Commit transaction: opcode, version (8 bytes)
173 * Truncate file: opcode, file number, length
174 * Abort transaction: opcode
175 * Write data: opcode, file, position, length, <length> data bytes
177 * A very simple routine, it just replays the log. Note that this is a new-value only log, which
178 * implies that no uncommitted data is written to the dbase: one writes data to the log, including
179 * the commit record, then we allow data to be written through to the dbase. In our particular
180 * implementation, once a transaction is done, we write out the pages to the database, so that
181 * our buffer package doesn't have to know about stable and uncommitted data in the memory buffers:
182 * any changed data while there is an uncommitted write transaction can be zapped during an
183 * abort and the remaining dbase on the disk is exactly the right dbase, without having to read
189 ReplayLog(register struct ubik_dbase *adbase)
192 register afs_int32 code, tpos;
194 afs_int32 len, thisSize, tfile, filePos;
196 afs_int32 syncFile = -1;
197 afs_int32 data[1024];
199 /* read the lock twice, once to see whether we have a transaction to deal
200 * with that committed, (theoretically, we should support more than one
201 * trans in the log at once, but not yet), and once replaying the
205 /* for now, assume that all ops in log pertain to one transaction; see if there's a commit */
208 (*adbase->read) (adbase, LOGFILE, (char *)&opcode, tpos,
210 if (code != sizeof(afs_int32))
212 if (opcode == LOGNEW) {
213 /* handle begin trans */
214 tpos += sizeof(afs_int32);
215 } else if (opcode == LOGABORT)
217 else if (opcode == LOGEND) {
220 } else if (opcode == LOGTRUNCATE) {
223 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
224 2 * sizeof(afs_int32));
225 if (code != 2 * sizeof(afs_int32))
226 break; /* premature eof or io error */
227 tpos += 2 * sizeof(afs_int32);
228 } else if (opcode == LOGDATA) {
231 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
232 3 * sizeof(afs_int32));
233 if (code != 3 * sizeof(afs_int32))
235 /* otherwise, skip over the data bytes, too */
236 tpos += buffer[2] + 3 * sizeof(afs_int32);
238 ubik_dprint("corrupt log opcode (%d) at position %d\n", opcode,
240 break; /* corrupt log! */
244 /* actually do the replay; log should go all the way through the commit record, since
245 * we just read it above. */
251 (*adbase->read) (adbase, LOGFILE, (char *)&opcode, tpos,
253 if (code != sizeof(afs_int32))
255 if (opcode == LOGNEW) {
256 /* handle begin trans */
257 tpos += sizeof(afs_int32);
258 } else if (opcode == LOGABORT)
259 panic("log abort\n");
260 else if (opcode == LOGEND) {
263 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
264 2 * sizeof(afs_int32));
265 if (code != 2 * sizeof(afs_int32))
267 code = (*adbase->setlabel) (adbase, 0, (ubik_version *)buffer);
271 break; /* all done now */
272 } else if (opcode == LOGTRUNCATE) {
275 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
276 2 * sizeof(afs_int32));
277 if (code != 2 * sizeof(afs_int32))
278 break; /* premature eof or io error */
279 tpos += 2 * sizeof(afs_int32);
281 (*adbase->truncate) (adbase, ntohl(buffer[0]),
285 } else if (opcode == LOGDATA) {
288 (*adbase->read) (adbase, LOGFILE, (char *)buffer, tpos,
289 3 * sizeof(afs_int32));
290 if (code != 3 * sizeof(afs_int32))
292 tpos += 3 * sizeof(afs_int32);
293 /* otherwise, skip over the data bytes, too */
294 len = ntohl(buffer[2]); /* total number of bytes to copy */
295 filePos = ntohl(buffer[1]);
296 tfile = ntohl(buffer[0]);
297 /* try to minimize file syncs */
298 if (syncFile != tfile) {
300 code = (*adbase->sync) (adbase, syncFile);
308 thisSize = (len > sizeof(data) ? sizeof(data) : len);
309 /* copy sizeof(data) buffer bytes at a time */
311 (*adbase->read) (adbase, LOGFILE, (char *)data, tpos,
313 if (code != thisSize)
316 (*adbase->write) (adbase, tfile, (char *)data, filePos,
318 if (code != thisSize)
325 ubik_dprint("corrupt log opcode (%d) at position %d\n",
327 break; /* corrupt log! */
332 code = (*adbase->sync) (adbase, syncFile);
336 ubik_dprint("Log read error on pass 2\n");
341 /* now truncate the log, we're done with it */
342 code = (*adbase->truncate) (adbase, LOGFILE, 0);
346 /* Called at initialization to figure out version of the dbase we really have.
347 * This routine is called after replaying the log; it reads the restored labels.
350 InitializeDB(register struct ubik_dbase *adbase)
352 register afs_int32 code;
354 code = (*adbase->getlabel) (adbase, 0, &adbase->version);
356 /* try setting the label to a new value */
357 adbase->version.epoch = 1; /* value for newly-initialized db */
358 adbase->version.counter = 1;
359 code = (*adbase->setlabel) (adbase, 0, &adbase->version);
361 /* failed, try to set it back */
362 adbase->version.epoch = 0;
363 adbase->version.counter = 0;
364 (*adbase->setlabel) (adbase, 0, &adbase->version);
366 LWP_NoYieldSignal(&adbase->version);
371 /* initialize the local dbase
372 * We replay the logs and then read the resulting file to figure out what version we've really got.
375 urecovery_Initialize(register struct ubik_dbase *adbase)
377 register afs_int32 code;
379 code = ReplayLog(adbase);
382 code = InitializeDB(adbase);
386 /* Main interaction loop for the recovery manager
387 * The recovery light-weight process only runs when you're the
388 * synchronization site. It performs the following tasks, if and only
389 * if the prerequisite tasks have been performed successfully (it
390 * keeps track of which ones have been performed in its bit map,
393 * First, it is responsible for probing that all servers are up. This
394 * is the only operation that must be performed even if this is not
395 * yet the sync site, since otherwise this site may not notice that
396 * enough other machines are running to even elect this guy to be the
399 * After that, the recovery process does nothing until the beacon and
400 * voting modules manage to get this site elected sync site.
402 * After becoming sync site, recovery first attempts to find the best
403 * database available in the network (it must do this in order to
404 * ensure finding the latest committed data). After finding the right
405 * database, it must fetch this dbase to the sync site.
407 * After fetching the dbase, it relabels it with a new version number,
408 * to ensure that everyone recognizes this dbase as the most recent
411 * One the dbase has been relabelled, this machine can start handling
412 * requests. However, the recovery module still has one more task:
413 * propagating the dbase out to everyone who is up in the network.
416 urecovery_Interact(void *dummy)
418 afs_int32 code, tcode;
419 struct ubik_server *bestServer = NULL;
420 struct ubik_server *ts;
421 int dbok, doingRPC, now;
422 afs_int32 lastProbeTime, lastDBVCheck;
423 /* if we're the sync site, the best db version we've found yet */
424 static struct ubik_version bestDBVersion;
425 struct ubik_version tversion;
427 int length, tlen, offset, file, nbytes;
428 struct rx_call *rxcall;
430 struct ubik_stat ubikstat;
431 struct in_addr inAddr;
432 #ifndef OLD_URECOVERY
435 afs_int32 epoch, pass;
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. */
540 urecovery_AbortAll(ubik_dbase);
542 /* Rx code to do the Bulk fetch */
545 rxcall = rx_NewCall(bestServer->disk_rxcid);
547 ubik_print("Ubik: Synchronize database with server %s\n",
548 afs_inet_ntoa(bestServer->addr[0]));
550 code = StartDISK_GetFile(rxcall, file);
552 ubik_dprint("StartDiskGetFile failed=%d\n", code);
555 nbytes = rx_Read(rxcall, (char *)&length, sizeof(afs_int32));
556 length = ntohl(length);
557 if (nbytes != sizeof(afs_int32)) {
558 ubik_dprint("Rx-read length error=%d\n", code = BULK_ERROR);
564 /* Truncate the file first */
565 code = (*ubik_dbase->truncate) (ubik_dbase, file, 0);
567 ubik_dprint("truncate io error=%d\n", code);
570 tversion.counter = 0;
572 /* give invalid label during file transit */
574 code = (*ubik_dbase->setlabel) (ubik_dbase, file, &tversion);
576 ubik_dprint("setlabel io error=%d\n", code);
579 #ifndef OLD_URECOVERY
581 afs_snprintf(pbuffer, sizeof(pbuffer), "%s.DB0.TMP", ubik_dbase->pathName);
582 fd = open(pbuffer, O_CREAT | O_RDWR | O_TRUNC, 0600);
587 code = lseek(fd, HDRSIZE, 0);
588 if (code != HDRSIZE) {
595 tlen = (length > sizeof(tbuffer) ? sizeof(tbuffer) : length);
596 #ifndef AFS_PTHREAD_ENV
600 nbytes = rx_Read(rxcall, tbuffer, tlen);
601 if (nbytes != tlen) {
602 ubik_dprint("Rx-read bulk error=%d\n", code = BULK_ERROR);
609 (*ubik_dbase->write) (ubik_dbase, file, tbuffer, offset,
612 nbytes = write(fd, tbuffer, tlen);
615 if (nbytes != tlen) {
623 #ifndef OLD_URECOVERY
628 code = EndDISK_GetFile(rxcall, &tversion);
630 tcode = rx_EndCall(rxcall, code);
634 /* we got a new file, set up its header */
635 urecovery_state |= UBIK_RECHAVEDB;
636 memcpy(&ubik_dbase->version, &tversion,
637 sizeof(struct ubik_version));
639 (*ubik_dbase->sync) (ubik_dbase, 0); /* get data out first */
641 afs_snprintf(tbuffer, sizeof(tbuffer), "%s.DB0", ubik_dbase->pathName);
643 afs_snprintf(pbuffer, sizeof(pbuffer), "%s.DB0.OLD", ubik_dbase->pathName);
644 code = unlink(pbuffer);
646 code = rename(tbuffer, pbuffer);
647 afs_snprintf(pbuffer, sizeof(pbuffer), "%s.DB0.TMP", ubik_dbase->pathName);
650 code = rename(pbuffer, tbuffer);
653 /* after data is good, sync disk with correct label */
655 (*ubik_dbase->setlabel) (ubik_dbase, 0,
656 &ubik_dbase->version);
657 #ifndef OLD_URECOVERY
659 afs_snprintf(pbuffer, sizeof(pbuffer), "%s.DB0.OLD", ubik_dbase->pathName);
665 #ifndef OLD_URECOVERY
668 * We will effectively invalidate the old data forever now.
669 * Unclear if we *should* but we do.
672 ubik_dbase->version.epoch = 0;
673 ubik_dbase->version.counter = 0;
674 ubik_print("Ubik: Synchronize database failed (error = %d)\n",
677 ubik_print("Ubik: Synchronize database completed\n");
678 urecovery_state |= UBIK_RECHAVEDB;
680 udisk_Invalidate(ubik_dbase, 0); /* data has changed */
681 LWP_NoYieldSignal(&ubik_dbase->version);
684 #if defined(UBIK_PAUSE)
685 if (!(urecovery_state & UBIK_RECSYNCSITE))
686 continue; /* lost sync */
687 #endif /* UBIK_PAUSE */
688 if (!(urecovery_state & UBIK_RECHAVEDB))
689 continue; /* not ready */
691 /* If the database was newly initialized, then when we establish quorum, write
692 * a new label. This allows urecovery_AllBetter() to allow access for reads.
693 * Setting it to 2 also allows another site to come along with a newer
694 * database and overwrite this one.
696 if (ubik_dbase->version.epoch == 1) {
698 urecovery_AbortAll(ubik_dbase);
700 ubik_dbase->version.epoch = ubik_epochTime;
701 ubik_dbase->version.counter = 1;
703 (*ubik_dbase->setlabel) (ubik_dbase, 0, &ubik_dbase->version);
704 udisk_Invalidate(ubik_dbase, 0); /* data may have changed */
705 LWP_NoYieldSignal(&ubik_dbase->version);
709 /* Check the other sites and send the database to them if they
710 * do not have the current db.
712 if (!(urecovery_state & UBIK_RECSENTDB)) {
713 /* now propagate out new version to everyone else */
714 dbok = 1; /* start off assuming they all worked */
718 * Check if a write transaction is in progress. We can't send the
719 * db when a write is in progress here because the db would be
720 * obsolete as soon as it goes there. Also, ops after the begin
721 * trans would reach the recepient and wouldn't find a transaction
722 * pending there. Frankly, I don't think it's possible to get past
723 * the write-lock above if there is a write transaction in progress,
724 * but then, it won't hurt to check, will it?
726 if (ubik_dbase->flags & DBWRITING) {
731 while ((ubik_dbase->flags & DBWRITING) && (safety < 500)) {
733 /* sleep for a little while */
734 IOMGR_Select(0, 0, 0, 0, &tv);
741 for (ts = ubik_servers; ts; ts = ts->next) {
742 inAddr.s_addr = ts->addr[0];
744 ubik_dprint("recovery cannot send version to %s\n",
745 afs_inet_ntoa(inAddr.s_addr));
749 ubik_dprint("recovery sending version to %s\n",
750 afs_inet_ntoa(inAddr.s_addr));
751 if (vcmp(ts->version, ubik_dbase->version) != 0) {
752 ubik_dprint("recovery stating local database\n");
754 /* Rx code to do the Bulk Store */
755 code = (*ubik_dbase->stat) (ubik_dbase, 0, &ubikstat);
757 length = ubikstat.size;
759 rxcall = rx_NewCall(ts->disk_rxcid);
761 StartDISK_SendFile(rxcall, file, length,
762 &ubik_dbase->version);
764 ubik_dprint("StartDiskSendFile failed=%d\n",
771 sizeof(tbuffer) ? sizeof(tbuffer) : length);
773 (*ubik_dbase->read) (ubik_dbase, file,
774 tbuffer, offset, tlen);
775 if (nbytes != tlen) {
776 ubik_dprint("Local disk read error=%d\n",
780 nbytes = rx_Write(rxcall, tbuffer, tlen);
781 if (nbytes != tlen) {
782 ubik_dprint("Rx-write bulk error=%d\n", code =
789 code = EndDISK_SendFile(rxcall);
791 code = rx_EndCall(rxcall, code);
794 /* we set a new file, process its header */
795 ts->version = ubik_dbase->version;
800 /* mark file up to date */
806 urecovery_state |= UBIK_RECSENTDB;
812 ** send a Probe to all the network address of this server
813 ** Return 0 if success, else return 1
816 DoProbe(struct ubik_server *server)
818 struct rx_connection *conns[UBIK_MAX_INTERFACE_ADDR];
819 struct rx_connection *connSuccess = 0;
823 extern afs_int32 ubikSecIndex;
824 extern struct rx_securityClass *ubikSecClass;
826 for (i = 0; (addr = server->addr[i]) && (i < UBIK_MAX_INTERFACE_ADDR);
829 rx_NewConnection(addr, ubik_callPortal, DISK_SERVICE_ID,
830 ubikSecClass, ubikSecIndex);
832 /* user requirement to use only the primary interface */
833 if (ubikPrimaryAddrOnly) {
838 assert(i); /* at least one interface address for this server */
842 if (!multi_error) { /* first success */
843 addr = server->addr[multi_i]; /* successful interface addr */
845 if (server->disk_rxcid) /* destroy existing conn */
846 rx_DestroyConnection(server->disk_rxcid);
847 if (server->vote_rxcid)
848 rx_DestroyConnection(server->vote_rxcid);
850 /* make new connections */
851 server->disk_rxcid = conns[multi_i];
852 server->vote_rxcid = rx_NewConnection(addr, ubik_callPortal, VOTE_SERVICE_ID, ubikSecClass, ubikSecIndex); /* for vote reqs */
854 connSuccess = conns[multi_i];
855 strcpy(buffer, (char *)afs_inet_ntoa(server->addr[0]));
857 ("ubik:server %s is back up: will be contacted through %s\n",
858 buffer, afs_inet_ntoa(addr));
864 /* Destroy all connections except the one on which we succeeded */
865 for (j = 0; j < i; j++)
866 if (conns[j] != connSuccess)
867 rx_DestroyConnection(conns[j]);
870 ubik_dprint("ubik:server %s still down\n",
871 afs_inet_ntoa(server->addr[0]));
874 return 0; /* success */
876 return 1; /* failure */