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>
14 #include <sys/types.h>
19 #include <netinet/in.h>
25 #include <afs/afsutil.h>
29 #define UBIK_INTERNALS
35 * The goal is to provide reliable operation among N servers, such that any
36 * server can crash with the remaining servers continuing operation within a
37 * short period of time. While a \b short outage is acceptable, this time
38 * should be order of 3 minutes or less.
40 * Theory of operation:
42 * Note: #SMALLTIME and #BIGTIME are essentially the same time value, separated
43 * only by the clock skew, #MAXSKEW. In general, if you are making guarantees
44 * for someone else, promise them no more than #SMALLTIME seconds of whatever
45 * invariant you provide. If you are waiting to be sure some invariant is now
46 * \b false, wait at least #BIGTIME seconds to be sure that #SMALLTIME seconds
47 * has passed at the other site.
49 * Now, back to the design:
50 * One site in the collection is a special site, designated the \b sync site.
51 * The sync site sends periodic messages, which can be thought of as
52 * keep-alive messages. When a non-sync site hears from the sync site, it
53 * knows that it is getting updates for the next #SMALLTIME seconds from that
56 * If a server does not hear from the sync site in #SMALLTIME seconds, it
57 * determines that it no longer is getting updates, and thus refuses to give
58 * out potentially out-of-date data. If a sync site can not muster a majority
59 * of servers to agree that it is the sync site, then there is a possibility
60 * that a network partition has occurred, allowing another server to claim to
61 * be the sync site. Thus, any time that the sync site has not heard from a
62 * majority of the servers in the last #SMALLTIME seconds, it voluntarily
63 * relinquishes its role as sync site.
65 * While attempting to nominate a new sync site, certain rules apply. First,
66 * a server can not reply "ok" (return 1 from ServBeacon) to two different
67 * hosts in less than #BIGTIME seconds; this allows a server that has heard
68 * affirmative replies from a majority of the servers to know that no other
69 * server in the network has heard enough affirmative replies in the last
70 * #BIGTIME seconds to become sync site, too. The variables #ubik_lastYesTime
71 * and #lastYesHost are used by all servers to keep track of which host they
72 * have last replied affirmatively to, when queried by a potential new sync
75 * Once a sync site has become a sync site, it periodically sends beacon
76 * messages with a parameter of 1, indicating that it already has determined
77 * it is supposed to be the sync site. The servers treat such a message as a
78 * guarantee that no other site will become sync site for the next #SMALLTIME
79 * seconds. In the interim, these servers can answer a query concerning which
80 * site is the sync site without any communication with any server. The
81 * variables #lastBeaconArrival and #lastBeaconHost are used by all servers to
82 * keep track of which sync site has last contacted them.
84 * One complication occurs while nominating a new sync site: each site may be
85 * trying to nominate a different site (based on the value of #lastYesHost),
86 * yet we must nominate the smallest host (under some order), to prevent this
87 * process from looping. The process could loop by having each server give
88 * one vote to another server, but with no server getting a majority of the
89 * votes. To avoid this, we try to withhold our votes for the server with the
90 * lowest internet address (an easy-to-generate order). To this effect, we
91 * keep track (in #lowestTime and #lowestHost) of the lowest server trying to
92 * become a sync site. We wait for this server unless there is already a sync
93 * site (indicated by ServBeacon's parameter being 1).
96 afs_int32 ubik_debugFlag = 0; /*!< print out debugging messages? */
98 /*! \name these statics are used by all sites in nominating new sync sites */
99 afs_int32 ubik_lastYesTime = 0; /*!< time we sent the last \b yes vote */
100 static afs_uint32 lastYesHost = 0xffffffff; /*!< host to which we sent \b yes vote */
102 /*! \name Next is time sync site began this vote: guarantees sync site until this + SMALLTIME */
103 static afs_int32 lastYesClaim = 0;
104 static int lastYesState = 0; /*!< did last site we voted for claim to be sync site? */
107 /*! \name used to guarantee that nomination process doesn't loop */
108 static afs_int32 lowestTime = 0;
109 static afs_uint32 lowestHost = 0xffffffff;
110 static afs_int32 syncTime = 0;
111 static afs_int32 syncHost = 0;
114 /*! \name used to remember which dbase version is the one at the sync site (for non-sync sites) */
115 struct ubik_version ubik_dbVersion; /*!< sync site's dbase version */
116 struct ubik_tid ubik_dbTid; /*!< sync site's tid, or 0 if none */
120 * \brief Decide if we should try to become sync site.
122 * The basic rule is that we
123 * don't run if there is a valid sync site and it ain't us (we have to run if
124 * it is us, in order to keep our votes). If there is no sync site, then we
125 * want to run if we're the lowest numbered host running, otherwise we defer to
126 * the lowest host. However, if the lowest host hasn't been heard from for a
127 * while, then we start running again, in case he crashed.
129 * \return true if we should run, and false otherwise.
132 uvote_ShouldIRun(void)
134 register afs_int32 now;
136 now = FT_ApproxTime();
137 if (BIGTIME + ubik_lastYesTime < now)
138 return 1; /* no valid guy even trying */
139 if (lastYesState && lastYesHost != ubik_host[0])
140 return 0; /* other guy is sync site, leave him alone */
141 if (ntohl((afs_uint32) lastYesHost) < ntohl((afs_uint32) ubik_host[0]))
142 return 0; /* if someone is valid and better than us, don't run */
143 /* otherwise we should run */
148 * \brief Return the current synchronization site, if any.
150 * Simple approach: if the
151 * last guy we voted yes for claims to be the sync site, then we we're happy to
152 * use that guy for a sync site until the time his mandate expires. If the guy
153 * does not claim to be sync site, then, of course, there's none.
155 * In addition, if we lost the sync, we set #urecovery_syncSite to an invalid
156 * value, indicating that we no longer know which version of the dbase is the
157 * one we should have. We'll get a new one when we next hear from the sync
160 * \return 0 or currently valid sync site. It can return our own
161 * address, if we're the sync site.
164 uvote_GetSyncSite(void)
166 register afs_int32 now;
167 register afs_int32 code;
172 now = FT_ApproxTime();
173 if (SMALLTIME + lastYesClaim < now)
174 code = 0; /* last guy timed out */
182 * \brief called by the sync site to handle vote beacons; if aconn is null, this is a
185 * \returns 0 or time when the vote was sent. It returns 0 if we are
186 * not voting for this sync site, or the time we actually voted yes, if
190 SVOTE_Beacon(register struct rx_call * rxcall, afs_int32 astate,
191 afs_int32 astart, struct ubik_version * avers,
192 struct ubik_tid * atid)
194 register afs_int32 otherHost;
195 register afs_int32 now;
197 struct rx_connection *aconn;
199 struct ubik_server *ts;
202 now = FT_ApproxTime(); /* close to current time */
203 if (rxcall) { /* caller's host */
204 aconn = rx_ConnectionOf(rxcall);
205 rxp = rx_PeerOf(aconn);
206 otherHost = rx_HostOf(rxp);
208 /* get the primary interface address for this host. */
209 /* This is the identifier that ubik uses. */
210 otherHost = ubikGetPrimaryInterfaceAddr(otherHost);
212 ubik_dprint("Received beacon from unknown host %s\n",
213 afs_inet_ntoa(rx_HostOf(rxp)));
214 return 0; /* I don't know about you: vote no */
216 for (ts = ubik_servers; ts; ts = ts->next) {
217 if (ts->addr[0] == otherHost)
221 ubik_dprint("Unknown host %x has sent a beacon\n", otherHost);
222 if (ts && ts->isClone)
225 otherHost = ubik_host[0]; /* this host */
229 ubik_dprint("Received beacon type %d from host %s\n", astate,
230 afs_inet_ntoa(otherHost));
232 /* compute the lowest server we've heard from. We'll try to only vote for
233 * this dude if we don't already have a synchronization site. Also, don't
234 * let a very old lowestHost confusing things forever. We pick a new
235 * lowestHost after BIGTIME seconds to limit the damage if this host
236 * actually crashes. Finally, we also count in this computation: don't
237 * pick someone else if we're even better!
239 * Note that the test below must be <=, not <, so that we keep refreshing
240 * lowestTime. Otherwise it will look like we haven't heard from
241 * lowestHost in a while and another host could slip in. */
244 /* First compute the lowest host we've heard from, whether we want them
245 * for a sync site or not. If we haven't heard from a site in BIGTIME
246 * seconds, we ignore its presence in lowestHost: it may have crashed.
247 * Note that we don't ever let anyone appear in our lowestHost if we're
248 * lower than them, 'cause we know we're up. */
249 /* But do not consider clones for lowesHost since they never may become
252 && (ntohl((afs_uint32) otherHost) <= ntohl((afs_uint32) lowestHost)
253 || lowestTime + BIGTIME < now)) {
255 lowestHost = otherHost;
257 /* why do we need this next check? Consider the case where each of two
258 * servers decides the other is lowestHost. Each stops sending beacons
259 * 'cause the other is there. Not obvious that this process terminates:
260 * i.e. each guy could restart procedure and again think other side is
261 * lowest. Need to prove: if one guy in the system is lowest and knows
262 * he's lowest, these loops don't occur. because if someone knows he's
263 * lowest, he will send out beacons telling others to vote for him. */
265 && (ntohl((afs_uint32) ubik_host[0]) <= ntohl((afs_uint32) lowestHost)
266 || lowestTime + BIGTIME < now)) {
268 lowestHost = ubik_host[0];
271 /* tell if we've heard from a sync site recently (even if we're not voting
272 * for this dude yet). After a while, time the guy out. */
273 if (astate) { /* this guy is a sync site */
274 syncHost = otherHost;
276 } else if (syncTime + BIGTIME < now) {
279 ("Ubik: Lost contact with sync-site %s (NOT in quorum)\n",
280 afs_inet_ntoa(syncHost));
285 /* decide how to vote */
286 vote = 0; /* start off voting no */
288 /* if we this guy isn't a sync site, we don't really have to vote for him.
289 * We get to apply some heuristics to try to avoid weird oscillation sates
290 * in the voting procedure. */
292 /* in here only if this guy doesn't claim to be a sync site */
294 /* lowestHost is also trying for our votes, then just say no. */
295 if (ntohl(lowestHost) != ntohl(otherHost)) {
299 /* someone else *is* a sync site, just say no */
300 if (syncHost && syncHost != otherHost)
302 } else /* fast startup if this is the only non-clone */ if (lastYesHost ==
309 for (ts = ubik_servers; ts; ts = ts->next) {
310 if (ts->addr[0] == otherHost)
316 lastYesHost = otherHost;
321 return 0; /* clone never can become sync site */
323 /* Don't promise sync site support to more than one host every BIGTIME
324 * seconds. This is the heart of our invariants in this system. */
325 if (ubik_lastYesTime + BIGTIME < now || otherHost == lastYesHost) {
326 if ((ubik_lastYesTime + BIGTIME < now) || (otherHost != lastYesHost)
327 || (lastYesState != astate)) {
328 /* A new vote or a change in the vote or changed quorum */
329 ubik_dprint("Ubik: vote 'yes' for %s %s\n",
330 afs_inet_ntoa(otherHost),
331 (astate ? "(in quorum)" : "(NOT in quorum)"));
334 vote = now; /* vote yes */
335 ubik_lastYesTime = now; /* remember when we voted yes */
336 lastYesClaim = astart; /* remember for computing when sync site expires */
337 lastYesHost = otherHost; /* and who for */
338 lastYesState = astate; /* remember if site is a sync site */
339 ubik_dbVersion = *avers; /* resync value */
340 ubik_dbTid = *atid; /* transaction id, if any, of active trans */
341 urecovery_CheckTid(atid); /* check if current write trans needs aborted */
347 * \brief Handle per-server debug command, where 0 is the first server.
349 * Basic network debugging hooks.
352 SVOTE_SDebug(struct rx_call * rxcall, afs_int32 awhich,
353 register struct ubik_sdebug * aparm)
355 afs_int32 code, isClone;
356 code = SVOTE_XSDebug(rxcall, awhich, aparm, &isClone);
361 SVOTE_XSDebug(struct rx_call * rxcall, afs_int32 awhich,
362 register struct ubik_sdebug * aparm, afs_int32 * isclone)
364 register struct ubik_server *ts;
366 for (ts = ubik_servers; ts; ts = ts->next) {
369 aparm->addr = ntohl(ts->addr[0]); /* primary interface */
370 for (i = 0; i < UBIK_MAX_INTERFACE_ADDR - 1; i++)
371 aparm->altAddr[i] = ntohl(ts->addr[i + 1]);
372 aparm->lastVoteTime = ts->lastVoteTime;
373 aparm->lastBeaconSent = ts->lastBeaconSent;
374 memcpy(&aparm->remoteVersion, &ts->version,
375 sizeof(struct ubik_version));
376 aparm->lastVote = ts->lastVote;
378 aparm->beaconSinceDown = ts->beaconSinceDown;
379 aparm->currentDB = ts->currentDB;
380 *isclone = ts->isClone;
388 SVOTE_XDebug(struct rx_call * rxcall, register struct ubik_debug * aparm,
393 code = SVOTE_Debug(rxcall, aparm);
399 * \brief Handle basic network debug command. This is the global state dumper.
402 SVOTE_Debug(struct rx_call * rxcall, register struct ubik_debug * aparm)
405 /* fill in the basic debug structure. Note the the RPC protocol transfers,
406 * integers in host order. */
408 aparm->now = FT_ApproxTime();
409 aparm->lastYesTime = ubik_lastYesTime;
410 aparm->lastYesHost = ntohl(lastYesHost);
411 aparm->lastYesState = lastYesState;
412 aparm->lastYesClaim = lastYesClaim;
413 aparm->lowestHost = ntohl(lowestHost);
414 aparm->lowestTime = lowestTime;
415 aparm->syncHost = ntohl(syncHost);
416 aparm->syncTime = syncTime;
418 /* fill in all interface addresses of myself in hostbyte order */
419 for (i = 0; i < UBIK_MAX_INTERFACE_ADDR; i++)
420 aparm->interfaceAddr[i] = ntohl(ubik_host[i]);
422 aparm->amSyncSite = ubik_amSyncSite;
423 ubeacon_Debug(aparm);
429 /* Get the recovery state. The label of the database may not have
430 * been written yet but set the flag so udebug behavior remains.
433 aparm->recoveryState = urecovery_state;
434 if ((urecovery_state & UBIK_RECSYNCSITE)
435 && (urecovery_state & UBIK_RECFOUNDDB)
436 && (urecovery_state & UBIK_RECHAVEDB)) {
437 aparm->recoveryState |= UBIK_RECLABELDB;
439 memcpy(&aparm->syncVersion, &ubik_dbVersion, sizeof(struct ubik_version));
440 memcpy(&aparm->syncTid, &ubik_dbTid, sizeof(struct ubik_tid));
441 aparm->activeWrite = (ubik_dbase->flags & DBWRITING);
442 aparm->tidCounter = ubik_dbase->tidCounter;
444 if (ubik_currentTrans) {
445 aparm->currentTrans = 1;
446 if (ubik_currentTrans->type == UBIK_WRITETRANS)
447 aparm->writeTrans = 1;
449 aparm->writeTrans = 0;
451 aparm->currentTrans = 0;
454 aparm->epochTime = ubik_epochTime;
460 SVOTE_SDebugOld(struct rx_call * rxcall, afs_int32 awhich,
461 register struct ubik_sdebug_old * aparm)
463 register struct ubik_server *ts;
465 for (ts = ubik_servers; ts; ts = ts->next) {
468 aparm->addr = ntohl(ts->addr[0]); /* primary interface */
469 aparm->lastVoteTime = ts->lastVoteTime;
470 aparm->lastBeaconSent = ts->lastBeaconSent;
471 memcpy(&aparm->remoteVersion, &ts->version,
472 sizeof(struct ubik_version));
473 aparm->lastVote = ts->lastVote;
475 aparm->beaconSinceDown = ts->beaconSinceDown;
476 aparm->currentDB = ts->currentDB;
485 * \brief Handle basic network debug command. This is the global state dumper.
488 SVOTE_DebugOld(struct rx_call * rxcall,
489 register struct ubik_debug_old * aparm)
492 /* fill in the basic debug structure. Note the the RPC protocol transfers,
493 * integers in host order. */
495 aparm->now = FT_ApproxTime();
496 aparm->lastYesTime = ubik_lastYesTime;
497 aparm->lastYesHost = ntohl(lastYesHost);
498 aparm->lastYesState = lastYesState;
499 aparm->lastYesClaim = lastYesClaim;
500 aparm->lowestHost = ntohl(lowestHost);
501 aparm->lowestTime = lowestTime;
502 aparm->syncHost = ntohl(syncHost);
503 aparm->syncTime = syncTime;
505 aparm->amSyncSite = ubik_amSyncSite;
506 ubeacon_Debug((ubik_debug *)aparm);
508 udisk_Debug((ubik_debug *)aparm);
510 ulock_Debug((ubik_debug *)aparm);
512 /* Get the recovery state. The label of the database may not have
513 * been written yet but set the flag so udebug behavior remains.
516 aparm->recoveryState = urecovery_state;
517 if ((urecovery_state & UBIK_RECSYNCSITE)
518 && (urecovery_state & UBIK_RECFOUNDDB)
519 && (urecovery_state & UBIK_RECHAVEDB)) {
520 aparm->recoveryState |= UBIK_RECLABELDB;
522 memcpy(&aparm->syncVersion, &ubik_dbVersion, sizeof(struct ubik_version));
523 memcpy(&aparm->syncTid, &ubik_dbTid, sizeof(struct ubik_tid));
524 aparm->activeWrite = (ubik_dbase->flags & DBWRITING);
525 aparm->tidCounter = ubik_dbase->tidCounter;
527 if (ubik_currentTrans) {
528 aparm->currentTrans = 1;
529 if (ubik_currentTrans->type == UBIK_WRITETRANS)
530 aparm->writeTrans = 1;
532 aparm->writeTrans = 0;
534 aparm->currentTrans = 0;
537 aparm->epochTime = ubik_epochTime;
544 * \brief Get the sync site; called by remote servers to find where they should go.
547 SVOTE_GetSyncSite(register struct rx_call * rxcall,
548 register afs_int32 * ahost)
550 register afs_int32 temp;
552 temp = uvote_GetSyncSite();
553 *ahost = ntohl(temp);
558 ubik_dprint_25(const char *format, ...)
562 va_start(ap, format);
563 vViceLog(25, (format, ap));
568 ubik_dprint(const char *format, ...)
572 va_start(ap, format);
573 vViceLog(5, (format, ap));
578 ubik_vprint(const char *format, va_list ap)
580 vViceLog(0, (format, ap));
584 ubik_print(const char *format, ...)
588 va_start(ap, format);
589 ubik_vprint(format, ap);
594 * \brief Called once/run to init the vote module
599 /* pretend we just voted for someone else, since we just restarted */
600 ubik_lastYesTime = FT_ApproxTime();