2 * (C) COPYRIGHT IBM CORPORATION 1987, 1988
3 * LICENSED MATERIALS - PROPERTY OF IBM
13 #include <sys/socket.h>
14 #include <netinet/in.h>
21 #include <rx/rx_multi.h>
23 #include <afs/afsutil.h>
24 #include <afs/netutils.h>
27 #define UBIK_INTERNALS
31 /* statics used to determine if we're the sync site */
32 static afs_int32 syncSiteUntil = 0; /* valid only if amSyncSite */
33 int ubik_amSyncSite = 0; /* flag telling if I'm sync site */
34 static nServers; /* total number of servers */
35 static char amIMagic=0; /* is this host the magic host */
36 extern struct rx_securityClass *rxnull_NewClientSecurityObject();
37 int (*ubik_CRXSecurityProc)();
38 char *ubik_CRXSecurityRock;
39 afs_int32 ubikSecIndex;
40 struct rx_securityClass *ubikSecClass;
43 /* Module responsible for both deciding if we're currently the sync site,
44 * and keeping collecting votes so as to stay sync site.
46 * The basic module contacts all of the servers it can, trying to get them to vote
47 * for this server for sync site. The vote request message (called a beacon message)
48 * also specifies until which time this site claims to be the sync site, if at all, thus enabling
49 * receiving sites to know how long the sync site guarantee is made for.
51 * Each of these beacon messages is thus both a declaration of how long this site will
52 * remain sync site, and an attempt to extend that time by collecting votes for a later
53 * sync site extension.
55 * The voting module is responsible for choosing a reasonable time until which it promises
56 * not to vote for someone else. This parameter (BIG seconds) is not actually passed in
57 * the interface (perhaps it should be?) but is instead a compile time constant that both
60 * The beacon and vote modules work intimately together; the vote module decides how long
61 * it should promise the beacon module its vote, and the beacon module takes all of these
62 * votes and decides for how long it is the synchronization site.
65 /* procedure called from debug rpc call to get this module's state for debugging */
67 register struct ubik_debug *aparm; {
68 /* fill in beacon's state fields in the ubik_debug structure */
69 aparm->syncSiteUntil = syncSiteUntil;
70 aparm->nServers = nServers;
73 /* procedure that determines whether this site has enough current votes to remain sync site.
74 * called from higher-level modules (everything but the vote module).
76 * If we're the sync site, check that our guarantees, obtained by the ubeacon_Interact
77 * light-weight process, haven't expired. We're sync site as long as a majority of the
78 * servers in existence have promised us unexpired guarantees. The variable ubik_syncSiteUntil
79 * contains the time at which the latest of the majority of the sync site guarantees expires
80 * (if the variable ubik_amSyncSite is true)
81 * This module also calls up to the recovery module if it thinks that the recovery module
82 * may have to pick up a new database (which offucr sif we lose the sync site votes).
84 ubeacon_AmSyncSite() {
85 register afs_int32 now;
86 register afs_int32 rcode;
88 /* special case for fast startup */
90 return 1; /* one guy is always the sync site */
93 if (ubik_amSyncSite == 0) rcode = 0; /* if I don't think I'm the sync site, say so */
95 now = FT_ApproxTime();
96 if (syncSiteUntil <= now) { /* if my votes have expired, say so */
97 if (ubik_amSyncSite) ubik_dprint("Ubik: I am no longer the sync site\n");
102 rcode = 1; /* otherwise still have the required votes */
105 if (rcode == 0) urecovery_ResetState(); /* force recovery to re-execute */
106 ubik_dprint("beacon: amSyncSite is %d\n", rcode);
110 /* setup server list; called with two parms, first is my address, second is list of other servers
111 * called only at initialization to set up the list of servers to contact for votes. Just creates
112 * the server structure. Note that there are two connections in every server structure, one for
113 * vote calls (which must always go through quickly) and one for database operations, which
114 * are subject to waiting for locks. If we used only one, the votes would sometimes get
115 * held up behind database operations, and the sync site guarantees would timeout
116 * even though the host would be up for communication.
118 * The "magic" host is the one with the lowest internet address. It is
119 * magic because its vote counts epsilon more than the others. This acts
120 * as a tie-breaker when we have an even number of hosts in the system.
121 * For example, if the "magic" host is up in a 2 site system, then it
122 * is sync site. Without the magic host hack, if anyone crashed in a 2
123 * site system, we'd be out of business.
125 ubeacon_InitServerList(ame, aservers)
127 register afs_int32 aservers[]; {
128 register struct ubik_server *ts;
129 register afs_int32 servAddr;
130 register afs_int32 i, code;
132 struct ubik_server *magicServer;
134 /* verify that the addresses passed in are correct */
135 if ( code = verifyInterfaceAddress(&ame, aservers ))
138 /* get the security index to use, if we can */
139 if (ubik_CRXSecurityProc) {
140 i = (*ubik_CRXSecurityProc)(ubik_CRXSecurityRock, &ubikSecClass, &ubikSecIndex);
144 /* don't have sec module yet */
146 ubikSecClass = rxnull_NewClientSecurityObject();
149 magicHost = ntohl(ame); /* do comparisons in host order */
150 magicServer = (struct ubik_server *) 0;
151 while (servAddr = *aservers++) {
152 if (i >= MAXSERVERS) return UNHOSTS; /* too many hosts */
153 ts = (struct ubik_server *) malloc(sizeof(struct ubik_server));
154 bzero(ts, sizeof(struct ubik_server));
155 ts->next = ubik_servers;
157 ts->addr[0] = servAddr; /* primary address in net byte order */
158 ts->vote_rxcid = rx_NewConnection(servAddr, ubik_callPortal, VOTE_SERVICE_ID, ubikSecClass, ubikSecIndex); /* for vote reqs */
159 ts->disk_rxcid = rx_NewConnection(servAddr, ubik_callPortal, DISK_SERVICE_ID, ubikSecClass, ubikSecIndex); /* for disk reqs */
161 if (ntohl((afs_uint32) servAddr) < (afs_uint32) magicHost) {
162 magicHost = ntohl(servAddr);
167 if (magicServer) magicServer->magic = 1; /* remember for when counting votes */
169 nServers = i+1; /* count this server as well as the remotes */
170 ubik_quorum = (nServers>>1)+1; /* compute the majority figure */
171 /* send addrs to all other servers */
172 code = updateUbikNetworkAddress(ubik_host);
176 /* Shoud we set some defaults for RX??
178 r_nRetries = (RPCTIMEOUT/r_retryInterval);
180 if (nServers == 1) { /* special case 1 server */
181 if (!ubik_amSyncSite) ubik_dprint("Ubik: I am the sync site - 1 server\n");
183 syncSiteUntil = 0x7fffffff; /* quite a while */
188 /* main lwp loop for code that sends out beacons. This code only runs while
189 * we're sync site or we want to be the sync site. It runs in its very own light-weight
193 register afs_int32 code;
195 struct rx_connection *connections[MAXSERVERS];
196 struct ubik_server *servers[MAXSERVERS];
197 register afs_int32 i;
198 register struct ubik_server *ts;
199 afs_int32 temp, yesVotes, lastWakeupTime, oldestYesVote, syncsite;
200 struct ubik_tid ttid;
203 /* loop forever getting votes */
204 lastWakeupTime = 0; /* keep track of time we last started a vote collection */
207 /* don't wakeup more than every POLLTIME seconds */
208 temp = (lastWakeupTime + POLLTIME) - FT_ApproxTime();
209 /* don't sleep if last collection phase took too long (probably timed someone out ) */
211 if (temp > POLLTIME) temp = POLLTIME;
214 code = IOMGR_Select(0, 0, 0, 0, &tt);
218 lastWakeupTime = FT_ApproxTime(); /* started a new collection phase */
220 if (nServers == 1) continue; /* special-case 1 server for speedy startup */
222 if (!uvote_ShouldIRun()) continue; /* if voter has heard from a better candidate than us, don't bother running */
224 /* otherwise we should run for election, or we're the sync site (and have already won);
225 send out the beacon packets */
226 /* build list of all up hosts (noticing dead hosts are running again
227 is a task for the recovery module, not the beacon module), and
228 prepare to send them an r multi-call containing the beacon message */
229 i = 0; /* collect connections */
230 for(ts = ubik_servers; ts; ts=ts->next) {
231 if (ts->up && ts->addr[0] != ubik_host[0]) {
233 connections[i++] = ts->vote_rxcid;
236 servers[i] = (struct ubik_server *) 0; /* end of list */
237 /* note that we assume in the vote module that we'll always get at least BIGTIME
238 seconds of vote from anyone who votes for us, which means we can conservatively
239 assume we'll be fine until SMALLTIME seconds after we start collecting votes */
240 /* this next is essentially an expansion of rgen's ServBeacon routine */
242 ttid.epoch = ubik_epochTime;
243 if (ubik_dbase->flags & DBWRITING) {
245 * if a write is in progress, we have to send the writeTidCounter
246 * which holds the tid counter of the write transaction , and not
247 * send the tidCounter value which holds the tid counter of the
250 ttid.counter = ubik_dbase->writeTidCounter;
253 ttid.counter = ubik_dbase->tidCounter+1;
255 /* now analyze return codes, counting up our votes */
256 yesVotes = 0; /* count how many to ensure we have quorum */
257 oldestYesVote = 0x3fffffff; /* time quorum expires */
258 syncsite= ubeacon_AmSyncSite();
259 startTime = FT_ApproxTime();
261 * Don't waste time using mult Rx calls if there are no connections out there
264 multi_Rx(connections, i) {
265 multi_VOTE_Beacon(syncsite, startTime, &ubik_dbase->version, &ttid);
266 temp = FT_ApproxTime(); /* now, more or less */
267 ts = servers[multi_i];
268 ts->lastBeaconSent = temp;
270 /* note that the vote time (the return code) represents the time
271 the vote was computed, *not* the time the vote expires. We compute
272 the latter down below if we got enough votes to go with */
274 ts->lastVoteTime = code;
275 if (code < oldestYesVote) oldestYesVote = code;
278 if (ts->magic) yesVotes++; /* the extra epsilon */
279 ts->up = 1; /* server is up (not really necessary: recovery does this for real) */
280 ts->beaconSinceDown = 1;
281 ubik_dprint("yes vote from host %s\n",afs_inet_ntoa(ts->addr[0]));
283 else if (code == 0) {
284 ts->lastVoteTime = temp;
286 ts->beaconSinceDown = 1;
287 ubik_dprint("no vote from %s\n", afs_inet_ntoa(ts->addr[0]));
291 ts->beaconSinceDown = 0;
292 urecovery_LostServer();
293 ubik_dprint("time out from %s\n", afs_inet_ntoa(ts->addr[0]));
297 /* now call our own voter module to see if we'll vote for ourself. Note that
298 the same restrictions apply for our voting for ourself as for our voting
300 i = SVOTE_Beacon((struct rx_connection *) 0, ubeacon_AmSyncSite(), startTime, &ubik_dbase->version, &ttid);
303 if (amIMagic) yesVotes++; /* extra epsilon */
304 if (i < oldestYesVote) oldestYesVote = i;
307 /* now decide if we have enough votes to become sync site.
308 Note that we can still get enough votes even if we didn't for ourself. */
309 if (yesVotes > nServers) { /* yesVotes is bumped by 2 or 3 for each site */
310 if (!ubik_amSyncSite) ubik_dprint("Ubik: I am the sync site\n");
312 syncSiteUntil = oldestYesVote + SMALLTIME;
313 LWP_NoYieldSignal(&ubik_amSyncSite);
316 if (ubik_amSyncSite) ubik_dprint("Ubik: I am no longer the sync site\n");
318 urecovery_ResetState(); /* tell recovery we're no longer the sync site */
325 * Input Param : ame is the pointer to my IP address specified in the
326 * CellServDB file. aservers is an array containing IP
327 * addresses of remote ubik servers. The array is
328 * terminated by a zero address.
330 * Algorithm : Verify that my IP addresses 'ame' does actually exist
331 * on this machine. If any of my IP addresses are there
332 * in the remote server list 'aserver', remove them from
333 * this list. Update global variable ubik_host[] with
336 * Return Values : 0 on success, non-zero on failure
338 verifyInterfaceAddress(ame, aservers)
339 afs_uint32 *ame; /* one of my interface addr in net byte order */
340 afs_uint32 aservers[]; /* list of all possible server addresses */
342 afs_uint32 myAddr[UBIK_MAX_INTERFACE_ADDR], *servList;
343 int count, index, found, i, j, totalServers, start, end;
345 /* count the number of servers */
346 for ( totalServers=0, servList = aservers; *servList; servList++)
350 /* for now use getaddr(). use getAllAddr when implemented */
351 myAddr[0] = rxi_getaddr();
352 count = (myAddr[0] != 0);
354 if(AFSDIR_SERVER_NETRESTRICT_FILEPATH || AFSDIR_SERVER_NETINFO_FILEPATH) {
356 * Find addresses we are supposed to register as per the netrestrict file
357 * if it exists, else just register all the addresses we find on this
358 * host as returned by rx_getAllAddr (in NBO)
361 count=parseNetFiles(myAddr,NULL,NULL,UBIK_MAX_INTERFACE_ADDR,
362 reason,AFSDIR_SERVER_NETINFO_FILEPATH,
363 AFSDIR_SERVER_NETRESTRICT_FILEPATH);
365 ubik_print("ubik: Can't register any valid addresses:%s\n",reason);
366 ubik_print("Aborting..\n");
371 /* get all my interface addresses in net byte order */
372 count = rx_getAllAddr(myAddr, UBIK_MAX_INTERFACE_ADDR);
376 if ( count <= 0 ) /* no address found */
378 ubik_print("ubik: No network addresses found, aborting..");
382 /* verify that the My-address passed in by ubik is correct */
383 for ( j=0, found = 0; j < count; j++)
385 if ( *ame == myAddr[j] ) /* both in net byte order */
394 ubik_print("ubik: primary address %s does not exist\n",
395 afs_inet_ntoa(*ame));
399 /* if any of my addresses are there in serverList, then
400 ** use that as my primary addresses : the higher level
401 ** application screwed up in dealing with multihomed concepts
403 for ( j=0, found = 0; j < count; j++)
405 for ( i=0; i < totalServers; i++)
406 if ( myAddr[j] == aservers[i] )
414 ubik_print("Using %s as my primary address\n", afs_inet_ntoa(*ame) );
416 /* get rid of servers which were purged because all
417 ** those interface addresses are myself
419 for ( start=0, end=totalServers-1; (start<end) ; start++, end--)
421 /* find the first zero entry from the beginning */
422 for ( ; (start < end) && ( aservers[start] ); start++);
424 /* find the last non-zero entry from the end */
425 for ( ; (end >= 0) && ( !aservers[end] ); end-- );
427 /* if there is nothing more to purge, exit from loop */
428 if ( start >= end ) break;
431 aservers[start] = aservers[end];
432 aservers[end] = 0; /* this entry was moved */
435 /* update all my addresses in ubik_host in such a way
436 ** that ubik_host[0] has the primary address
439 for ( j=0, i=1; j < count; j++)
440 if ( *ame != myAddr[j] )
441 ubik_host[i++] = myAddr[j];
443 return 0; /* return success */
448 * Input Param : ubik_host is an array containing all my IP addresses.
450 * Algorithm : Do an RPC to all remote ubik servers infroming them
451 * about my IP addresses. Get their IP addresses and
452 * update my linked list of ubik servers 'ubik_servers'
454 * Return Values : 0 on success, non-zero on failure
457 updateUbikNetworkAddress(ubik_host)
458 afs_uint32 ubik_host[UBIK_MAX_INTERFACE_ADDR];
460 int j, count, found, index, code = 0;
461 UbikInterfaceAddr inAddr, outAddr;
462 struct rx_connection *conns[MAXSERVERS];
463 struct ubik_server *ts, *server[MAXSERVERS];
466 for ( count = 0, ts=ubik_servers; ts; count++, ts = ts->next )
468 conns[count] = ts->disk_rxcid;
473 /* inform all other servers only if there are more than one
474 database servers in the cell */
478 for ( j=0; j < UBIK_MAX_INTERFACE_ADDR; j++)
479 inAddr.hostAddr[j] = ntohl(ubik_host[j]);
482 /* do the multi-RX RPC to all other servers */
483 multi_Rx(conns, count) {
484 multi_DISK_UpdateInterfaceAddr(&inAddr, &outAddr);
485 ts = server[multi_i]; /* reply received from this server */
486 if ( !multi_error ) {
487 if ( ts->addr[0] != htonl(outAddr.hostAddr[0]) ) {
489 strcpy(buffer, (char*)afs_inet_ntoa(ts->addr[0]));
490 ubik_print("ubik:Two primary addresses for same server \
491 %s %s\n", buffer, afs_inet_ntoa(htonl(outAddr.hostAddr[0])));
494 for ( j=1; j < UBIK_MAX_INTERFACE_ADDR; j++)
495 ts->addr[j] = htonl(outAddr.hostAddr[j]);
498 else if ( multi_error == RXGEN_OPCODE ) {/* pre 3.5 remote server */
499 ubik_print("ubik server %s does not support UpdateInterfaceAddr RPC\n", afs_inet_ntoa(ts->addr[0]));
501 else if ( multi_error == UBADHOST ) {
502 code = UBADHOST; /* remote CellServDB inconsistency */
503 ubik_print("Inconsistent Cell Info on server: ");
504 for ( j=0; j < UBIK_MAX_INTERFACE_ADDR && ts->addr[j]; j++)
505 printf("%s ", afs_inet_ntoa(ts->addr[j]));
507 fflush(stdout); fflush(stderr);
510 ts->up= 0; /* mark the remote server as down */