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>
17 #ifdef AFS_PTHREAD_ENV
18 # include <opr/lock.h>
20 # include <opr/lockstub.h>
25 #include <afs/cellconfig.h>
26 #include <afs/afsutil.h>
29 #define UBIK_INTERNALS
33 #include <lwp.h> /* temporary hack by klm */
35 #define ERROR_EXIT(code) do { \
42 * This system is organized in a hierarchical set of related modules. Modules
43 * at one level can only call modules at the same level or below.
45 * At the bottom level (0) we have R, RFTP, LWP and IOMGR, i.e. the basic
46 * operating system primitives.
48 * At the next level (1) we have
50 * \li VOTER--The module responsible for casting votes when asked. It is also
51 * responsible for determining whether this server should try to become
52 * a synchronization site.
53 * \li BEACONER--The module responsible for sending keep-alives out when a
54 * server is actually the sync site, or trying to become a sync site.
55 * \li DISK--The module responsible for representing atomic transactions
56 * on the local disk. It maintains a new-value only log.
57 * \li LOCK--The module responsible for locking byte ranges in the database file.
59 * At the next level (2) we have
61 * \li RECOVERY--The module responsible for ensuring that all members of a quorum
62 * have the same up-to-date database after a new synchronization site is
63 * elected. This module runs only on the synchronization site.
65 * At the next level (3) we have
67 * \li REMOTE--The module responsible for interpreting requests from the sync
68 * site and applying them to the database, after obtaining the appropriate
71 * At the next level (4) we have
73 * \li UBIK--The module users call to perform operations on the database.
78 afs_int32 ubik_quorum = 0;
79 struct ubik_dbase *ubik_dbase = 0;
80 struct ubik_stats ubik_stats;
81 afs_uint32 ubik_host[UBIK_MAX_INTERFACE_ADDR];
82 afs_int32 urecovery_state = 0;
83 int (*ubik_SyncWriterCacheProc) (void);
84 struct ubik_server *ubik_servers;
85 short ubik_callPortal;
87 /* These global variables were used to control the server security layers.
88 * They are retained for backwards compatibility with legacy callers.
90 * The ubik_SetServerSecurityProcs() interface should be used instead.
93 int (*ubik_SRXSecurityProc) (void *, struct rx_securityClass **, afs_int32 *);
94 void *ubik_SRXSecurityRock;
95 int (*ubik_CheckRXSecurityProc) (void *, struct rx_call *);
96 void *ubik_CheckRXSecurityRock;
100 static int BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
101 struct ubik_trans **transPtr, int readAny);
103 static struct rx_securityClass **ubik_sc = NULL;
104 static void (*buildSecClassesProc)(void *, struct rx_securityClass ***,
106 static int (*checkSecurityProc)(void *, struct rx_call *) = NULL;
107 static void *securityRock = NULL;
109 struct version_data version_globals;
111 #define CStampVersion 1 /* meaning set ts->version */
112 #define CCheckSyncAdvertised 2 /* check if the remote knows we are the sync-site */
114 static_inline struct rx_connection *
115 Quorum_StartIO(struct ubik_trans *atrans, struct ubik_server *as)
117 struct rx_connection *conn;
120 conn = as->disk_rxcid;
122 #ifdef AFS_PTHREAD_ENV
123 rx_GetConnection(conn);
125 DBRELE(atrans->dbase);
128 #endif /* AFS_PTHREAD_ENV */
134 Quorum_EndIO(struct ubik_trans *atrans, struct rx_connection *aconn)
136 #ifdef AFS_PTHREAD_ENV
137 DBHOLD(atrans->dbase);
138 rx_PutConnection(aconn);
139 #endif /* AFS_PTHREAD_ENV */
144 * Iterate over all servers. Callers pass in *ts which is used to track
145 * the current server.
146 * - Returns 1 if there are no more servers
147 * - Returns 0 with conn set to the connection for the current server if
148 * it's up and current
151 ContactQuorum_iterate(struct ubik_trans *atrans, int aflags, struct ubik_server **ts,
152 struct rx_connection **conn, afs_int32 *rcode,
153 afs_int32 *okcalls, afs_int32 code, const char *procname)
156 /* Initial call - start iterating over servers */
163 Quorum_EndIO(atrans, *conn);
165 if (code) { /* failure */
170 (*ts)->up = 0; /* mark as down now; beacons will no longer be sent */
171 (*ts)->beaconSinceDown = 0;
173 (*ts)->currentDB = 0;
174 urecovery_LostServer(*ts); /* tell recovery to try to resend dbase later */
175 ViceLog(0, ("Server %s is marked down due to %s code %d\n",
176 afs_inet_ntoa_r((*ts)->addr[0], hoststr), procname, *rcode));
177 } else { /* success */
179 (*okcalls)++; /* count up how many worked */
180 if (aflags & CStampVersion) {
181 (*ts)->version = atrans->dbase->version;
190 if (!(*ts)->up || !(*ts)->currentDB ||
191 /* do not call DISK_Begin until we know that lastYesState is set on the
192 * remote in question; otherwise, DISK_Begin will fail. */
193 ((aflags & CCheckSyncAdvertised) && !((*ts)->beaconSinceDown && (*ts)->lastVote))) {
195 (*ts)->currentDB = 0; /* db is no longer current; we just missed an update */
196 return 0; /* not up-to-date, don't bother. NULL conn will tell caller not to use */
199 *conn = Quorum_StartIO(atrans, *ts);
204 ContactQuorum_rcode(int okcalls, afs_int32 rcode)
207 * return 0 if we successfully contacted a quorum, otherwise return error code.
208 * We don't have to contact ourselves (that was done locally)
210 if (okcalls + 1 >= ubik_quorum)
213 return (rcode != 0) ? rcode : UNOQUORUM;
217 * \brief Perform an operation at a quorum, handling error conditions.
218 * \return 0 if all worked and a quorum was contacted successfully
219 * \return otherwise mark failing server as down and return #UERROR
221 * \note If any server misses an update, we must wait #BIGTIME seconds before
222 * allowing the transaction to commit, to ensure that the missing and
223 * possibly still functioning server times out and stops handing out old
224 * data. This is done in the commit code, where we wait for a server marked
225 * down to have stayed down for #BIGTIME seconds before we allow a transaction
226 * to commit. A server that fails but comes back up won't give out old data
227 * because it is sent the sync count along with the beacon message that
228 * marks it as \b really up (\p beaconSinceDown).
231 ContactQuorum_NoArguments(afs_int32 (*proc)(struct rx_connection *, ubik_tid *),
232 struct ubik_trans *atrans, int aflags, const char *procname)
234 struct ubik_server *ts = NULL;
235 afs_int32 code = 0, rcode, okcalls;
236 struct rx_connection *conn;
239 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
242 code = (*proc)(conn, &atrans->tid);
243 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
245 return ContactQuorum_rcode(okcalls, rcode);
250 ContactQuorum_DISK_Lock(struct ubik_trans *atrans, int aflags,afs_int32 file,
251 afs_int32 position, afs_int32 length, afs_int32 type)
253 struct ubik_server *ts = NULL;
254 afs_int32 code = 0, rcode, okcalls;
255 struct rx_connection *conn;
257 char *procname = "DISK_Lock";
259 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
262 code = DISK_Lock(conn, &atrans->tid, file, position, length, type);
263 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
265 return ContactQuorum_rcode(okcalls, rcode);
269 ContactQuorum_DISK_Truncate(struct ubik_trans *atrans, int aflags,
270 afs_int32 file, afs_int32 length)
272 struct ubik_server *ts = NULL;
273 afs_int32 code = 0, rcode, okcalls;
274 struct rx_connection *conn;
276 char *procname = "DISK_Truncate";
278 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
281 code = DISK_Truncate(conn, &atrans->tid, file, length);
282 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
284 return ContactQuorum_rcode(okcalls, rcode);
289 ContactQuorum_DISK_WriteV(struct ubik_trans *atrans, int aflags,
290 iovec_wrt * io_vector, iovec_buf *io_buffer)
292 struct ubik_server *ts = NULL;
293 afs_int32 code = 0, rcode, okcalls;
294 struct rx_connection *conn;
296 char *procname = "DISK_WriteV";
298 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
301 procname = "DISK_WriteV"; /* in case previous fallback to DISK_Write */
302 code = DISK_WriteV(conn, &atrans->tid, io_vector, io_buffer);
303 if ((code <= -450) && (code > -500)) {
304 /* An RPC interface mismatch (as defined in comerr/error_msg.c).
305 * Un-bulk the entries and do individual DISK_Write calls
306 * instead of DISK_WriteV.
308 struct ubik_iovec *iovec =
309 (struct ubik_iovec *)io_vector->iovec_wrt_val;
310 char *iobuf = (char *)io_buffer->iovec_buf_val;
314 procname = "DISK_Write"; /* for accurate error msg, if any */
315 for (i = 0, offset = 0; i < io_vector->iovec_wrt_len; i++) {
316 /* Sanity check for going off end of buffer */
317 if ((offset + iovec[i].length) > io_buffer->iovec_buf_len) {
321 tcbs.bulkdata_len = iovec[i].length;
322 tcbs.bulkdata_val = &iobuf[offset];
323 code = DISK_Write(conn, &atrans->tid, iovec[i].file,
324 iovec[i].position, &tcbs);
327 offset += iovec[i].length;
331 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
333 return ContactQuorum_rcode(okcalls, rcode);
338 ContactQuorum_DISK_SetVersion(struct ubik_trans *atrans, int aflags,
339 ubik_version *OldVersion,
340 ubik_version *NewVersion)
342 struct ubik_server *ts = NULL;
343 afs_int32 code = 0, rcode, okcalls;
344 struct rx_connection *conn;
346 char *procname = "DISK_SetVersion";
348 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
351 code = DISK_SetVersion(conn, &atrans->tid, OldVersion, NewVersion);
352 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code, procname);
354 return ContactQuorum_rcode(okcalls, rcode);
357 #if defined(AFS_PTHREAD_ENV)
359 ubik_thread_create(pthread_attr_t *tattr, pthread_t *thread, void *proc) {
360 opr_Verify(pthread_attr_init(tattr) == 0);
361 opr_Verify(pthread_attr_setdetachstate(tattr,
362 PTHREAD_CREATE_DETACHED) == 0);
363 opr_Verify(pthread_create(thread, tattr, proc, NULL) == 0);
369 * \brief This routine initializes the ubik system for a set of servers.
370 * \return 0 for success, or an error code on failure.
371 * \param serverList set of servers specified; nServers gives the number of entries in this array.
372 * \param pathName provides an initial prefix used for naming storage files used by this system.
373 * \param dbase the returned structure representing this instance of an ubik; it is passed to various calls below.
375 * \todo This routine should perhaps be generalized to a low-level disk interface providing read, write, file enumeration and sync operations.
377 * \warning The host named by myHost should not also be listed in serverList.
379 * \see ubik_ServerInit(), ubik_ServerInitByInfo()
382 ubik_ServerInitCommon(afs_uint32 myHost, short myPort,
383 struct afsconf_cell *info, char clones[],
384 afs_uint32 serverList[], const char *pathName,
385 struct ubik_dbase **dbase)
387 struct ubik_dbase *tdb;
389 #ifdef AFS_PTHREAD_ENV
390 pthread_t rxServerThread; /* pthread variables */
391 pthread_t ubeacon_InteractThread;
392 pthread_t urecovery_InteractThread;
393 pthread_attr_t rxServer_tattr;
394 pthread_attr_t ubeacon_Interact_tattr;
395 pthread_attr_t urecovery_Interact_tattr;
398 extern int rx_stackSize;
402 struct rx_securityClass *secClass;
405 struct rx_service *tservice;
407 initialize_U_error_table();
409 tdb = malloc(sizeof(struct ubik_dbase));
410 tdb->pathName = strdup(pathName);
411 tdb->activeTrans = (struct ubik_trans *)0;
412 memset(&tdb->version, 0, sizeof(struct ubik_version));
413 memset(&tdb->cachedVersion, 0, sizeof(struct ubik_version));
414 #ifdef AFS_PTHREAD_ENV
415 opr_mutex_init(&tdb->versionLock);
416 opr_mutex_init(&beacon_globals.beacon_lock);
417 opr_mutex_init(&vote_globals.vote_lock);
418 opr_mutex_init(&addr_globals.addr_lock);
419 opr_mutex_init(&version_globals.version_lock);
421 Lock_Init(&tdb->versionLock);
423 Lock_Init(&tdb->cache_lock);
425 tdb->read = uphys_read;
426 tdb->write = uphys_write;
427 tdb->truncate = uphys_truncate;
428 tdb->open = uphys_invalidate; /* this function isn't used any more */
429 tdb->sync = uphys_sync;
430 tdb->stat = uphys_stat;
431 tdb->getlabel = uphys_getlabel;
432 tdb->setlabel = uphys_setlabel;
433 tdb->getnfiles = uphys_getnfiles;
435 tdb->tidCounter = tdb->writeTidCounter = 0;
437 ubik_dbase = tdb; /* for now, only one db per server; can fix later when we have names for the other dbases */
439 #ifdef AFS_PTHREAD_ENV
440 opr_cv_init(&tdb->version_cond);
441 opr_cv_init(&tdb->flags_cond);
442 #endif /* AFS_PTHREAD_ENV */
446 /* the following call is idempotent so when/if it got called earlier,
447 * by whatever called us, it doesn't really matter -- klm */
448 code = rx_Init(myPort);
452 ubik_callPortal = myPort;
454 udisk_Init(ubik_nBuffers);
460 code = urecovery_Initialize(tdb);
464 code = ubeacon_InitServerListByInfo(myHost, info, clones);
466 code = ubeacon_InitServerList(myHost, serverList);
470 /* try to get an additional security object */
471 if (buildSecClassesProc == NULL) {
473 ubik_sc = calloc(numClasses, sizeof(struct rx_securityClass *));
474 ubik_sc[0] = rxnull_NewServerSecurityObject();
475 if (ubik_SRXSecurityProc) {
476 code = (*ubik_SRXSecurityProc) (ubik_SRXSecurityRock,
480 ubik_sc[secIndex] = secClass;
484 (*buildSecClassesProc) (securityRock, &ubik_sc, &numClasses);
486 /* for backwards compat this should keep working as it does now
490 rx_NewService(0, VOTE_SERVICE_ID, "VOTE", ubik_sc, numClasses,
491 VOTE_ExecuteRequest);
492 if (tservice == (struct rx_service *)0) {
493 ViceLog(0, ("Could not create VOTE rx service!\n"));
496 rx_SetMinProcs(tservice, 2);
497 rx_SetMaxProcs(tservice, 3);
500 rx_NewService(0, DISK_SERVICE_ID, "DISK", ubik_sc, numClasses,
501 DISK_ExecuteRequest);
502 if (tservice == (struct rx_service *)0) {
503 ViceLog(0, ("Could not create DISK rx service!\n"));
506 rx_SetMinProcs(tservice, 2);
507 rx_SetMaxProcs(tservice, 3);
509 /* start an rx_ServerProc to handle incoming RPC's in particular the
510 * UpdateInterfaceAddr RPC that occurs in ubeacon_InitServerList. This avoids
511 * the "steplock" problem in ubik initialization. Defect 11037.
513 #ifdef AFS_PTHREAD_ENV
514 ubik_thread_create(&rxServer_tattr, &rxServerThread, (void *)rx_ServerProc);
516 LWP_CreateProcess(rx_ServerProc, rx_stackSize, RX_PROCESS_PRIORITY,
517 NULL, "rx_ServerProc", &junk);
520 /* send addrs to all other servers */
521 code = ubeacon_updateUbikNetworkAddress(ubik_host);
525 /* now start up async processes */
526 #ifdef AFS_PTHREAD_ENV
527 ubik_thread_create(&ubeacon_Interact_tattr, &ubeacon_InteractThread,
528 (void *)ubeacon_Interact);
530 code = LWP_CreateProcess(ubeacon_Interact, 16384 /*8192 */ ,
531 LWP_MAX_PRIORITY - 1, (void *)0, "beacon",
537 #ifdef AFS_PTHREAD_ENV
538 ubik_thread_create(&urecovery_Interact_tattr, &urecovery_InteractThread,
539 (void *)urecovery_Interact);
540 return 0; /* is this correct? - klm */
542 code = LWP_CreateProcess(urecovery_Interact, 16384 /*8192 */ ,
543 LWP_MAX_PRIORITY - 1, (void *)0, "recovery",
551 * \see ubik_ServerInitCommon()
554 ubik_ServerInitByInfo(afs_uint32 myHost, short myPort,
555 struct afsconf_cell *info, char clones[],
556 const char *pathName, struct ubik_dbase **dbase)
561 ubik_ServerInitCommon(myHost, myPort, info, clones, 0, pathName,
567 * \see ubik_ServerInitCommon()
570 ubik_ServerInit(afs_uint32 myHost, short myPort, afs_uint32 serverList[],
571 const char *pathName, struct ubik_dbase **dbase)
576 ubik_ServerInitCommon(myHost, myPort, (struct afsconf_cell *)0, 0,
577 serverList, pathName, dbase);
582 * \brief This routine begins a read or write transaction on the transaction
583 * identified by transPtr, in the dbase named by dbase.
585 * An open mode of ubik_READTRANS identifies this as a read transaction,
586 * while a mode of ubik_WRITETRANS identifies this as a write transaction.
587 * transPtr is set to the returned transaction control block.
588 * The readAny flag is set to 0 or 1 or 2 by the wrapper functions
589 * ubik_BeginTrans() or ubik_BeginTransReadAny() or
590 * ubik_BeginTransReadAnyWrite() below.
592 * \note We can only begin transaction when we have an up-to-date database.
595 BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
596 struct ubik_trans **transPtr, int readAny)
598 struct ubik_trans *jt;
599 struct ubik_trans *tt;
602 if (readAny > 1 && ubik_SyncWriterCacheProc == NULL) {
603 /* it's not safe to use ubik_BeginTransReadAnyWrite without a
604 * cache-syncing function; fall back to ubik_BeginTransReadAny,
605 * which is safe but slower */
606 ViceLog(0, ("ubik_BeginTransReadAnyWrite called, but "
607 "ubik_SyncWriterCacheProc not set; pretending "
608 "ubik_BeginTransReadAny was called instead\n"));
612 if ((transMode != UBIK_READTRANS) && readAny)
615 if (urecovery_AllBetter(dbase, readAny) == 0) {
619 /* otherwise we have a quorum, use it */
621 /* make sure that at most one write transaction occurs at any one time. This
622 * has nothing to do with transaction locking; that's enforced by the lock package. However,
623 * we can't even handle two non-conflicting writes, since our log and recovery modules
624 * don't know how to restore one without possibly picking up some data from the other. */
625 if (transMode == UBIK_WRITETRANS) {
626 /* if we're writing already, wait */
627 while (dbase->flags & DBWRITING) {
628 #ifdef AFS_PTHREAD_ENV
629 opr_cv_wait(&dbase->flags_cond, &dbase->versionLock);
632 LWP_WaitProcess(&dbase->flags);
637 if (!ubeacon_AmSyncSite()) {
641 if (!ubeacon_SyncSiteAdvertised()) {
642 /* i am the sync-site but the remotes are not aware yet */
648 /* create the transaction */
649 code = udisk_begin(dbase, transMode, &jt); /* can't take address of register var */
650 tt = jt; /* move to a register */
651 if (code || tt == NULL) {
657 tt->flags |= TRREADANY;
659 tt->flags |= TRREADWRITE;
662 /* label trans and dbase with new tid */
663 tt->tid.epoch = version_globals.ubik_epochTime;
664 /* bump by two, since tidCounter+1 means trans id'd by tidCounter has finished */
665 tt->tid.counter = (dbase->tidCounter += 2);
667 if (transMode == UBIK_WRITETRANS) {
668 /* for a write trans, we have to keep track of the write tid counter too */
669 dbase->writeTidCounter = tt->tid.counter;
674 if (transMode == UBIK_WRITETRANS) {
675 /* next try to start transaction on appropriate number of machines */
676 code = ContactQuorum_NoArguments(DISK_Begin, tt, CCheckSyncAdvertised, "DISK_Begin");
678 /* we must abort the operation */
680 /* force aborts to the others */
681 ContactQuorum_NoArguments(DISK_Abort, tt, 0, "DISK_Abort");
697 ubik_BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
698 struct ubik_trans **transPtr)
700 return BeginTrans(dbase, transMode, transPtr, 0);
707 ubik_BeginTransReadAny(struct ubik_dbase *dbase, afs_int32 transMode,
708 struct ubik_trans **transPtr)
710 return BeginTrans(dbase, transMode, transPtr, 1);
717 ubik_BeginTransReadAnyWrite(struct ubik_dbase *dbase, afs_int32 transMode,
718 struct ubik_trans **transPtr)
720 return BeginTrans(dbase, transMode, transPtr, 2);
724 * \brief This routine ends a read or write transaction by aborting it.
727 ubik_AbortTrans(struct ubik_trans *transPtr)
731 struct ubik_dbase *dbase;
733 dbase = transPtr->dbase;
735 if (transPtr->flags & TRCACHELOCKED) {
736 ReleaseReadLock(&dbase->cache_lock);
737 transPtr->flags &= ~TRCACHELOCKED;
740 ObtainWriteLock(&dbase->cache_lock);
743 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
745 ReleaseWriteLock(&dbase->cache_lock);
747 /* see if we're still up-to-date */
748 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
749 udisk_abort(transPtr);
755 if (transPtr->type == UBIK_READTRANS) {
756 code = udisk_abort(transPtr);
762 /* below here, we know we're doing a write transaction */
763 if (!ubeacon_AmSyncSite()) {
764 udisk_abort(transPtr);
770 /* now it is safe to try remote abort */
771 code = ContactQuorum_NoArguments(DISK_Abort, transPtr, 0, "DISK_Abort");
772 code2 = udisk_abort(transPtr);
775 return (code ? code : code2);
779 WritebackApplicationCache(struct ubik_dbase *dbase)
782 if (ubik_SyncWriterCacheProc) {
783 code = ubik_SyncWriterCacheProc();
786 /* we failed to sync the local cache, so just invalidate the cache;
787 * we'll try to read the cache in again on the next read */
788 memset(&dbase->cachedVersion, 0, sizeof(dbase->cachedVersion));
790 memcpy(&dbase->cachedVersion, &dbase->version,
791 sizeof(dbase->cachedVersion));
796 * \brief This routine ends a read or write transaction on the open transaction identified by transPtr.
797 * \return an error code.
800 ubik_EndTrans(struct ubik_trans *transPtr)
805 struct ubik_server *ts;
808 struct ubik_dbase *dbase;
810 if (transPtr->type == UBIK_WRITETRANS) {
811 code = ubik_Flush(transPtr);
813 ubik_AbortTrans(transPtr);
818 dbase = transPtr->dbase;
820 if (transPtr->flags & TRCACHELOCKED) {
821 ReleaseReadLock(&dbase->cache_lock);
822 transPtr->flags &= ~TRCACHELOCKED;
825 if (transPtr->type != UBIK_READTRANS) {
826 /* must hold cache_lock before DBHOLD'ing */
827 ObtainWriteLock(&dbase->cache_lock);
833 /* give up if no longer current */
834 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
835 udisk_abort(transPtr);
842 if (transPtr->type == UBIK_READTRANS) { /* reads are easy */
843 code = udisk_commit(transPtr);
845 goto success; /* update cachedVersion correctly */
851 if (!ubeacon_AmSyncSite()) { /* no longer sync site */
852 udisk_abort(transPtr);
859 /* now it is safe to do commit */
860 code = udisk_commit(transPtr);
862 /* db data has been committed locally; update the local cache so
863 * readers can get at it */
864 WritebackApplicationCache(dbase);
866 ReleaseWriteLock(&dbase->cache_lock);
868 code = ContactQuorum_NoArguments(DISK_Commit, transPtr, CStampVersion, "DISK_Commit");
871 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
872 ReleaseWriteLock(&dbase->cache_lock);
876 /* failed to commit, so must return failure. Try to clear locks first, just for fun
877 * Note that we don't know if this transaction will eventually commit at this point.
878 * If it made it to a site that will be present in the next quorum, we win, otherwise
879 * we lose. If we contact a majority of sites, then we won't be here: contacting
880 * a majority guarantees commit, since it guarantees that one dude will be a
881 * member of the next quorum. */
882 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0, "DISK_ReleaseLocks");
887 /* before we can start sending unlock messages, we must wait until all servers
888 * that are possibly still functioning on the other side of a network partition
889 * have timed out. Check the server structures, compute how long to wait, then
890 * start the unlocks */
891 realStart = FT_ApproxTime();
893 /* wait for all servers to time out */
895 now = FT_ApproxTime();
896 /* check if we're still sync site, the guy should either come up
897 * to us, or timeout. Put safety check in anyway */
898 if (now - realStart > 10 * BIGTIME) {
899 ubik_stats.escapes++;
900 ViceLog(0, ("ubik escaping from commit wait\n"));
903 for (ts = ubik_servers; ts; ts = ts->next) {
905 if (!ts->beaconSinceDown && now <= ts->lastBeaconSent + BIGTIME) {
908 /* this guy could have some damaged data, wait for him */
910 tv.tv_sec = 1; /* try again after a while (ha ha) */
913 #ifdef AFS_PTHREAD_ENV
914 /* we could release the dbase outside of the loop, but we do
915 * it here, in the loop, to avoid an unnecessary RELE/HOLD
916 * if all sites are up */
918 select(0, 0, 0, 0, &tv);
921 IOMGR_Select(0, 0, 0, 0, &tv); /* poll, should we wait on something? */
929 break; /* no down ones still pseudo-active */
932 /* finally, unlock all the dudes. We can return success independent of the number of servers
933 * that really unlock the dbase; the others will do it if/when they elect a new sync site.
934 * The transaction is committed anyway, since we succeeded in contacting a quorum
935 * at the start (when invoking the DiskCommit function).
937 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0, "DISK_ReleaseLocks");
941 /* don't update cachedVersion here; it should have been updated way back
942 * in ubik_CheckCache, and earlier in this function for writes */
945 ReleaseWriteLock(&dbase->cache_lock);
951 ObtainWriteLock(&dbase->cache_lock);
953 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
954 ReleaseWriteLock(&dbase->cache_lock);
959 * \brief This routine reads length bytes into buffer from the current position in the database.
961 * The file pointer is updated appropriately (by adding the number of bytes actually transferred), and the length actually transferred is stored in the long integer pointed to by length. A short read returns zero for an error code.
963 * \note *length is an INOUT parameter: at the start it represents the size of the buffer, and when done, it contains the number of bytes actually transferred.
966 ubik_Read(struct ubik_trans *transPtr, void *buffer,
971 /* reads are easy to do: handle locally */
972 DBHOLD(transPtr->dbase);
973 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
974 DBRELE(transPtr->dbase);
979 udisk_read(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
982 transPtr->seekPos += length;
984 DBRELE(transPtr->dbase);
989 * \brief This routine will flush the io data in the iovec structures.
991 * It first flushes to the local disk and then uses ContactQuorum to write it
992 * to the other servers.
995 ubik_Flush(struct ubik_trans *transPtr)
997 afs_int32 code, error = 0;
999 if (transPtr->type != UBIK_WRITETRANS)
1002 DBHOLD(transPtr->dbase);
1003 if (!transPtr->iovec_info.iovec_wrt_len
1004 || !transPtr->iovec_info.iovec_wrt_val) {
1005 DBRELE(transPtr->dbase);
1009 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1010 ERROR_EXIT(UNOQUORUM);
1011 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1012 ERROR_EXIT(UNOTSYNC);
1014 /* Update the rest of the servers in the quorum */
1016 ContactQuorum_DISK_WriteV(transPtr, 0, &transPtr->iovec_info,
1017 &transPtr->iovec_data);
1019 udisk_abort(transPtr);
1020 /* force aborts to the others */
1021 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0, "DISK_Abort");
1022 transPtr->iovec_info.iovec_wrt_len = 0;
1023 transPtr->iovec_data.iovec_buf_len = 0;
1027 /* Wrote the buffers out, so start at scratch again */
1028 transPtr->iovec_info.iovec_wrt_len = 0;
1029 transPtr->iovec_data.iovec_buf_len = 0;
1032 DBRELE(transPtr->dbase);
1037 ubik_Write(struct ubik_trans *transPtr, void *vbuffer,
1040 struct ubik_iovec *iovec;
1041 afs_int32 code, error = 0;
1042 afs_int32 pos, len, size;
1043 char * buffer = (char *)vbuffer;
1045 if (transPtr->type != UBIK_WRITETRANS)
1050 if (length > IOVEC_MAXBUF) {
1051 for (pos = 0, len = length; len > 0; len -= size, pos += size) {
1052 size = ((len < IOVEC_MAXBUF) ? len : IOVEC_MAXBUF);
1053 code = ubik_Write(transPtr, buffer+pos, size);
1060 DBHOLD(transPtr->dbase);
1061 if (!transPtr->iovec_info.iovec_wrt_val) {
1062 transPtr->iovec_info.iovec_wrt_len = 0;
1063 transPtr->iovec_info.iovec_wrt_val =
1064 malloc(IOVEC_MAXWRT * sizeof(struct ubik_iovec));
1065 transPtr->iovec_data.iovec_buf_len = 0;
1066 transPtr->iovec_data.iovec_buf_val = malloc(IOVEC_MAXBUF);
1067 if (!transPtr->iovec_info.iovec_wrt_val
1068 || !transPtr->iovec_data.iovec_buf_val) {
1069 if (transPtr->iovec_info.iovec_wrt_val)
1070 free(transPtr->iovec_info.iovec_wrt_val);
1071 transPtr->iovec_info.iovec_wrt_val = 0;
1072 if (transPtr->iovec_data.iovec_buf_val)
1073 free(transPtr->iovec_data.iovec_buf_val);
1074 transPtr->iovec_data.iovec_buf_val = 0;
1075 DBRELE(transPtr->dbase);
1080 /* If this write won't fit in the structure, then flush it out and start anew */
1081 if ((transPtr->iovec_info.iovec_wrt_len >= IOVEC_MAXWRT)
1082 || ((length + transPtr->iovec_data.iovec_buf_len) > IOVEC_MAXBUF)) {
1083 /* Can't hold the DB lock over ubik_Flush */
1084 DBRELE(transPtr->dbase);
1085 code = ubik_Flush(transPtr);
1088 DBHOLD(transPtr->dbase);
1091 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1092 ERROR_EXIT(UNOQUORUM);
1093 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1094 ERROR_EXIT(UNOTSYNC);
1096 /* Write to the local disk */
1098 udisk_write(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
1101 udisk_abort(transPtr);
1102 transPtr->iovec_info.iovec_wrt_len = 0;
1103 transPtr->iovec_data.iovec_buf_len = 0;
1104 DBRELE(transPtr->dbase);
1108 /* Collect writes for the other ubik servers (to be done in bulk) */
1109 iovec = (struct ubik_iovec *)transPtr->iovec_info.iovec_wrt_val;
1110 iovec[transPtr->iovec_info.iovec_wrt_len].file = transPtr->seekFile;
1111 iovec[transPtr->iovec_info.iovec_wrt_len].position = transPtr->seekPos;
1112 iovec[transPtr->iovec_info.iovec_wrt_len].length = length;
1114 memcpy(&transPtr->iovec_data.
1115 iovec_buf_val[transPtr->iovec_data.iovec_buf_len], buffer, length);
1117 transPtr->iovec_info.iovec_wrt_len++;
1118 transPtr->iovec_data.iovec_buf_len += length;
1119 transPtr->seekPos += length;
1122 DBRELE(transPtr->dbase);
1127 * \brief This sets the file pointer associated with the current transaction
1128 * to the appropriate file and byte position.
1130 * Unlike Unix files, a transaction is labelled by both a file number \p fileid
1131 * and a byte position relative to the specified file \p position.
1134 ubik_Seek(struct ubik_trans *transPtr, afs_int32 fileid,
1139 DBHOLD(transPtr->dbase);
1140 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
1143 transPtr->seekFile = fileid;
1144 transPtr->seekPos = position;
1147 DBRELE(transPtr->dbase);
1152 * \brief This call returns the file pointer associated with the specified
1153 * transaction in \p fileid and \p position.
1156 ubik_Tell(struct ubik_trans *transPtr, afs_int32 * fileid,
1157 afs_int32 * position)
1159 DBHOLD(transPtr->dbase);
1160 *fileid = transPtr->seekFile;
1161 *position = transPtr->seekPos;
1162 DBRELE(transPtr->dbase);
1167 * \brief This sets the file size for the currently-selected file to \p length
1168 * bytes, if length is less than the file's current size.
1171 ubik_Truncate(struct ubik_trans *transPtr, afs_int32 length)
1173 afs_int32 code, error = 0;
1175 /* Will also catch if not UBIK_WRITETRANS */
1176 code = ubik_Flush(transPtr);
1180 DBHOLD(transPtr->dbase);
1181 /* first, check that quorum is still good, and that dbase is up-to-date */
1182 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1183 ERROR_EXIT(UNOQUORUM);
1184 if (!ubeacon_AmSyncSite())
1185 ERROR_EXIT(UNOTSYNC);
1187 /* now do the operation locally, and propagate it out */
1188 code = udisk_truncate(transPtr, transPtr->seekFile, length);
1191 ContactQuorum_DISK_Truncate(transPtr, 0, transPtr->seekFile,
1195 /* we must abort the operation */
1196 udisk_abort(transPtr);
1197 /* force aborts to the others */
1198 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0, "DISK_Abort");
1203 DBRELE(transPtr->dbase);
1208 * \brief set a lock; all locks are released on transaction end (commit/abort)
1211 ubik_SetLock(struct ubik_trans *atrans, afs_int32 apos, afs_int32 alen,
1214 afs_int32 code = 0, error = 0;
1216 if (atype == LOCKWRITE) {
1217 if (atrans->type == UBIK_READTRANS)
1219 code = ubik_Flush(atrans);
1224 DBHOLD(atrans->dbase);
1225 if (atype == LOCKREAD) {
1226 code = ulock_getLock(atrans, atype, 1);
1230 /* first, check that quorum is still good, and that dbase is up-to-date */
1231 if (!urecovery_AllBetter(atrans->dbase, atrans->flags & TRREADANY))
1232 ERROR_EXIT(UNOQUORUM);
1233 if (!ubeacon_AmSyncSite())
1234 ERROR_EXIT(UNOTSYNC);
1236 /* now do the operation locally, and propagate it out */
1237 code = ulock_getLock(atrans, atype, 1);
1239 code = ContactQuorum_DISK_Lock(atrans, 0, 0, 1 /*unused */ ,
1240 1 /*unused */ , LOCKWRITE);
1243 /* we must abort the operation */
1244 udisk_abort(atrans);
1245 /* force aborts to the others */
1246 ContactQuorum_NoArguments(DISK_Abort, atrans, 0, "DISK_Abort");
1252 DBRELE(atrans->dbase);
1257 * \brief utility to wait for a version # to change
1260 ubik_WaitVersion(struct ubik_dbase *adatabase,
1261 struct ubik_version *aversion)
1265 /* wait until version # changes, and then return */
1266 if (vcmp(*aversion, adatabase->version) != 0) {
1270 #ifdef AFS_PTHREAD_ENV
1271 opr_cv_wait(&adatabase->version_cond, &adatabase->versionLock);
1274 LWP_WaitProcess(&adatabase->version); /* same vers, just wait */
1281 * \brief utility to get the version of the dbase a transaction is dealing with
1284 ubik_GetVersion(struct ubik_trans *atrans,
1285 struct ubik_version *avers)
1287 DBHOLD(atrans->dbase);
1288 *avers = atrans->dbase->version;
1289 DBRELE(atrans->dbase);
1294 * \brief Facility to simplify database caching.
1295 * \return zero if last trans was done on the local server and was successful.
1296 * \return -1 means bad (NULL) argument.
1298 * If return value is non-zero and the caller is a server caching part of the
1299 * Ubik database, it should invalidate that cache.
1302 ubik_CacheUpdate(struct ubik_trans *atrans)
1304 if (!(atrans && atrans->dbase))
1306 return vcmp(atrans->dbase->cachedVersion, atrans->dbase->version) != 0;
1310 * check and possibly update cache of ubik db.
1312 * If the version of the cached db data is out of date, this calls (*check) to
1313 * update the cache. If (*check) returns success, we update the version of the
1316 * Checking the version of the cached db data is done under a read lock;
1317 * updating the cache (and thus calling (*check)) is done under a write lock
1318 * so is guaranteed not to interfere with another thread's (*check). On
1319 * successful return, a read lock on the cached db data is obtained, which
1320 * will be released by ubik_EndTrans or ubik_AbortTrans.
1322 * @param[in] atrans ubik transaction
1323 * @param[in] check function to call to check/update cache
1324 * @param[in] rock rock to pass to *check
1326 * @return operation status
1328 * @retval nonzero error; cachedVersion not updated
1330 * @post On success, application cache is read-locked, and cache data is
1334 ubik_CheckCache(struct ubik_trans *atrans, ubik_updatecache_func cbf, void *rock)
1338 if (!(atrans && atrans->dbase))
1341 ObtainReadLock(&atrans->dbase->cache_lock);
1343 while (ubik_CacheUpdate(atrans) != 0) {
1345 ReleaseReadLock(&atrans->dbase->cache_lock);
1346 ObtainSharedLock(&atrans->dbase->cache_lock);
1348 if (ubik_CacheUpdate(atrans) != 0) {
1350 BoostSharedLock(&atrans->dbase->cache_lock);
1352 ret = (*cbf) (atrans, rock);
1354 memcpy(&atrans->dbase->cachedVersion, &atrans->dbase->version,
1355 sizeof(atrans->dbase->cachedVersion));
1359 /* It would be nice if we could convert from a shared lock to a read
1360 * lock... instead, just release the shared and acquire the read */
1361 ReleaseSharedLock(&atrans->dbase->cache_lock);
1364 /* if we have an error, don't retry, and don't hold any locks */
1368 ObtainReadLock(&atrans->dbase->cache_lock);
1371 atrans->flags |= TRCACHELOCKED;
1377 * "Who said anything about panicking?" snapped Arthur.
1378 * "This is still just the culture shock. You wait till I've settled down
1379 * into the situation and found my bearings. \em Then I'll start panicking!"
1382 * \returns There is no return from panic.
1385 panic(char *format, ...)
1389 va_start(ap, format);
1390 ViceLog(0, ("Ubik PANIC:\n"));
1391 vViceLog(0, (format, ap));
1395 ViceLog(0, ("BACK FROM ABORT\n")); /* shouldn't come back */
1396 exit(1); /* never know, though */
1400 * This function takes an IP addresses as its parameter. It returns the
1401 * the primary IP address that is on the host passed in, or 0 if not found.
1404 ubikGetPrimaryInterfaceAddr(afs_uint32 addr)
1406 struct ubik_server *ts;
1410 for (ts = ubik_servers; ts; ts = ts->next)
1411 for (j = 0; j < UBIK_MAX_INTERFACE_ADDR; j++)
1412 if (ts->addr[j] == addr) {
1414 return ts->addr[0]; /* net byte order */
1417 return 0; /* if not in server database, return error */
1421 ubik_CheckAuth(struct rx_call *acall)
1423 if (checkSecurityProc)
1424 return (*checkSecurityProc) (securityRock, acall);
1425 else if (ubik_CheckRXSecurityProc) {
1426 return (*ubik_CheckRXSecurityProc) (ubik_CheckRXSecurityRock, acall);
1432 ubik_SetServerSecurityProcs(void (*buildproc) (void *,
1433 struct rx_securityClass ***,
1435 int (*checkproc) (void *, struct rx_call *),
1438 buildSecClassesProc = buildproc;
1439 checkSecurityProc = checkproc;
1440 securityRock = rock;