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)
156 /* Initial call - start iterating over servers */
163 Quorum_EndIO(atrans, *conn);
165 if (code) { /* failure */
168 (*ts)->up = 0; /* mark as down now; beacons will no longer be sent */
169 (*ts)->beaconSinceDown = 0;
171 (*ts)->currentDB = 0;
172 urecovery_LostServer(*ts); /* tell recovery to try to resend dbase later */
173 } else { /* success */
175 (*okcalls)++; /* count up how many worked */
176 if (aflags & CStampVersion) {
177 (*ts)->version = atrans->dbase->version;
186 if (!(*ts)->up || !(*ts)->currentDB ||
187 /* do not call DISK_Begin until we know that lastYesState is set on the
188 * remote in question; otherwise, DISK_Begin will fail. */
189 ((aflags & CCheckSyncAdvertised) && !((*ts)->beaconSinceDown && (*ts)->lastVote))) {
191 (*ts)->currentDB = 0; /* db is no longer current; we just missed an update */
192 return 0; /* not up-to-date, don't bother. NULL conn will tell caller not to use */
195 *conn = Quorum_StartIO(atrans, *ts);
200 ContactQuorum_rcode(int okcalls, afs_int32 rcode)
203 * return 0 if we successfully contacted a quorum, otherwise return error code.
204 * We don't have to contact ourselves (that was done locally)
206 if (okcalls + 1 >= ubik_quorum)
209 return (rcode != 0) ? rcode : UNOQUORUM;
213 * \brief Perform an operation at a quorum, handling error conditions.
214 * \return 0 if all worked and a quorum was contacted successfully
215 * \return otherwise mark failing server as down and return #UERROR
217 * \note If any server misses an update, we must wait #BIGTIME seconds before
218 * allowing the transaction to commit, to ensure that the missing and
219 * possibly still functioning server times out and stops handing out old
220 * data. This is done in the commit code, where we wait for a server marked
221 * down to have stayed down for #BIGTIME seconds before we allow a transaction
222 * to commit. A server that fails but comes back up won't give out old data
223 * because it is sent the sync count along with the beacon message that
224 * marks it as \b really up (\p beaconSinceDown).
227 ContactQuorum_NoArguments(afs_int32 (*proc)(struct rx_connection *, ubik_tid *),
228 struct ubik_trans *atrans, int aflags)
230 struct ubik_server *ts = NULL;
231 afs_int32 code = 0, rcode, okcalls;
232 struct rx_connection *conn;
235 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
238 code = (*proc)(conn, &atrans->tid);
239 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
241 return ContactQuorum_rcode(okcalls, rcode);
246 ContactQuorum_DISK_Lock(struct ubik_trans *atrans, int aflags,afs_int32 file,
247 afs_int32 position, afs_int32 length, afs_int32 type)
249 struct ubik_server *ts = NULL;
250 afs_int32 code = 0, rcode, okcalls;
251 struct rx_connection *conn;
254 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
257 code = DISK_Lock(conn, &atrans->tid, file, position, length, type);
258 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
260 return ContactQuorum_rcode(okcalls, rcode);
265 ContactQuorum_DISK_Write(struct ubik_trans *atrans, int aflags,
266 afs_int32 file, afs_int32 position, bulkdata *data)
268 struct ubik_server *ts = NULL;
269 afs_int32 code = 0, rcode, okcalls;
270 struct rx_connection *conn;
273 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
276 code = DISK_Write(conn, &atrans->tid, file, position, data);
277 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
279 return ContactQuorum_rcode(okcalls, rcode);
284 ContactQuorum_DISK_Truncate(struct ubik_trans *atrans, int aflags,
285 afs_int32 file, afs_int32 length)
287 struct ubik_server *ts = NULL;
288 afs_int32 code = 0, rcode, okcalls;
289 struct rx_connection *conn;
292 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
295 code = DISK_Truncate(conn, &atrans->tid, file, length);
296 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
298 return ContactQuorum_rcode(okcalls, rcode);
303 ContactQuorum_DISK_WriteV(struct ubik_trans *atrans, int aflags,
304 iovec_wrt * io_vector, iovec_buf *io_buffer)
306 struct ubik_server *ts = NULL;
307 afs_int32 code = 0, rcode, okcalls;
308 struct rx_connection *conn;
311 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
314 code = DISK_WriteV(conn, &atrans->tid, io_vector, io_buffer);
315 if ((code <= -450) && (code > -500)) {
316 /* An RPC interface mismatch (as defined in comerr/error_msg.c).
317 * Un-bulk the entries and do individual DISK_Write calls
318 * instead of DISK_WriteV.
320 struct ubik_iovec *iovec =
321 (struct ubik_iovec *)io_vector->iovec_wrt_val;
322 char *iobuf = (char *)io_buffer->iovec_buf_val;
326 for (i = 0, offset = 0; i < io_vector->iovec_wrt_len; i++) {
327 /* Sanity check for going off end of buffer */
328 if ((offset + iovec[i].length) > io_buffer->iovec_buf_len) {
332 tcbs.bulkdata_len = iovec[i].length;
333 tcbs.bulkdata_val = &iobuf[offset];
334 code = DISK_Write(conn, &atrans->tid, iovec[i].file,
335 iovec[i].position, &tcbs);
338 offset += iovec[i].length;
342 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
344 return ContactQuorum_rcode(okcalls, rcode);
349 ContactQuorum_DISK_SetVersion(struct ubik_trans *atrans, int aflags,
350 ubik_version *OldVersion,
351 ubik_version *NewVersion)
353 struct ubik_server *ts = NULL;
354 afs_int32 code = 0, rcode, okcalls;
355 struct rx_connection *conn;
358 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
361 code = DISK_SetVersion(conn, &atrans->tid, OldVersion, NewVersion);
362 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
364 return ContactQuorum_rcode(okcalls, rcode);
367 #if defined(AFS_PTHREAD_ENV)
369 ubik_thread_create(pthread_attr_t *tattr, pthread_t *thread, void *proc) {
370 opr_Verify(pthread_attr_init(tattr) == 0);
371 opr_Verify(pthread_attr_setdetachstate(tattr,
372 PTHREAD_CREATE_DETACHED) == 0);
373 opr_Verify(pthread_create(thread, tattr, proc, NULL) == 0);
379 * \brief This routine initializes the ubik system for a set of servers.
380 * \return 0 for success, or an error code on failure.
381 * \param serverList set of servers specified; nServers gives the number of entries in this array.
382 * \param pathName provides an initial prefix used for naming storage files used by this system.
383 * \param dbase the returned structure representing this instance of an ubik; it is passed to various calls below.
385 * \todo This routine should perhaps be generalized to a low-level disk interface providing read, write, file enumeration and sync operations.
387 * \warning The host named by myHost should not also be listed in serverList.
389 * \see ubik_ServerInit(), ubik_ServerInitByInfo()
392 ubik_ServerInitCommon(afs_uint32 myHost, short myPort,
393 struct afsconf_cell *info, char clones[],
394 afs_uint32 serverList[], const char *pathName,
395 struct ubik_dbase **dbase)
397 struct ubik_dbase *tdb;
399 #ifdef AFS_PTHREAD_ENV
400 pthread_t rxServerThread; /* pthread variables */
401 pthread_t ubeacon_InteractThread;
402 pthread_t urecovery_InteractThread;
403 pthread_attr_t rxServer_tattr;
404 pthread_attr_t ubeacon_Interact_tattr;
405 pthread_attr_t urecovery_Interact_tattr;
408 extern int rx_stackSize;
412 struct rx_securityClass *secClass;
415 struct rx_service *tservice;
417 initialize_U_error_table();
419 tdb = malloc(sizeof(struct ubik_dbase));
420 tdb->pathName = strdup(pathName);
421 tdb->activeTrans = (struct ubik_trans *)0;
422 memset(&tdb->version, 0, sizeof(struct ubik_version));
423 memset(&tdb->cachedVersion, 0, sizeof(struct ubik_version));
424 #ifdef AFS_PTHREAD_ENV
425 opr_mutex_init(&tdb->versionLock);
426 opr_mutex_init(&beacon_globals.beacon_lock);
427 opr_mutex_init(&vote_globals.vote_lock);
428 opr_mutex_init(&addr_globals.addr_lock);
429 opr_mutex_init(&version_globals.version_lock);
431 Lock_Init(&tdb->versionLock);
433 Lock_Init(&tdb->cache_lock);
435 tdb->read = uphys_read;
436 tdb->write = uphys_write;
437 tdb->truncate = uphys_truncate;
438 tdb->open = uphys_invalidate; /* this function isn't used any more */
439 tdb->sync = uphys_sync;
440 tdb->stat = uphys_stat;
441 tdb->getlabel = uphys_getlabel;
442 tdb->setlabel = uphys_setlabel;
443 tdb->getnfiles = uphys_getnfiles;
445 tdb->tidCounter = tdb->writeTidCounter = 0;
447 ubik_dbase = tdb; /* for now, only one db per server; can fix later when we have names for the other dbases */
449 #ifdef AFS_PTHREAD_ENV
450 opr_cv_init(&tdb->version_cond);
451 opr_cv_init(&tdb->flags_cond);
452 #endif /* AFS_PTHREAD_ENV */
456 /* the following call is idempotent so when/if it got called earlier,
457 * by whatever called us, it doesn't really matter -- klm */
458 code = rx_Init(myPort);
462 ubik_callPortal = myPort;
464 udisk_Init(ubik_nBuffers);
470 code = urecovery_Initialize(tdb);
474 code = ubeacon_InitServerListByInfo(myHost, info, clones);
476 code = ubeacon_InitServerList(myHost, serverList);
480 /* try to get an additional security object */
481 if (buildSecClassesProc == NULL) {
483 ubik_sc = calloc(numClasses, sizeof(struct rx_securityClass *));
484 ubik_sc[0] = rxnull_NewServerSecurityObject();
485 if (ubik_SRXSecurityProc) {
486 code = (*ubik_SRXSecurityProc) (ubik_SRXSecurityRock,
490 ubik_sc[secIndex] = secClass;
494 (*buildSecClassesProc) (securityRock, &ubik_sc, &numClasses);
496 /* for backwards compat this should keep working as it does now
500 rx_NewService(0, VOTE_SERVICE_ID, "VOTE", ubik_sc, numClasses,
501 VOTE_ExecuteRequest);
502 if (tservice == (struct rx_service *)0) {
503 ViceLog(5, ("Could not create VOTE rx service!\n"));
506 rx_SetMinProcs(tservice, 2);
507 rx_SetMaxProcs(tservice, 3);
510 rx_NewService(0, DISK_SERVICE_ID, "DISK", ubik_sc, numClasses,
511 DISK_ExecuteRequest);
512 if (tservice == (struct rx_service *)0) {
513 ViceLog(5, ("Could not create DISK rx service!\n"));
516 rx_SetMinProcs(tservice, 2);
517 rx_SetMaxProcs(tservice, 3);
519 /* start an rx_ServerProc to handle incoming RPC's in particular the
520 * UpdateInterfaceAddr RPC that occurs in ubeacon_InitServerList. This avoids
521 * the "steplock" problem in ubik initialization. Defect 11037.
523 #ifdef AFS_PTHREAD_ENV
524 ubik_thread_create(&rxServer_tattr, &rxServerThread, (void *)rx_ServerProc);
526 LWP_CreateProcess(rx_ServerProc, rx_stackSize, RX_PROCESS_PRIORITY,
527 NULL, "rx_ServerProc", &junk);
530 /* send addrs to all other servers */
531 code = ubeacon_updateUbikNetworkAddress(ubik_host);
535 /* now start up async processes */
536 #ifdef AFS_PTHREAD_ENV
537 ubik_thread_create(&ubeacon_Interact_tattr, &ubeacon_InteractThread,
538 (void *)ubeacon_Interact);
540 code = LWP_CreateProcess(ubeacon_Interact, 16384 /*8192 */ ,
541 LWP_MAX_PRIORITY - 1, (void *)0, "beacon",
547 #ifdef AFS_PTHREAD_ENV
548 ubik_thread_create(&urecovery_Interact_tattr, &urecovery_InteractThread,
549 (void *)urecovery_Interact);
550 return 0; /* is this correct? - klm */
552 code = LWP_CreateProcess(urecovery_Interact, 16384 /*8192 */ ,
553 LWP_MAX_PRIORITY - 1, (void *)0, "recovery",
561 * \see ubik_ServerInitCommon()
564 ubik_ServerInitByInfo(afs_uint32 myHost, short myPort,
565 struct afsconf_cell *info, char clones[],
566 const char *pathName, struct ubik_dbase **dbase)
571 ubik_ServerInitCommon(myHost, myPort, info, clones, 0, pathName,
577 * \see ubik_ServerInitCommon()
580 ubik_ServerInit(afs_uint32 myHost, short myPort, afs_uint32 serverList[],
581 const char *pathName, struct ubik_dbase **dbase)
586 ubik_ServerInitCommon(myHost, myPort, (struct afsconf_cell *)0, 0,
587 serverList, pathName, dbase);
592 * \brief This routine begins a read or write transaction on the transaction
593 * identified by transPtr, in the dbase named by dbase.
595 * An open mode of ubik_READTRANS identifies this as a read transaction,
596 * while a mode of ubik_WRITETRANS identifies this as a write transaction.
597 * transPtr is set to the returned transaction control block.
598 * The readAny flag is set to 0 or 1 or 2 by the wrapper functions
599 * ubik_BeginTrans() or ubik_BeginTransReadAny() or
600 * ubik_BeginTransReadAnyWrite() below.
602 * \note We can only begin transaction when we have an up-to-date database.
605 BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
606 struct ubik_trans **transPtr, int readAny)
608 struct ubik_trans *jt;
609 struct ubik_trans *tt;
612 if (readAny > 1 && ubik_SyncWriterCacheProc == NULL) {
613 /* it's not safe to use ubik_BeginTransReadAnyWrite without a
614 * cache-syncing function; fall back to ubik_BeginTransReadAny,
615 * which is safe but slower */
616 ViceLog(0, ("ubik_BeginTransReadAnyWrite called, but "
617 "ubik_SyncWriterCacheProc not set; pretending "
618 "ubik_BeginTransReadAny was called instead\n"));
622 if ((transMode != UBIK_READTRANS) && readAny)
625 if (urecovery_AllBetter(dbase, readAny) == 0) {
629 /* otherwise we have a quorum, use it */
631 /* make sure that at most one write transaction occurs at any one time. This
632 * has nothing to do with transaction locking; that's enforced by the lock package. However,
633 * we can't even handle two non-conflicting writes, since our log and recovery modules
634 * don't know how to restore one without possibly picking up some data from the other. */
635 if (transMode == UBIK_WRITETRANS) {
636 /* if we're writing already, wait */
637 while (dbase->flags & DBWRITING) {
638 #ifdef AFS_PTHREAD_ENV
639 opr_cv_wait(&dbase->flags_cond, &dbase->versionLock);
642 LWP_WaitProcess(&dbase->flags);
647 if (!ubeacon_AmSyncSite()) {
651 if (!ubeacon_SyncSiteAdvertised()) {
652 /* i am the sync-site but the remotes are not aware yet */
658 /* create the transaction */
659 code = udisk_begin(dbase, transMode, &jt); /* can't take address of register var */
660 tt = jt; /* move to a register */
661 if (code || tt == NULL) {
667 tt->flags |= TRREADANY;
669 tt->flags |= TRREADWRITE;
672 /* label trans and dbase with new tid */
673 tt->tid.epoch = version_globals.ubik_epochTime;
674 /* bump by two, since tidCounter+1 means trans id'd by tidCounter has finished */
675 tt->tid.counter = (dbase->tidCounter += 2);
677 if (transMode == UBIK_WRITETRANS) {
678 /* for a write trans, we have to keep track of the write tid counter too */
679 dbase->writeTidCounter = tt->tid.counter;
684 if (transMode == UBIK_WRITETRANS) {
685 /* next try to start transaction on appropriate number of machines */
686 code = ContactQuorum_NoArguments(DISK_Begin, tt, CCheckSyncAdvertised);
688 /* we must abort the operation */
690 ContactQuorum_NoArguments(DISK_Abort, tt, 0); /* force aborts to the others */
706 ubik_BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
707 struct ubik_trans **transPtr)
709 return BeginTrans(dbase, transMode, transPtr, 0);
716 ubik_BeginTransReadAny(struct ubik_dbase *dbase, afs_int32 transMode,
717 struct ubik_trans **transPtr)
719 return BeginTrans(dbase, transMode, transPtr, 1);
726 ubik_BeginTransReadAnyWrite(struct ubik_dbase *dbase, afs_int32 transMode,
727 struct ubik_trans **transPtr)
729 return BeginTrans(dbase, transMode, transPtr, 2);
733 * \brief This routine ends a read or write transaction by aborting it.
736 ubik_AbortTrans(struct ubik_trans *transPtr)
740 struct ubik_dbase *dbase;
742 dbase = transPtr->dbase;
744 if (transPtr->flags & TRCACHELOCKED) {
745 ReleaseReadLock(&dbase->cache_lock);
746 transPtr->flags &= ~TRCACHELOCKED;
749 ObtainWriteLock(&dbase->cache_lock);
752 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
754 ReleaseWriteLock(&dbase->cache_lock);
756 /* see if we're still up-to-date */
757 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
758 udisk_abort(transPtr);
764 if (transPtr->type == UBIK_READTRANS) {
765 code = udisk_abort(transPtr);
771 /* below here, we know we're doing a write transaction */
772 if (!ubeacon_AmSyncSite()) {
773 udisk_abort(transPtr);
779 /* now it is safe to try remote abort */
780 code = ContactQuorum_NoArguments(DISK_Abort, transPtr, 0);
781 code2 = udisk_abort(transPtr);
784 return (code ? code : code2);
788 WritebackApplicationCache(struct ubik_dbase *dbase)
791 if (ubik_SyncWriterCacheProc) {
792 code = ubik_SyncWriterCacheProc();
795 /* we failed to sync the local cache, so just invalidate the cache;
796 * we'll try to read the cache in again on the next read */
797 memset(&dbase->cachedVersion, 0, sizeof(dbase->cachedVersion));
799 memcpy(&dbase->cachedVersion, &dbase->version,
800 sizeof(dbase->cachedVersion));
805 * \brief This routine ends a read or write transaction on the open transaction identified by transPtr.
806 * \return an error code.
809 ubik_EndTrans(struct ubik_trans *transPtr)
814 struct ubik_server *ts;
817 struct ubik_dbase *dbase;
819 if (transPtr->type == UBIK_WRITETRANS) {
820 code = ubik_Flush(transPtr);
822 ubik_AbortTrans(transPtr);
827 dbase = transPtr->dbase;
829 if (transPtr->flags & TRCACHELOCKED) {
830 ReleaseReadLock(&dbase->cache_lock);
831 transPtr->flags &= ~TRCACHELOCKED;
834 if (transPtr->type != UBIK_READTRANS) {
835 /* must hold cache_lock before DBHOLD'ing */
836 ObtainWriteLock(&dbase->cache_lock);
842 /* give up if no longer current */
843 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
844 udisk_abort(transPtr);
851 if (transPtr->type == UBIK_READTRANS) { /* reads are easy */
852 code = udisk_commit(transPtr);
854 goto success; /* update cachedVersion correctly */
860 if (!ubeacon_AmSyncSite()) { /* no longer sync site */
861 udisk_abort(transPtr);
868 /* now it is safe to do commit */
869 code = udisk_commit(transPtr);
871 /* db data has been committed locally; update the local cache so
872 * readers can get at it */
873 WritebackApplicationCache(dbase);
875 ReleaseWriteLock(&dbase->cache_lock);
877 code = ContactQuorum_NoArguments(DISK_Commit, transPtr, CStampVersion);
880 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
881 ReleaseWriteLock(&dbase->cache_lock);
885 /* failed to commit, so must return failure. Try to clear locks first, just for fun
886 * Note that we don't know if this transaction will eventually commit at this point.
887 * If it made it to a site that will be present in the next quorum, we win, otherwise
888 * we lose. If we contact a majority of sites, then we won't be here: contacting
889 * a majority guarantees commit, since it guarantees that one dude will be a
890 * member of the next quorum. */
891 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0);
896 /* before we can start sending unlock messages, we must wait until all servers
897 * that are possibly still functioning on the other side of a network partition
898 * have timed out. Check the server structures, compute how long to wait, then
899 * start the unlocks */
900 realStart = FT_ApproxTime();
902 /* wait for all servers to time out */
904 now = FT_ApproxTime();
905 /* check if we're still sync site, the guy should either come up
906 * to us, or timeout. Put safety check in anyway */
907 if (now - realStart > 10 * BIGTIME) {
908 ubik_stats.escapes++;
909 ViceLog(0, ("ubik escaping from commit wait\n"));
912 for (ts = ubik_servers; ts; ts = ts->next) {
914 if (!ts->beaconSinceDown && now <= ts->lastBeaconSent + BIGTIME) {
917 /* this guy could have some damaged data, wait for him */
919 tv.tv_sec = 1; /* try again after a while (ha ha) */
922 #ifdef AFS_PTHREAD_ENV
923 /* we could release the dbase outside of the loop, but we do
924 * it here, in the loop, to avoid an unnecessary RELE/HOLD
925 * if all sites are up */
927 select(0, 0, 0, 0, &tv);
930 IOMGR_Select(0, 0, 0, 0, &tv); /* poll, should we wait on something? */
938 break; /* no down ones still pseudo-active */
941 /* finally, unlock all the dudes. We can return success independent of the number of servers
942 * that really unlock the dbase; the others will do it if/when they elect a new sync site.
943 * The transaction is committed anyway, since we succeeded in contacting a quorum
944 * at the start (when invoking the DiskCommit function).
946 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0);
950 /* don't update cachedVersion here; it should have been updated way back
951 * in ubik_CheckCache, and earlier in this function for writes */
954 ReleaseWriteLock(&dbase->cache_lock);
960 ObtainWriteLock(&dbase->cache_lock);
962 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
963 ReleaseWriteLock(&dbase->cache_lock);
968 * \brief This routine reads length bytes into buffer from the current position in the database.
970 * 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.
972 * \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.
975 ubik_Read(struct ubik_trans *transPtr, void *buffer,
980 /* reads are easy to do: handle locally */
981 DBHOLD(transPtr->dbase);
982 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
983 DBRELE(transPtr->dbase);
988 udisk_read(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
991 transPtr->seekPos += length;
993 DBRELE(transPtr->dbase);
998 * \brief This routine will flush the io data in the iovec structures.
1000 * It first flushes to the local disk and then uses ContactQuorum to write it
1001 * to the other servers.
1004 ubik_Flush(struct ubik_trans *transPtr)
1006 afs_int32 code, error = 0;
1008 if (transPtr->type != UBIK_WRITETRANS)
1011 DBHOLD(transPtr->dbase);
1012 if (!transPtr->iovec_info.iovec_wrt_len
1013 || !transPtr->iovec_info.iovec_wrt_val) {
1014 DBRELE(transPtr->dbase);
1018 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1019 ERROR_EXIT(UNOQUORUM);
1020 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1021 ERROR_EXIT(UNOTSYNC);
1023 /* Update the rest of the servers in the quorum */
1025 ContactQuorum_DISK_WriteV(transPtr, 0, &transPtr->iovec_info,
1026 &transPtr->iovec_data);
1028 udisk_abort(transPtr);
1029 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0); /* force aborts to the others */
1030 transPtr->iovec_info.iovec_wrt_len = 0;
1031 transPtr->iovec_data.iovec_buf_len = 0;
1035 /* Wrote the buffers out, so start at scratch again */
1036 transPtr->iovec_info.iovec_wrt_len = 0;
1037 transPtr->iovec_data.iovec_buf_len = 0;
1040 DBRELE(transPtr->dbase);
1045 ubik_Write(struct ubik_trans *transPtr, void *vbuffer,
1048 struct ubik_iovec *iovec;
1049 afs_int32 code, error = 0;
1050 afs_int32 pos, len, size;
1051 char * buffer = (char *)vbuffer;
1053 if (transPtr->type != UBIK_WRITETRANS)
1058 if (length > IOVEC_MAXBUF) {
1059 for (pos = 0, len = length; len > 0; len -= size, pos += size) {
1060 size = ((len < IOVEC_MAXBUF) ? len : IOVEC_MAXBUF);
1061 code = ubik_Write(transPtr, buffer+pos, size);
1068 DBHOLD(transPtr->dbase);
1069 if (!transPtr->iovec_info.iovec_wrt_val) {
1070 transPtr->iovec_info.iovec_wrt_len = 0;
1071 transPtr->iovec_info.iovec_wrt_val =
1072 malloc(IOVEC_MAXWRT * sizeof(struct ubik_iovec));
1073 transPtr->iovec_data.iovec_buf_len = 0;
1074 transPtr->iovec_data.iovec_buf_val = malloc(IOVEC_MAXBUF);
1075 if (!transPtr->iovec_info.iovec_wrt_val
1076 || !transPtr->iovec_data.iovec_buf_val) {
1077 if (transPtr->iovec_info.iovec_wrt_val)
1078 free(transPtr->iovec_info.iovec_wrt_val);
1079 transPtr->iovec_info.iovec_wrt_val = 0;
1080 if (transPtr->iovec_data.iovec_buf_val)
1081 free(transPtr->iovec_data.iovec_buf_val);
1082 transPtr->iovec_data.iovec_buf_val = 0;
1083 DBRELE(transPtr->dbase);
1088 /* If this write won't fit in the structure, then flush it out and start anew */
1089 if ((transPtr->iovec_info.iovec_wrt_len >= IOVEC_MAXWRT)
1090 || ((length + transPtr->iovec_data.iovec_buf_len) > IOVEC_MAXBUF)) {
1091 /* Can't hold the DB lock over ubik_Flush */
1092 DBRELE(transPtr->dbase);
1093 code = ubik_Flush(transPtr);
1096 DBHOLD(transPtr->dbase);
1099 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1100 ERROR_EXIT(UNOQUORUM);
1101 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1102 ERROR_EXIT(UNOTSYNC);
1104 /* Write to the local disk */
1106 udisk_write(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
1109 udisk_abort(transPtr);
1110 transPtr->iovec_info.iovec_wrt_len = 0;
1111 transPtr->iovec_data.iovec_buf_len = 0;
1112 DBRELE(transPtr->dbase);
1116 /* Collect writes for the other ubik servers (to be done in bulk) */
1117 iovec = (struct ubik_iovec *)transPtr->iovec_info.iovec_wrt_val;
1118 iovec[transPtr->iovec_info.iovec_wrt_len].file = transPtr->seekFile;
1119 iovec[transPtr->iovec_info.iovec_wrt_len].position = transPtr->seekPos;
1120 iovec[transPtr->iovec_info.iovec_wrt_len].length = length;
1122 memcpy(&transPtr->iovec_data.
1123 iovec_buf_val[transPtr->iovec_data.iovec_buf_len], buffer, length);
1125 transPtr->iovec_info.iovec_wrt_len++;
1126 transPtr->iovec_data.iovec_buf_len += length;
1127 transPtr->seekPos += length;
1130 DBRELE(transPtr->dbase);
1135 * \brief This sets the file pointer associated with the current transaction
1136 * to the appropriate file and byte position.
1138 * Unlike Unix files, a transaction is labelled by both a file number \p fileid
1139 * and a byte position relative to the specified file \p position.
1142 ubik_Seek(struct ubik_trans *transPtr, afs_int32 fileid,
1147 DBHOLD(transPtr->dbase);
1148 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
1151 transPtr->seekFile = fileid;
1152 transPtr->seekPos = position;
1155 DBRELE(transPtr->dbase);
1160 * \brief This call returns the file pointer associated with the specified
1161 * transaction in \p fileid and \p position.
1164 ubik_Tell(struct ubik_trans *transPtr, afs_int32 * fileid,
1165 afs_int32 * position)
1167 DBHOLD(transPtr->dbase);
1168 *fileid = transPtr->seekFile;
1169 *position = transPtr->seekPos;
1170 DBRELE(transPtr->dbase);
1175 * \brief This sets the file size for the currently-selected file to \p length
1176 * bytes, if length is less than the file's current size.
1179 ubik_Truncate(struct ubik_trans *transPtr, afs_int32 length)
1181 afs_int32 code, error = 0;
1183 /* Will also catch if not UBIK_WRITETRANS */
1184 code = ubik_Flush(transPtr);
1188 DBHOLD(transPtr->dbase);
1189 /* first, check that quorum is still good, and that dbase is up-to-date */
1190 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1191 ERROR_EXIT(UNOQUORUM);
1192 if (!ubeacon_AmSyncSite())
1193 ERROR_EXIT(UNOTSYNC);
1195 /* now do the operation locally, and propagate it out */
1196 code = udisk_truncate(transPtr, transPtr->seekFile, length);
1199 ContactQuorum_DISK_Truncate(transPtr, 0, transPtr->seekFile,
1203 /* we must abort the operation */
1204 udisk_abort(transPtr);
1205 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0); /* force aborts to the others */
1210 DBRELE(transPtr->dbase);
1215 * \brief set a lock; all locks are released on transaction end (commit/abort)
1218 ubik_SetLock(struct ubik_trans *atrans, afs_int32 apos, afs_int32 alen,
1221 afs_int32 code = 0, error = 0;
1223 if (atype == LOCKWRITE) {
1224 if (atrans->type == UBIK_READTRANS)
1226 code = ubik_Flush(atrans);
1231 DBHOLD(atrans->dbase);
1232 if (atype == LOCKREAD) {
1233 code = ulock_getLock(atrans, atype, 1);
1237 /* first, check that quorum is still good, and that dbase is up-to-date */
1238 if (!urecovery_AllBetter(atrans->dbase, atrans->flags & TRREADANY))
1239 ERROR_EXIT(UNOQUORUM);
1240 if (!ubeacon_AmSyncSite())
1241 ERROR_EXIT(UNOTSYNC);
1243 /* now do the operation locally, and propagate it out */
1244 code = ulock_getLock(atrans, atype, 1);
1246 code = ContactQuorum_DISK_Lock(atrans, 0, 0, 1 /*unused */ ,
1247 1 /*unused */ , LOCKWRITE);
1250 /* we must abort the operation */
1251 udisk_abort(atrans);
1252 ContactQuorum_NoArguments(DISK_Abort, atrans, 0); /* force aborts to the others */
1258 DBRELE(atrans->dbase);
1263 * \brief utility to wait for a version # to change
1266 ubik_WaitVersion(struct ubik_dbase *adatabase,
1267 struct ubik_version *aversion)
1271 /* wait until version # changes, and then return */
1272 if (vcmp(*aversion, adatabase->version) != 0) {
1276 #ifdef AFS_PTHREAD_ENV
1277 opr_cv_wait(&adatabase->version_cond, &adatabase->versionLock);
1280 LWP_WaitProcess(&adatabase->version); /* same vers, just wait */
1287 * \brief utility to get the version of the dbase a transaction is dealing with
1290 ubik_GetVersion(struct ubik_trans *atrans,
1291 struct ubik_version *avers)
1293 DBHOLD(atrans->dbase);
1294 *avers = atrans->dbase->version;
1295 DBRELE(atrans->dbase);
1300 * \brief Facility to simplify database caching.
1301 * \return zero if last trans was done on the local server and was successful.
1302 * \return -1 means bad (NULL) argument.
1304 * If return value is non-zero and the caller is a server caching part of the
1305 * Ubik database, it should invalidate that cache.
1308 ubik_CacheUpdate(struct ubik_trans *atrans)
1310 if (!(atrans && atrans->dbase))
1312 return vcmp(atrans->dbase->cachedVersion, atrans->dbase->version) != 0;
1316 * check and possibly update cache of ubik db.
1318 * If the version of the cached db data is out of date, this calls (*check) to
1319 * update the cache. If (*check) returns success, we update the version of the
1322 * Checking the version of the cached db data is done under a read lock;
1323 * updating the cache (and thus calling (*check)) is done under a write lock
1324 * so is guaranteed not to interfere with another thread's (*check). On
1325 * successful return, a read lock on the cached db data is obtained, which
1326 * will be released by ubik_EndTrans or ubik_AbortTrans.
1328 * @param[in] atrans ubik transaction
1329 * @param[in] check function to call to check/update cache
1330 * @param[in] rock rock to pass to *check
1332 * @return operation status
1334 * @retval nonzero error; cachedVersion not updated
1336 * @post On success, application cache is read-locked, and cache data is
1340 ubik_CheckCache(struct ubik_trans *atrans, ubik_updatecache_func cbf, void *rock)
1344 if (!(atrans && atrans->dbase))
1347 ObtainReadLock(&atrans->dbase->cache_lock);
1349 while (ubik_CacheUpdate(atrans) != 0) {
1351 ReleaseReadLock(&atrans->dbase->cache_lock);
1352 ObtainSharedLock(&atrans->dbase->cache_lock);
1354 if (ubik_CacheUpdate(atrans) != 0) {
1356 BoostSharedLock(&atrans->dbase->cache_lock);
1358 ret = (*cbf) (atrans, rock);
1360 memcpy(&atrans->dbase->cachedVersion, &atrans->dbase->version,
1361 sizeof(atrans->dbase->cachedVersion));
1365 /* It would be nice if we could convert from a shared lock to a read
1366 * lock... instead, just release the shared and acquire the read */
1367 ReleaseSharedLock(&atrans->dbase->cache_lock);
1370 /* if we have an error, don't retry, and don't hold any locks */
1374 ObtainReadLock(&atrans->dbase->cache_lock);
1377 atrans->flags |= TRCACHELOCKED;
1383 * "Who said anything about panicking?" snapped Arthur.
1384 * "This is still just the culture shock. You wait till I've settled down
1385 * into the situation and found my bearings. \em Then I'll start panicking!"
1388 * \returns There is no return from panic.
1391 panic(char *format, ...)
1395 va_start(ap, format);
1396 ViceLog(0, ("Ubik PANIC:\n"));
1397 vViceLog(0, (format, ap));
1401 ViceLog(0, ("BACK FROM ABORT\n")); /* shouldn't come back */
1402 exit(1); /* never know, though */
1406 * This function takes an IP addresses as its parameter. It returns the
1407 * the primary IP address that is on the host passed in, or 0 if not found.
1410 ubikGetPrimaryInterfaceAddr(afs_uint32 addr)
1412 struct ubik_server *ts;
1416 for (ts = ubik_servers; ts; ts = ts->next)
1417 for (j = 0; j < UBIK_MAX_INTERFACE_ADDR; j++)
1418 if (ts->addr[j] == addr) {
1420 return ts->addr[0]; /* net byte order */
1423 return 0; /* if not in server database, return error */
1427 ubik_CheckAuth(struct rx_call *acall)
1429 if (checkSecurityProc)
1430 return (*checkSecurityProc) (securityRock, acall);
1431 else if (ubik_CheckRXSecurityProc) {
1432 return (*ubik_CheckRXSecurityProc) (ubik_CheckRXSecurityRock, acall);
1438 ubik_SetServerSecurityProcs(void (*buildproc) (void *,
1439 struct rx_securityClass ***,
1441 int (*checkproc) (void *, struct rx_call *),
1444 buildSecClassesProc = buildproc;
1445 checkSecurityProc = checkproc;
1446 securityRock = rock;