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>
18 #include <afs/cellconfig.h>
20 #define UBIK_INTERNALS
24 #include <lwp.h> /* temporary hack by klm */
26 #define ERROR_EXIT(code) do { \
33 * This system is organized in a hierarchical set of related modules. Modules
34 * at one level can only call modules at the same level or below.
36 * At the bottom level (0) we have R, RFTP, LWP and IOMGR, i.e. the basic
37 * operating system primitives.
39 * At the next level (1) we have
41 * \li VOTER--The module responsible for casting votes when asked. It is also
42 * responsible for determining whether this server should try to become
43 * a synchronization site.
44 * \li BEACONER--The module responsible for sending keep-alives out when a
45 * server is actually the sync site, or trying to become a sync site.
46 * \li DISK--The module responsible for representing atomic transactions
47 * on the local disk. It maintains a new-value only log.
48 * \li LOCK--The module responsible for locking byte ranges in the database file.
50 * At the next level (2) we have
52 * \li RECOVERY--The module responsible for ensuring that all members of a quorum
53 * have the same up-to-date database after a new synchronization site is
54 * elected. This module runs only on the synchronization site.
56 * At the next level (3) we have
58 * \li REMOTE--The module responsible for interpreting requests from the sync
59 * site and applying them to the database, after obtaining the appropriate
62 * At the next level (4) we have
64 * \li UBIK--The module users call to perform operations on the database.
69 afs_int32 ubik_quorum = 0;
70 struct ubik_dbase *ubik_dbase = 0;
71 struct ubik_stats ubik_stats;
72 afs_uint32 ubik_host[UBIK_MAX_INTERFACE_ADDR];
73 afs_int32 ubik_epochTime = 0;
74 afs_int32 urecovery_state = 0;
75 int (*ubik_SyncWriterCacheProc) (void);
76 struct ubik_server *ubik_servers;
77 short ubik_callPortal;
79 /* These global variables were used to control the server security layers.
80 * They are retained for backwards compatibility with legacy callers.
82 * The ubik_SetServerSecurityProcs() interface should be used instead.
85 int (*ubik_SRXSecurityProc) (void *, struct rx_securityClass **, afs_int32 *);
86 void *ubik_SRXSecurityRock;
87 int (*ubik_CheckRXSecurityProc) (void *, struct rx_call *);
88 void *ubik_CheckRXSecurityRock;
92 static int BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
93 struct ubik_trans **transPtr, int readAny);
95 static struct rx_securityClass **ubik_sc = NULL;
96 static void (*buildSecClassesProc)(void *, struct rx_securityClass ***,
98 static int (*checkSecurityProc)(void *, struct rx_call *) = NULL;
99 static void *securityRock = NULL;
101 #define CStampVersion 1 /* meaning set ts->version */
103 static_inline struct rx_connection *
104 Quorum_StartIO(struct ubik_trans *atrans, struct ubik_server *as)
106 struct rx_connection *conn;
108 conn = as->disk_rxcid;
110 #ifdef AFS_PTHREAD_ENV
111 rx_GetConnection(conn);
112 DBRELE(atrans->dbase);
113 #endif /* AFS_PTHREAD_ENV */
119 Quorum_EndIO(struct ubik_trans *atrans, struct rx_connection *aconn)
121 #ifdef AFS_PTHREAD_ENV
122 DBHOLD(atrans->dbase);
123 rx_PutConnection(aconn);
124 #endif /* AFS_PTHREAD_ENV */
129 * Iterate over all servers. Callers pass in *ts which is used to track
130 * the current server.
131 * - Returns 1 if there are no more servers
132 * - Returns 0 with conn set to the connection for the current server if
133 * it's up and current
136 ContactQuorum_iterate(struct ubik_trans *atrans, int aflags, struct ubik_server **ts,
137 struct rx_connection **conn, afs_int32 *rcode,
138 afs_int32 *okcalls, afs_int32 code)
141 /* Initial call - start iterating over servers */
148 Quorum_EndIO(atrans, *conn);
150 if (code) { /* failure */
153 (*ts)->up = 0; /* mark as down now; beacons will no longer be sent */
154 (*ts)->beaconSinceDown = 0;
156 (*ts)->currentDB = 0;
157 urecovery_LostServer(*ts); /* tell recovery to try to resend dbase later */
158 } else { /* success */
160 (*okcalls)++; /* count up how many worked */
161 if (aflags & CStampVersion) {
162 (*ts)->version = atrans->dbase->version;
171 if (!(*ts)->up || !(*ts)->currentDB) {
173 (*ts)->currentDB = 0; /* db is no longer current; we just missed an update */
174 return 0; /* not up-to-date, don't bother. NULL conn will tell caller not to use */
177 *conn = Quorum_StartIO(atrans, *ts);
182 ContactQuorum_rcode(int okcalls, afs_int32 rcode)
185 * return 0 if we successfully contacted a quorum, otherwise return error code.
186 * We don't have to contact ourselves (that was done locally)
188 if (okcalls + 1 >= ubik_quorum)
195 * \brief Perform an operation at a quorum, handling error conditions.
196 * \return 0 if all worked and a quorum was contacted successfully
197 * \return otherwise mark failing server as down and return #UERROR
199 * \note If any server misses an update, we must wait #BIGTIME seconds before
200 * allowing the transaction to commit, to ensure that the missing and
201 * possibly still functioning server times out and stops handing out old
202 * data. This is done in the commit code, where we wait for a server marked
203 * down to have stayed down for #BIGTIME seconds before we allow a transaction
204 * to commit. A server that fails but comes back up won't give out old data
205 * because it is sent the sync count along with the beacon message that
206 * marks it as \b really up (\p beaconSinceDown).
209 ContactQuorum_NoArguments(afs_int32 (*proc)(struct rx_connection *, ubik_tid *),
210 struct ubik_trans *atrans, int aflags)
212 struct ubik_server *ts = NULL;
213 afs_int32 code = 0, rcode, okcalls;
214 struct rx_connection *conn;
217 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
220 code = (*proc)(conn, &atrans->tid);
221 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
223 return ContactQuorum_rcode(okcalls, rcode);
228 ContactQuorum_DISK_Lock(struct ubik_trans *atrans, int aflags,afs_int32 file,
229 afs_int32 position, afs_int32 length, afs_int32 type)
231 struct ubik_server *ts = NULL;
232 afs_int32 code = 0, rcode, okcalls;
233 struct rx_connection *conn;
236 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
239 code = DISK_Lock(conn, &atrans->tid, file, position, length, type);
240 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
242 return ContactQuorum_rcode(okcalls, rcode);
247 ContactQuorum_DISK_Write(struct ubik_trans *atrans, int aflags,
248 afs_int32 file, afs_int32 position, bulkdata *data)
250 struct ubik_server *ts = NULL;
251 afs_int32 code = 0, rcode, okcalls;
252 struct rx_connection *conn;
255 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
258 code = DISK_Write(conn, &atrans->tid, file, position, data);
259 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
261 return ContactQuorum_rcode(okcalls, rcode);
266 ContactQuorum_DISK_Truncate(struct ubik_trans *atrans, int aflags,
267 afs_int32 file, afs_int32 length)
269 struct ubik_server *ts = NULL;
270 afs_int32 code = 0, rcode, okcalls;
271 struct rx_connection *conn;
274 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
277 code = DISK_Truncate(conn, &atrans->tid, file, length);
278 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
280 return ContactQuorum_rcode(okcalls, rcode);
285 ContactQuorum_DISK_WriteV(struct ubik_trans *atrans, int aflags,
286 iovec_wrt * io_vector, iovec_buf *io_buffer)
288 struct ubik_server *ts = NULL;
289 afs_int32 code = 0, rcode, okcalls;
290 struct rx_connection *conn;
293 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
296 code = DISK_WriteV(conn, &atrans->tid, io_vector, io_buffer);
297 if ((code <= -450) && (code > -500)) {
298 /* An RPC interface mismatch (as defined in comerr/error_msg.c).
299 * Un-bulk the entries and do individual DISK_Write calls
300 * instead of DISK_WriteV.
302 struct ubik_iovec *iovec =
303 (struct ubik_iovec *)io_vector->iovec_wrt_val;
304 char *iobuf = (char *)io_buffer->iovec_buf_val;
308 for (i = 0, offset = 0; i < io_vector->iovec_wrt_len; i++) {
309 /* Sanity check for going off end of buffer */
310 if ((offset + iovec[i].length) > io_buffer->iovec_buf_len) {
314 tcbs.bulkdata_len = iovec[i].length;
315 tcbs.bulkdata_val = &iobuf[offset];
316 code = DISK_Write(conn, &atrans->tid, iovec[i].file,
317 iovec[i].position, &tcbs);
320 offset += iovec[i].length;
324 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
326 return ContactQuorum_rcode(okcalls, rcode);
331 ContactQuorum_DISK_SetVersion(struct ubik_trans *atrans, int aflags,
332 ubik_version *OldVersion,
333 ubik_version *NewVersion)
335 struct ubik_server *ts = NULL;
336 afs_int32 code = 0, rcode, okcalls;
337 struct rx_connection *conn;
340 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
343 code = DISK_SetVersion(conn, &atrans->tid, OldVersion, NewVersion);
344 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
346 return ContactQuorum_rcode(okcalls, rcode);
349 #if defined(AFS_PTHREAD_ENV)
351 ubik_thread_create(pthread_attr_t *tattr, pthread_t *thread, void *proc) {
352 osi_Assert(pthread_attr_init(tattr) == 0);
353 osi_Assert(pthread_attr_setdetachstate(tattr, PTHREAD_CREATE_DETACHED) == 0);
354 osi_Assert(pthread_create(thread, tattr, proc, NULL) == 0);
360 * \brief This routine initializes the ubik system for a set of servers.
361 * \return 0 for success, or an error code on failure.
362 * \param serverList set of servers specified; nServers gives the number of entries in this array.
363 * \param pathName provides an initial prefix used for naming storage files used by this system.
364 * \param dbase the returned structure representing this instance of an ubik; it is passed to various calls below.
366 * \todo This routine should perhaps be generalized to a low-level disk interface providing read, write, file enumeration and sync operations.
368 * \warning The host named by myHost should not also be listed in serverList.
370 * \see ubik_ServerInit(), ubik_ServerInitByInfo()
373 ubik_ServerInitCommon(afs_uint32 myHost, short myPort,
374 struct afsconf_cell *info, char clones[],
375 afs_uint32 serverList[], const char *pathName,
376 struct ubik_dbase **dbase)
378 struct ubik_dbase *tdb;
380 #ifdef AFS_PTHREAD_ENV
381 pthread_t rxServerThread; /* pthread variables */
382 pthread_t ubeacon_InteractThread;
383 pthread_t urecovery_InteractThread;
384 pthread_attr_t rxServer_tattr;
385 pthread_attr_t ubeacon_Interact_tattr;
386 pthread_attr_t urecovery_Interact_tattr;
389 extern int rx_stackSize;
393 struct rx_securityClass *secClass;
396 struct rx_service *tservice;
398 initialize_U_error_table();
400 tdb = (struct ubik_dbase *)malloc(sizeof(struct ubik_dbase));
401 tdb->pathName = (char *)malloc(strlen(pathName) + 1);
402 strcpy(tdb->pathName, pathName);
403 tdb->activeTrans = (struct ubik_trans *)0;
404 memset(&tdb->version, 0, sizeof(struct ubik_version));
405 memset(&tdb->cachedVersion, 0, sizeof(struct ubik_version));
406 #ifdef AFS_PTHREAD_ENV
407 MUTEX_INIT(&tdb->versionLock, "version lock", MUTEX_DEFAULT, 0);
408 MUTEX_INIT(&beacon_globals.beacon_lock, "beacon lock", MUTEX_DEFAULT, 0);
410 Lock_Init(&tdb->versionLock);
412 Lock_Init(&tdb->cache_lock);
414 tdb->read = uphys_read;
415 tdb->write = uphys_write;
416 tdb->truncate = uphys_truncate;
417 tdb->open = uphys_invalidate; /* this function isn't used any more */
418 tdb->sync = uphys_sync;
419 tdb->stat = uphys_stat;
420 tdb->getlabel = uphys_getlabel;
421 tdb->setlabel = uphys_setlabel;
422 tdb->getnfiles = uphys_getnfiles;
424 tdb->tidCounter = tdb->writeTidCounter = 0;
426 ubik_dbase = tdb; /* for now, only one db per server; can fix later when we have names for the other dbases */
428 #ifdef AFS_PTHREAD_ENV
429 CV_INIT(&tdb->version_cond, "version", CV_DEFAULT, 0);
430 CV_INIT(&tdb->flags_cond, "flags", CV_DEFAULT, 0);
431 #endif /* AFS_PTHREAD_ENV */
435 /* the following call is idempotent so when/if it got called earlier,
436 * by whatever called us, it doesn't really matter -- klm */
437 code = rx_Init(myPort);
441 udisk_Init(ubik_nBuffers);
447 code = urecovery_Initialize(tdb);
451 code = ubeacon_InitServerListByInfo(myHost, info, clones);
453 code = ubeacon_InitServerList(myHost, serverList);
457 ubik_callPortal = myPort;
458 /* try to get an additional security object */
459 if (buildSecClassesProc == NULL) {
461 ubik_sc = calloc(numClasses, sizeof(struct rx_securityClass *));
462 ubik_sc[0] = rxnull_NewServerSecurityObject();
463 if (ubik_SRXSecurityProc) {
464 code = (*ubik_SRXSecurityProc) (ubik_SRXSecurityRock,
468 ubik_sc[secIndex] = secClass;
472 (*buildSecClassesProc) (securityRock, &ubik_sc, &numClasses);
474 /* for backwards compat this should keep working as it does now
478 rx_NewService(0, VOTE_SERVICE_ID, "VOTE", ubik_sc, numClasses,
479 VOTE_ExecuteRequest);
480 if (tservice == (struct rx_service *)0) {
481 ubik_dprint("Could not create VOTE rx service!\n");
484 rx_SetMinProcs(tservice, 2);
485 rx_SetMaxProcs(tservice, 3);
488 rx_NewService(0, DISK_SERVICE_ID, "DISK", ubik_sc, numClasses,
489 DISK_ExecuteRequest);
490 if (tservice == (struct rx_service *)0) {
491 ubik_dprint("Could not create DISK rx service!\n");
494 rx_SetMinProcs(tservice, 2);
495 rx_SetMaxProcs(tservice, 3);
497 /* start an rx_ServerProc to handle incoming RPC's in particular the
498 * UpdateInterfaceAddr RPC that occurs in ubeacon_InitServerList. This avoids
499 * the "steplock" problem in ubik initialization. Defect 11037.
501 #ifdef AFS_PTHREAD_ENV
502 ubik_thread_create(&rxServer_tattr, &rxServerThread, (void *)rx_ServerProc);
504 LWP_CreateProcess(rx_ServerProc, rx_stackSize, RX_PROCESS_PRIORITY,
505 NULL, "rx_ServerProc", &junk);
508 /* now start up async processes */
509 #ifdef AFS_PTHREAD_ENV
510 ubik_thread_create(&ubeacon_Interact_tattr, &ubeacon_InteractThread,
511 (void *)ubeacon_Interact);
513 code = LWP_CreateProcess(ubeacon_Interact, 16384 /*8192 */ ,
514 LWP_MAX_PRIORITY - 1, (void *)0, "beacon",
520 #ifdef AFS_PTHREAD_ENV
521 ubik_thread_create(&urecovery_Interact_tattr, &urecovery_InteractThread,
522 (void *)urecovery_Interact);
523 return 0; /* is this correct? - klm */
525 code = LWP_CreateProcess(urecovery_Interact, 16384 /*8192 */ ,
526 LWP_MAX_PRIORITY - 1, (void *)0, "recovery",
534 * \see ubik_ServerInitCommon()
537 ubik_ServerInitByInfo(afs_uint32 myHost, short myPort,
538 struct afsconf_cell *info, char clones[],
539 const char *pathName, struct ubik_dbase **dbase)
544 ubik_ServerInitCommon(myHost, myPort, info, clones, 0, pathName,
550 * \see ubik_ServerInitCommon()
553 ubik_ServerInit(afs_uint32 myHost, short myPort, afs_uint32 serverList[],
554 const char *pathName, struct ubik_dbase **dbase)
559 ubik_ServerInitCommon(myHost, myPort, (struct afsconf_cell *)0, 0,
560 serverList, pathName, dbase);
565 * \brief This routine begins a read or write transaction on the transaction
566 * identified by transPtr, in the dbase named by dbase.
568 * An open mode of ubik_READTRANS identifies this as a read transaction,
569 * while a mode of ubik_WRITETRANS identifies this as a write transaction.
570 * transPtr is set to the returned transaction control block.
571 * The readAny flag is set to 0 or 1 or 2 by the wrapper functions
572 * ubik_BeginTrans() or ubik_BeginTransReadAny() or
573 * ubik_BeginTransReadAnyWrite() below.
575 * \note We can only begin transaction when we have an up-to-date database.
578 BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
579 struct ubik_trans **transPtr, int readAny)
581 struct ubik_trans *jt;
582 struct ubik_trans *tt;
585 if (readAny > 1 && ubik_SyncWriterCacheProc == NULL) {
586 /* it's not safe to use ubik_BeginTransReadAnyWrite without a
587 * cache-syncing function; fall back to ubik_BeginTransReadAny,
588 * which is safe but slower */
589 ubik_print("ubik_BeginTransReadAnyWrite called, but "
590 "ubik_SyncWriterCacheProc not set; pretending "
591 "ubik_BeginTransReadAny was called instead\n");
595 if ((transMode != UBIK_READTRANS) && readAny)
598 if (urecovery_AllBetter(dbase, readAny) == 0) {
602 /* otherwise we have a quorum, use it */
604 /* make sure that at most one write transaction occurs at any one time. This
605 * has nothing to do with transaction locking; that's enforced by the lock package. However,
606 * we can't even handle two non-conflicting writes, since our log and recovery modules
607 * don't know how to restore one without possibly picking up some data from the other. */
608 if (transMode == UBIK_WRITETRANS) {
609 /* if we're writing already, wait */
610 while (dbase->flags & DBWRITING) {
611 #ifdef AFS_PTHREAD_ENV
612 CV_WAIT(&dbase->flags_cond, &dbase->versionLock);
615 LWP_WaitProcess(&dbase->flags);
620 if (!ubeacon_AmSyncSite()) {
626 /* create the transaction */
627 code = udisk_begin(dbase, transMode, &jt); /* can't take address of register var */
628 tt = jt; /* move to a register */
629 if (code || tt == (struct ubik_trans *)NULL) {
634 tt->flags |= TRREADANY;
636 tt->flags |= TRREADWRITE;
639 /* label trans and dbase with new tid */
640 tt->tid.epoch = ubik_epochTime;
641 /* bump by two, since tidCounter+1 means trans id'd by tidCounter has finished */
642 tt->tid.counter = (dbase->tidCounter += 2);
644 if (transMode == UBIK_WRITETRANS) {
645 /* for a write trans, we have to keep track of the write tid counter too */
646 dbase->writeTidCounter = tt->tid.counter;
648 /* next try to start transaction on appropriate number of machines */
649 code = ContactQuorum_NoArguments(DISK_Begin, tt, 0);
651 /* we must abort the operation */
653 ContactQuorum_NoArguments(DISK_Abort, tt, 0); /* force aborts to the others */
669 ubik_BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
670 struct ubik_trans **transPtr)
672 return BeginTrans(dbase, transMode, transPtr, 0);
679 ubik_BeginTransReadAny(struct ubik_dbase *dbase, afs_int32 transMode,
680 struct ubik_trans **transPtr)
682 return BeginTrans(dbase, transMode, transPtr, 1);
689 ubik_BeginTransReadAnyWrite(struct ubik_dbase *dbase, afs_int32 transMode,
690 struct ubik_trans **transPtr)
692 return BeginTrans(dbase, transMode, transPtr, 2);
696 * \brief This routine ends a read or write transaction by aborting it.
699 ubik_AbortTrans(struct ubik_trans *transPtr)
703 struct ubik_dbase *dbase;
705 dbase = transPtr->dbase;
707 if (transPtr->flags & TRCACHELOCKED) {
708 ReleaseReadLock(&dbase->cache_lock);
709 transPtr->flags &= ~TRCACHELOCKED;
712 ObtainWriteLock(&dbase->cache_lock);
715 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
717 ReleaseWriteLock(&dbase->cache_lock);
719 /* see if we're still up-to-date */
720 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
721 udisk_abort(transPtr);
727 if (transPtr->type == UBIK_READTRANS) {
728 code = udisk_abort(transPtr);
734 /* below here, we know we're doing a write transaction */
735 if (!ubeacon_AmSyncSite()) {
736 udisk_abort(transPtr);
742 /* now it is safe to try remote abort */
743 code = ContactQuorum_NoArguments(DISK_Abort, transPtr, 0);
744 code2 = udisk_abort(transPtr);
747 return (code ? code : code2);
751 WritebackApplicationCache(struct ubik_dbase *dbase)
754 if (ubik_SyncWriterCacheProc) {
755 code = ubik_SyncWriterCacheProc();
758 /* we failed to sync the local cache, so just invalidate the cache;
759 * we'll try to read the cache in again on the next read */
760 memset(&dbase->cachedVersion, 0, sizeof(dbase->cachedVersion));
762 memcpy(&dbase->cachedVersion, &dbase->version,
763 sizeof(dbase->cachedVersion));
768 * \brief This routine ends a read or write transaction on the open transaction identified by transPtr.
769 * \return an error code.
772 ubik_EndTrans(struct ubik_trans *transPtr)
777 struct ubik_server *ts;
780 struct ubik_dbase *dbase;
782 if (transPtr->type == UBIK_WRITETRANS) {
783 code = ubik_Flush(transPtr);
785 ubik_AbortTrans(transPtr);
790 dbase = transPtr->dbase;
792 if (transPtr->flags & TRCACHELOCKED) {
793 ReleaseReadLock(&dbase->cache_lock);
794 transPtr->flags &= ~TRCACHELOCKED;
797 if (transPtr->type != UBIK_READTRANS) {
798 /* must hold cache_lock before DBHOLD'ing */
799 ObtainWriteLock(&dbase->cache_lock);
805 /* give up if no longer current */
806 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
807 udisk_abort(transPtr);
814 if (transPtr->type == UBIK_READTRANS) { /* reads are easy */
815 code = udisk_commit(transPtr);
817 goto success; /* update cachedVersion correctly */
823 if (!ubeacon_AmSyncSite()) { /* no longer sync site */
824 udisk_abort(transPtr);
831 /* now it is safe to do commit */
832 code = udisk_commit(transPtr);
834 /* db data has been committed locally; update the local cache so
835 * readers can get at it */
836 WritebackApplicationCache(dbase);
838 ReleaseWriteLock(&dbase->cache_lock);
840 code = ContactQuorum_NoArguments(DISK_Commit, transPtr, CStampVersion);
843 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
844 ReleaseWriteLock(&dbase->cache_lock);
848 /* failed to commit, so must return failure. Try to clear locks first, just for fun
849 * Note that we don't know if this transaction will eventually commit at this point.
850 * If it made it to a site that will be present in the next quorum, we win, otherwise
851 * we lose. If we contact a majority of sites, then we won't be here: contacting
852 * a majority guarantees commit, since it guarantees that one dude will be a
853 * member of the next quorum. */
854 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0);
859 /* before we can start sending unlock messages, we must wait until all servers
860 * that are possibly still functioning on the other side of a network partition
861 * have timed out. Check the server structures, compute how long to wait, then
862 * start the unlocks */
863 realStart = FT_ApproxTime();
865 /* wait for all servers to time out */
867 now = FT_ApproxTime();
868 /* check if we're still sync site, the guy should either come up
869 * to us, or timeout. Put safety check in anyway */
870 if (now - realStart > 10 * BIGTIME) {
871 ubik_stats.escapes++;
872 ubik_print("ubik escaping from commit wait\n");
875 for (ts = ubik_servers; ts; ts = ts->next) {
877 if (!ts->beaconSinceDown && now <= ts->lastBeaconSent + BIGTIME) {
880 /* this guy could have some damaged data, wait for him */
882 tv.tv_sec = 1; /* try again after a while (ha ha) */
885 #ifdef AFS_PTHREAD_ENV
886 /* we could release the dbase outside of the loop, but we do
887 * it here, in the loop, to avoid an unnecessary RELE/HOLD
888 * if all sites are up */
890 select(0, 0, 0, 0, &tv);
893 IOMGR_Select(0, 0, 0, 0, &tv); /* poll, should we wait on something? */
901 break; /* no down ones still pseudo-active */
904 /* finally, unlock all the dudes. We can return success independent of the number of servers
905 * that really unlock the dbase; the others will do it if/when they elect a new sync site.
906 * The transaction is committed anyway, since we succeeded in contacting a quorum
907 * at the start (when invoking the DiskCommit function).
909 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0);
913 /* don't update cachedVersion here; it should have been updated way back
914 * in ubik_CheckCache, and earlier in this function for writes */
917 ReleaseWriteLock(&dbase->cache_lock);
923 ObtainWriteLock(&dbase->cache_lock);
925 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
926 ReleaseWriteLock(&dbase->cache_lock);
931 * \brief This routine reads length bytes into buffer from the current position in the database.
933 * 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.
935 * \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.
938 ubik_Read(struct ubik_trans *transPtr, void *buffer,
943 /* reads are easy to do: handle locally */
944 DBHOLD(transPtr->dbase);
945 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
946 DBRELE(transPtr->dbase);
951 udisk_read(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
954 transPtr->seekPos += length;
956 DBRELE(transPtr->dbase);
961 * \brief This routine will flush the io data in the iovec structures.
963 * It first flushes to the local disk and then uses ContactQuorum to write it
964 * to the other servers.
967 ubik_Flush(struct ubik_trans *transPtr)
969 afs_int32 code, error = 0;
971 if (transPtr->type != UBIK_WRITETRANS)
973 if (!transPtr->iovec_info.iovec_wrt_len
974 || !transPtr->iovec_info.iovec_wrt_val)
977 DBHOLD(transPtr->dbase);
978 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
979 ERROR_EXIT(UNOQUORUM);
980 if (!ubeacon_AmSyncSite()) /* only sync site can write */
981 ERROR_EXIT(UNOTSYNC);
983 /* Update the rest of the servers in the quorum */
985 ContactQuorum_DISK_WriteV(transPtr, 0, &transPtr->iovec_info,
986 &transPtr->iovec_data);
988 udisk_abort(transPtr);
989 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0); /* force aborts to the others */
990 transPtr->iovec_info.iovec_wrt_len = 0;
991 transPtr->iovec_data.iovec_buf_len = 0;
995 /* Wrote the buffers out, so start at scratch again */
996 transPtr->iovec_info.iovec_wrt_len = 0;
997 transPtr->iovec_data.iovec_buf_len = 0;
1000 DBRELE(transPtr->dbase);
1005 ubik_Write(struct ubik_trans *transPtr, void *vbuffer,
1008 struct ubik_iovec *iovec;
1009 afs_int32 code, error = 0;
1010 afs_int32 pos, len, size;
1011 char * buffer = (char *)vbuffer;
1013 if (transPtr->type != UBIK_WRITETRANS)
1018 if (length > IOVEC_MAXBUF) {
1019 for (pos = 0, len = length; len > 0; len -= size, pos += size) {
1020 size = ((len < IOVEC_MAXBUF) ? len : IOVEC_MAXBUF);
1021 code = ubik_Write(transPtr, buffer+pos, size);
1028 if (!transPtr->iovec_info.iovec_wrt_val) {
1029 transPtr->iovec_info.iovec_wrt_len = 0;
1030 transPtr->iovec_info.iovec_wrt_val =
1031 (struct ubik_iovec *)malloc(IOVEC_MAXWRT *
1032 sizeof(struct ubik_iovec));
1033 transPtr->iovec_data.iovec_buf_len = 0;
1034 transPtr->iovec_data.iovec_buf_val = (char *)malloc(IOVEC_MAXBUF);
1035 if (!transPtr->iovec_info.iovec_wrt_val
1036 || !transPtr->iovec_data.iovec_buf_val) {
1037 if (transPtr->iovec_info.iovec_wrt_val)
1038 free(transPtr->iovec_info.iovec_wrt_val);
1039 transPtr->iovec_info.iovec_wrt_val = 0;
1040 if (transPtr->iovec_data.iovec_buf_val)
1041 free(transPtr->iovec_data.iovec_buf_val);
1042 transPtr->iovec_data.iovec_buf_val = 0;
1047 /* If this write won't fit in the structure, then flush it out and start anew */
1048 if ((transPtr->iovec_info.iovec_wrt_len >= IOVEC_MAXWRT)
1049 || ((length + transPtr->iovec_data.iovec_buf_len) > IOVEC_MAXBUF)) {
1050 code = ubik_Flush(transPtr);
1055 DBHOLD(transPtr->dbase);
1056 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1057 ERROR_EXIT(UNOQUORUM);
1058 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1059 ERROR_EXIT(UNOTSYNC);
1061 /* Write to the local disk */
1063 udisk_write(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
1066 udisk_abort(transPtr);
1067 transPtr->iovec_info.iovec_wrt_len = 0;
1068 transPtr->iovec_data.iovec_buf_len = 0;
1069 DBRELE(transPtr->dbase);
1073 /* Collect writes for the other ubik servers (to be done in bulk) */
1074 iovec = (struct ubik_iovec *)transPtr->iovec_info.iovec_wrt_val;
1075 iovec[transPtr->iovec_info.iovec_wrt_len].file = transPtr->seekFile;
1076 iovec[transPtr->iovec_info.iovec_wrt_len].position = transPtr->seekPos;
1077 iovec[transPtr->iovec_info.iovec_wrt_len].length = length;
1079 memcpy(&transPtr->iovec_data.
1080 iovec_buf_val[transPtr->iovec_data.iovec_buf_len], buffer, length);
1082 transPtr->iovec_info.iovec_wrt_len++;
1083 transPtr->iovec_data.iovec_buf_len += length;
1084 transPtr->seekPos += length;
1087 DBRELE(transPtr->dbase);
1092 * \brief This sets the file pointer associated with the current transaction
1093 * to the appropriate file and byte position.
1095 * Unlike Unix files, a transaction is labelled by both a file number \p fileid
1096 * and a byte position relative to the specified file \p position.
1099 ubik_Seek(struct ubik_trans *transPtr, afs_int32 fileid,
1104 DBHOLD(transPtr->dbase);
1105 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
1108 transPtr->seekFile = fileid;
1109 transPtr->seekPos = position;
1112 DBRELE(transPtr->dbase);
1117 * \brief This call returns the file pointer associated with the specified
1118 * transaction in \p fileid and \p position.
1121 ubik_Tell(struct ubik_trans *transPtr, afs_int32 * fileid,
1122 afs_int32 * position)
1124 DBHOLD(transPtr->dbase);
1125 *fileid = transPtr->seekFile;
1126 *position = transPtr->seekPos;
1127 DBRELE(transPtr->dbase);
1132 * \brief This sets the file size for the currently-selected file to \p length
1133 * bytes, if length is less than the file's current size.
1136 ubik_Truncate(struct ubik_trans *transPtr, afs_int32 length)
1138 afs_int32 code, error = 0;
1140 /* Will also catch if not UBIK_WRITETRANS */
1141 code = ubik_Flush(transPtr);
1145 DBHOLD(transPtr->dbase);
1146 /* first, check that quorum is still good, and that dbase is up-to-date */
1147 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1148 ERROR_EXIT(UNOQUORUM);
1149 if (!ubeacon_AmSyncSite())
1150 ERROR_EXIT(UNOTSYNC);
1152 /* now do the operation locally, and propagate it out */
1153 code = udisk_truncate(transPtr, transPtr->seekFile, length);
1156 ContactQuorum_DISK_Truncate(transPtr, 0, transPtr->seekFile,
1160 /* we must abort the operation */
1161 udisk_abort(transPtr);
1162 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0); /* force aborts to the others */
1167 DBRELE(transPtr->dbase);
1172 * \brief set a lock; all locks are released on transaction end (commit/abort)
1175 ubik_SetLock(struct ubik_trans *atrans, afs_int32 apos, afs_int32 alen,
1178 afs_int32 code = 0, error = 0;
1180 if (atype == LOCKWRITE) {
1181 if (atrans->type == UBIK_READTRANS)
1183 code = ubik_Flush(atrans);
1188 DBHOLD(atrans->dbase);
1189 if (atype == LOCKREAD) {
1190 code = ulock_getLock(atrans, atype, 1);
1194 /* first, check that quorum is still good, and that dbase is up-to-date */
1195 if (!urecovery_AllBetter(atrans->dbase, atrans->flags & TRREADANY))
1196 ERROR_EXIT(UNOQUORUM);
1197 if (!ubeacon_AmSyncSite())
1198 ERROR_EXIT(UNOTSYNC);
1200 /* now do the operation locally, and propagate it out */
1201 code = ulock_getLock(atrans, atype, 1);
1203 code = ContactQuorum_DISK_Lock(atrans, 0, 0, 1 /*unused */ ,
1204 1 /*unused */ , LOCKWRITE);
1207 /* we must abort the operation */
1208 udisk_abort(atrans);
1209 ContactQuorum_NoArguments(DISK_Abort, atrans, 0); /* force aborts to the others */
1215 DBRELE(atrans->dbase);
1220 * \brief utility to wait for a version # to change
1223 ubik_WaitVersion(struct ubik_dbase *adatabase,
1224 struct ubik_version *aversion)
1228 /* wait until version # changes, and then return */
1229 if (vcmp(*aversion, adatabase->version) != 0) {
1233 #ifdef AFS_PTHREAD_ENV
1234 CV_WAIT(&adatabase->version_cond, &adatabase->versionLock);
1237 LWP_WaitProcess(&adatabase->version); /* same vers, just wait */
1244 * \brief utility to get the version of the dbase a transaction is dealing with
1247 ubik_GetVersion(struct ubik_trans *atrans,
1248 struct ubik_version *avers)
1250 *avers = atrans->dbase->version;
1255 * \brief Facility to simplify database caching.
1256 * \return zero if last trans was done on the local server and was successful.
1257 * \return -1 means bad (NULL) argument.
1259 * If return value is non-zero and the caller is a server caching part of the
1260 * Ubik database, it should invalidate that cache.
1263 ubik_CacheUpdate(struct ubik_trans *atrans)
1265 if (!(atrans && atrans->dbase))
1267 return vcmp(atrans->dbase->cachedVersion, atrans->dbase->version) != 0;
1271 * check and possibly update cache of ubik db.
1273 * If the version of the cached db data is out of date, this calls (*check) to
1274 * update the cache. If (*check) returns success, we update the version of the
1277 * Checking the version of the cached db data is done under a read lock;
1278 * updating the cache (and thus calling (*check)) is done under a write lock
1279 * so is guaranteed not to interfere with another thread's (*check). On
1280 * successful return, a read lock on the cached db data is obtained, which
1281 * will be released by ubik_EndTrans or ubik_AbortTrans.
1283 * @param[in] atrans ubik transaction
1284 * @param[in] check function to call to check/update cache
1285 * @param[in] rock rock to pass to *check
1287 * @return operation status
1289 * @retval nonzero error; cachedVersion not updated
1291 * @post On success, application cache is read-locked, and cache data is
1295 ubik_CheckCache(struct ubik_trans *atrans, ubik_updatecache_func cbf, void *rock)
1299 if (!(atrans && atrans->dbase))
1302 ObtainReadLock(&atrans->dbase->cache_lock);
1304 while (ubik_CacheUpdate(atrans) != 0) {
1306 ReleaseReadLock(&atrans->dbase->cache_lock);
1307 ObtainSharedLock(&atrans->dbase->cache_lock);
1309 if (ubik_CacheUpdate(atrans) != 0) {
1311 BoostSharedLock(&atrans->dbase->cache_lock);
1313 ret = (*cbf) (atrans, rock);
1315 memcpy(&atrans->dbase->cachedVersion, &atrans->dbase->version,
1316 sizeof(atrans->dbase->cachedVersion));
1320 /* It would be nice if we could convert from a shared lock to a read
1321 * lock... instead, just release the shared and acquire the read */
1322 ReleaseSharedLock(&atrans->dbase->cache_lock);
1325 /* if we have an error, don't retry, and don't hold any locks */
1329 ObtainReadLock(&atrans->dbase->cache_lock);
1332 atrans->flags |= TRCACHELOCKED;
1338 * "Who said anything about panicking?" snapped Arthur.
1339 * "This is still just the culture shock. You wait till I've settled down
1340 * into the situation and found my bearings. \em Then I'll start panicking!"
1343 * \returns There is no return from panic.
1346 panic(char *format, ...)
1350 va_start(ap, format);
1351 ubik_print("Ubik PANIC: ");
1352 ubik_vprint(format, ap);
1356 ubik_print("BACK FROM ABORT\n"); /* shouldn't come back */
1357 exit(1); /* never know, though */
1361 * This function takes an IP addresses as its parameter. It returns the
1362 * the primary IP address that is on the host passed in, or 0 if not found.
1365 ubikGetPrimaryInterfaceAddr(afs_uint32 addr)
1367 struct ubik_server *ts;
1370 for (ts = ubik_servers; ts; ts = ts->next)
1371 for (j = 0; j < UBIK_MAX_INTERFACE_ADDR; j++)
1372 if (ts->addr[j] == addr)
1373 return ts->addr[0]; /* net byte order */
1374 return 0; /* if not in server database, return error */
1378 ubik_CheckAuth(struct rx_call *acall)
1380 if (checkSecurityProc)
1381 return (*checkSecurityProc) (securityRock, acall);
1382 else if (ubik_CheckRXSecurityProc) {
1383 return (*ubik_CheckRXSecurityProc) (ubik_CheckRXSecurityRock, acall);
1389 ubik_SetServerSecurityProcs(void (*buildproc) (void *,
1390 struct rx_securityClass ***,
1392 int (*checkproc) (void *, struct rx_call *),
1395 buildSecClassesProc = buildproc;
1396 checkSecurityProc = checkproc;
1397 securityRock = rock;