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>
28 #define UBIK_INTERNALS
32 #include <lwp.h> /* temporary hack by klm */
34 #define ERROR_EXIT(code) do { \
41 * This system is organized in a hierarchical set of related modules. Modules
42 * at one level can only call modules at the same level or below.
44 * At the bottom level (0) we have R, RFTP, LWP and IOMGR, i.e. the basic
45 * operating system primitives.
47 * At the next level (1) we have
49 * \li VOTER--The module responsible for casting votes when asked. It is also
50 * responsible for determining whether this server should try to become
51 * a synchronization site.
52 * \li BEACONER--The module responsible for sending keep-alives out when a
53 * server is actually the sync site, or trying to become a sync site.
54 * \li DISK--The module responsible for representing atomic transactions
55 * on the local disk. It maintains a new-value only log.
56 * \li LOCK--The module responsible for locking byte ranges in the database file.
58 * At the next level (2) we have
60 * \li RECOVERY--The module responsible for ensuring that all members of a quorum
61 * have the same up-to-date database after a new synchronization site is
62 * elected. This module runs only on the synchronization site.
64 * At the next level (3) we have
66 * \li REMOTE--The module responsible for interpreting requests from the sync
67 * site and applying them to the database, after obtaining the appropriate
70 * At the next level (4) we have
72 * \li UBIK--The module users call to perform operations on the database.
77 afs_int32 ubik_quorum = 0;
78 struct ubik_dbase *ubik_dbase = 0;
79 struct ubik_stats ubik_stats;
80 afs_uint32 ubik_host[UBIK_MAX_INTERFACE_ADDR];
81 afs_int32 urecovery_state = 0;
82 int (*ubik_SyncWriterCacheProc) (void);
83 struct ubik_server *ubik_servers;
84 short ubik_callPortal;
86 /* These global variables were used to control the server security layers.
87 * They are retained for backwards compatibility with legacy callers.
89 * The ubik_SetServerSecurityProcs() interface should be used instead.
92 int (*ubik_SRXSecurityProc) (void *, struct rx_securityClass **, afs_int32 *);
93 void *ubik_SRXSecurityRock;
94 int (*ubik_CheckRXSecurityProc) (void *, struct rx_call *);
95 void *ubik_CheckRXSecurityRock;
99 static int BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
100 struct ubik_trans **transPtr, int readAny);
102 static struct rx_securityClass **ubik_sc = NULL;
103 static void (*buildSecClassesProc)(void *, struct rx_securityClass ***,
105 static int (*checkSecurityProc)(void *, struct rx_call *) = NULL;
106 static void *securityRock = NULL;
108 struct version_data version_globals;
110 #define CStampVersion 1 /* meaning set ts->version */
112 static_inline struct rx_connection *
113 Quorum_StartIO(struct ubik_trans *atrans, struct ubik_server *as)
115 struct rx_connection *conn;
118 conn = as->disk_rxcid;
120 #ifdef AFS_PTHREAD_ENV
121 rx_GetConnection(conn);
123 DBRELE(atrans->dbase);
126 #endif /* AFS_PTHREAD_ENV */
132 Quorum_EndIO(struct ubik_trans *atrans, struct rx_connection *aconn)
134 #ifdef AFS_PTHREAD_ENV
135 DBHOLD(atrans->dbase);
136 rx_PutConnection(aconn);
137 #endif /* AFS_PTHREAD_ENV */
142 * Iterate over all servers. Callers pass in *ts which is used to track
143 * the current server.
144 * - Returns 1 if there are no more servers
145 * - Returns 0 with conn set to the connection for the current server if
146 * it's up and current
149 ContactQuorum_iterate(struct ubik_trans *atrans, int aflags, struct ubik_server **ts,
150 struct rx_connection **conn, afs_int32 *rcode,
151 afs_int32 *okcalls, afs_int32 code)
154 /* Initial call - start iterating over servers */
161 Quorum_EndIO(atrans, *conn);
163 if (code) { /* failure */
166 (*ts)->up = 0; /* mark as down now; beacons will no longer be sent */
167 (*ts)->beaconSinceDown = 0;
169 (*ts)->currentDB = 0;
170 urecovery_LostServer(*ts); /* tell recovery to try to resend dbase later */
171 } else { /* success */
173 (*okcalls)++; /* count up how many worked */
174 if (aflags & CStampVersion) {
175 (*ts)->version = atrans->dbase->version;
184 if (!(*ts)->up || !(*ts)->currentDB) {
186 (*ts)->currentDB = 0; /* db is no longer current; we just missed an update */
187 return 0; /* not up-to-date, don't bother. NULL conn will tell caller not to use */
190 *conn = Quorum_StartIO(atrans, *ts);
195 ContactQuorum_rcode(int okcalls, afs_int32 rcode)
198 * return 0 if we successfully contacted a quorum, otherwise return error code.
199 * We don't have to contact ourselves (that was done locally)
201 if (okcalls + 1 >= ubik_quorum)
208 * \brief Perform an operation at a quorum, handling error conditions.
209 * \return 0 if all worked and a quorum was contacted successfully
210 * \return otherwise mark failing server as down and return #UERROR
212 * \note If any server misses an update, we must wait #BIGTIME seconds before
213 * allowing the transaction to commit, to ensure that the missing and
214 * possibly still functioning server times out and stops handing out old
215 * data. This is done in the commit code, where we wait for a server marked
216 * down to have stayed down for #BIGTIME seconds before we allow a transaction
217 * to commit. A server that fails but comes back up won't give out old data
218 * because it is sent the sync count along with the beacon message that
219 * marks it as \b really up (\p beaconSinceDown).
222 ContactQuorum_NoArguments(afs_int32 (*proc)(struct rx_connection *, ubik_tid *),
223 struct ubik_trans *atrans, int aflags)
225 struct ubik_server *ts = NULL;
226 afs_int32 code = 0, rcode, okcalls;
227 struct rx_connection *conn;
230 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
233 code = (*proc)(conn, &atrans->tid);
234 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
236 return ContactQuorum_rcode(okcalls, rcode);
241 ContactQuorum_DISK_Lock(struct ubik_trans *atrans, int aflags,afs_int32 file,
242 afs_int32 position, afs_int32 length, afs_int32 type)
244 struct ubik_server *ts = NULL;
245 afs_int32 code = 0, rcode, okcalls;
246 struct rx_connection *conn;
249 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
252 code = DISK_Lock(conn, &atrans->tid, file, position, length, type);
253 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
255 return ContactQuorum_rcode(okcalls, rcode);
260 ContactQuorum_DISK_Write(struct ubik_trans *atrans, int aflags,
261 afs_int32 file, afs_int32 position, bulkdata *data)
263 struct ubik_server *ts = NULL;
264 afs_int32 code = 0, rcode, okcalls;
265 struct rx_connection *conn;
268 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
271 code = DISK_Write(conn, &atrans->tid, file, position, data);
272 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
274 return ContactQuorum_rcode(okcalls, rcode);
279 ContactQuorum_DISK_Truncate(struct ubik_trans *atrans, int aflags,
280 afs_int32 file, afs_int32 length)
282 struct ubik_server *ts = NULL;
283 afs_int32 code = 0, rcode, okcalls;
284 struct rx_connection *conn;
287 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
290 code = DISK_Truncate(conn, &atrans->tid, file, length);
291 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
293 return ContactQuorum_rcode(okcalls, rcode);
298 ContactQuorum_DISK_WriteV(struct ubik_trans *atrans, int aflags,
299 iovec_wrt * io_vector, iovec_buf *io_buffer)
301 struct ubik_server *ts = NULL;
302 afs_int32 code = 0, rcode, okcalls;
303 struct rx_connection *conn;
306 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
309 code = DISK_WriteV(conn, &atrans->tid, io_vector, io_buffer);
310 if ((code <= -450) && (code > -500)) {
311 /* An RPC interface mismatch (as defined in comerr/error_msg.c).
312 * Un-bulk the entries and do individual DISK_Write calls
313 * instead of DISK_WriteV.
315 struct ubik_iovec *iovec =
316 (struct ubik_iovec *)io_vector->iovec_wrt_val;
317 char *iobuf = (char *)io_buffer->iovec_buf_val;
321 for (i = 0, offset = 0; i < io_vector->iovec_wrt_len; i++) {
322 /* Sanity check for going off end of buffer */
323 if ((offset + iovec[i].length) > io_buffer->iovec_buf_len) {
327 tcbs.bulkdata_len = iovec[i].length;
328 tcbs.bulkdata_val = &iobuf[offset];
329 code = DISK_Write(conn, &atrans->tid, iovec[i].file,
330 iovec[i].position, &tcbs);
333 offset += iovec[i].length;
337 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
339 return ContactQuorum_rcode(okcalls, rcode);
344 ContactQuorum_DISK_SetVersion(struct ubik_trans *atrans, int aflags,
345 ubik_version *OldVersion,
346 ubik_version *NewVersion)
348 struct ubik_server *ts = NULL;
349 afs_int32 code = 0, rcode, okcalls;
350 struct rx_connection *conn;
353 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
356 code = DISK_SetVersion(conn, &atrans->tid, OldVersion, NewVersion);
357 done = ContactQuorum_iterate(atrans, aflags, &ts, &conn, &rcode, &okcalls, code);
359 return ContactQuorum_rcode(okcalls, rcode);
362 #if defined(AFS_PTHREAD_ENV)
364 ubik_thread_create(pthread_attr_t *tattr, pthread_t *thread, void *proc) {
365 opr_Verify(pthread_attr_init(tattr) == 0);
366 opr_Verify(pthread_attr_setdetachstate(tattr,
367 PTHREAD_CREATE_DETACHED) == 0);
368 opr_Verify(pthread_create(thread, tattr, proc, NULL) == 0);
374 * \brief This routine initializes the ubik system for a set of servers.
375 * \return 0 for success, or an error code on failure.
376 * \param serverList set of servers specified; nServers gives the number of entries in this array.
377 * \param pathName provides an initial prefix used for naming storage files used by this system.
378 * \param dbase the returned structure representing this instance of an ubik; it is passed to various calls below.
380 * \todo This routine should perhaps be generalized to a low-level disk interface providing read, write, file enumeration and sync operations.
382 * \warning The host named by myHost should not also be listed in serverList.
384 * \see ubik_ServerInit(), ubik_ServerInitByInfo()
387 ubik_ServerInitCommon(afs_uint32 myHost, short myPort,
388 struct afsconf_cell *info, char clones[],
389 afs_uint32 serverList[], const char *pathName,
390 struct ubik_dbase **dbase)
392 struct ubik_dbase *tdb;
394 #ifdef AFS_PTHREAD_ENV
395 pthread_t rxServerThread; /* pthread variables */
396 pthread_t ubeacon_InteractThread;
397 pthread_t urecovery_InteractThread;
398 pthread_attr_t rxServer_tattr;
399 pthread_attr_t ubeacon_Interact_tattr;
400 pthread_attr_t urecovery_Interact_tattr;
403 extern int rx_stackSize;
407 struct rx_securityClass *secClass;
410 struct rx_service *tservice;
412 initialize_U_error_table();
414 tdb = malloc(sizeof(struct ubik_dbase));
415 tdb->pathName = strdup(pathName);
416 tdb->activeTrans = (struct ubik_trans *)0;
417 memset(&tdb->version, 0, sizeof(struct ubik_version));
418 memset(&tdb->cachedVersion, 0, sizeof(struct ubik_version));
419 #ifdef AFS_PTHREAD_ENV
420 opr_mutex_init(&tdb->versionLock);
421 opr_mutex_init(&beacon_globals.beacon_lock);
422 opr_mutex_init(&vote_globals.vote_lock);
423 opr_mutex_init(&addr_globals.addr_lock);
424 opr_mutex_init(&version_globals.version_lock);
426 Lock_Init(&tdb->versionLock);
428 Lock_Init(&tdb->cache_lock);
430 tdb->read = uphys_read;
431 tdb->write = uphys_write;
432 tdb->truncate = uphys_truncate;
433 tdb->open = uphys_invalidate; /* this function isn't used any more */
434 tdb->sync = uphys_sync;
435 tdb->stat = uphys_stat;
436 tdb->getlabel = uphys_getlabel;
437 tdb->setlabel = uphys_setlabel;
438 tdb->getnfiles = uphys_getnfiles;
440 tdb->tidCounter = tdb->writeTidCounter = 0;
442 ubik_dbase = tdb; /* for now, only one db per server; can fix later when we have names for the other dbases */
444 #ifdef AFS_PTHREAD_ENV
445 opr_cv_init(&tdb->version_cond);
446 opr_cv_init(&tdb->flags_cond);
447 #endif /* AFS_PTHREAD_ENV */
451 /* the following call is idempotent so when/if it got called earlier,
452 * by whatever called us, it doesn't really matter -- klm */
453 code = rx_Init(myPort);
457 ubik_callPortal = myPort;
459 udisk_Init(ubik_nBuffers);
465 code = urecovery_Initialize(tdb);
469 code = ubeacon_InitServerListByInfo(myHost, info, clones);
471 code = ubeacon_InitServerList(myHost, serverList);
475 /* try to get an additional security object */
476 if (buildSecClassesProc == NULL) {
478 ubik_sc = calloc(numClasses, sizeof(struct rx_securityClass *));
479 ubik_sc[0] = rxnull_NewServerSecurityObject();
480 if (ubik_SRXSecurityProc) {
481 code = (*ubik_SRXSecurityProc) (ubik_SRXSecurityRock,
485 ubik_sc[secIndex] = secClass;
489 (*buildSecClassesProc) (securityRock, &ubik_sc, &numClasses);
491 /* for backwards compat this should keep working as it does now
495 rx_NewService(0, VOTE_SERVICE_ID, "VOTE", ubik_sc, numClasses,
496 VOTE_ExecuteRequest);
497 if (tservice == (struct rx_service *)0) {
498 ubik_dprint("Could not create VOTE rx service!\n");
501 rx_SetMinProcs(tservice, 2);
502 rx_SetMaxProcs(tservice, 3);
505 rx_NewService(0, DISK_SERVICE_ID, "DISK", ubik_sc, numClasses,
506 DISK_ExecuteRequest);
507 if (tservice == (struct rx_service *)0) {
508 ubik_dprint("Could not create DISK rx service!\n");
511 rx_SetMinProcs(tservice, 2);
512 rx_SetMaxProcs(tservice, 3);
514 /* start an rx_ServerProc to handle incoming RPC's in particular the
515 * UpdateInterfaceAddr RPC that occurs in ubeacon_InitServerList. This avoids
516 * the "steplock" problem in ubik initialization. Defect 11037.
518 #ifdef AFS_PTHREAD_ENV
519 ubik_thread_create(&rxServer_tattr, &rxServerThread, (void *)rx_ServerProc);
521 LWP_CreateProcess(rx_ServerProc, rx_stackSize, RX_PROCESS_PRIORITY,
522 NULL, "rx_ServerProc", &junk);
525 /* send addrs to all other servers */
526 code = ubeacon_updateUbikNetworkAddress(ubik_host);
530 /* now start up async processes */
531 #ifdef AFS_PTHREAD_ENV
532 ubik_thread_create(&ubeacon_Interact_tattr, &ubeacon_InteractThread,
533 (void *)ubeacon_Interact);
535 code = LWP_CreateProcess(ubeacon_Interact, 16384 /*8192 */ ,
536 LWP_MAX_PRIORITY - 1, (void *)0, "beacon",
542 #ifdef AFS_PTHREAD_ENV
543 ubik_thread_create(&urecovery_Interact_tattr, &urecovery_InteractThread,
544 (void *)urecovery_Interact);
545 return 0; /* is this correct? - klm */
547 code = LWP_CreateProcess(urecovery_Interact, 16384 /*8192 */ ,
548 LWP_MAX_PRIORITY - 1, (void *)0, "recovery",
556 * \see ubik_ServerInitCommon()
559 ubik_ServerInitByInfo(afs_uint32 myHost, short myPort,
560 struct afsconf_cell *info, char clones[],
561 const char *pathName, struct ubik_dbase **dbase)
566 ubik_ServerInitCommon(myHost, myPort, info, clones, 0, pathName,
572 * \see ubik_ServerInitCommon()
575 ubik_ServerInit(afs_uint32 myHost, short myPort, afs_uint32 serverList[],
576 const char *pathName, struct ubik_dbase **dbase)
581 ubik_ServerInitCommon(myHost, myPort, (struct afsconf_cell *)0, 0,
582 serverList, pathName, dbase);
587 * \brief This routine begins a read or write transaction on the transaction
588 * identified by transPtr, in the dbase named by dbase.
590 * An open mode of ubik_READTRANS identifies this as a read transaction,
591 * while a mode of ubik_WRITETRANS identifies this as a write transaction.
592 * transPtr is set to the returned transaction control block.
593 * The readAny flag is set to 0 or 1 or 2 by the wrapper functions
594 * ubik_BeginTrans() or ubik_BeginTransReadAny() or
595 * ubik_BeginTransReadAnyWrite() below.
597 * \note We can only begin transaction when we have an up-to-date database.
600 BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
601 struct ubik_trans **transPtr, int readAny)
603 struct ubik_trans *jt;
604 struct ubik_trans *tt;
607 if (readAny > 1 && ubik_SyncWriterCacheProc == NULL) {
608 /* it's not safe to use ubik_BeginTransReadAnyWrite without a
609 * cache-syncing function; fall back to ubik_BeginTransReadAny,
610 * which is safe but slower */
611 ubik_print("ubik_BeginTransReadAnyWrite called, but "
612 "ubik_SyncWriterCacheProc not set; pretending "
613 "ubik_BeginTransReadAny was called instead\n");
617 if ((transMode != UBIK_READTRANS) && readAny)
620 if (urecovery_AllBetter(dbase, readAny) == 0) {
624 /* otherwise we have a quorum, use it */
626 /* make sure that at most one write transaction occurs at any one time. This
627 * has nothing to do with transaction locking; that's enforced by the lock package. However,
628 * we can't even handle two non-conflicting writes, since our log and recovery modules
629 * don't know how to restore one without possibly picking up some data from the other. */
630 if (transMode == UBIK_WRITETRANS) {
631 /* if we're writing already, wait */
632 while (dbase->flags & DBWRITING) {
633 #ifdef AFS_PTHREAD_ENV
634 opr_cv_wait(&dbase->flags_cond, &dbase->versionLock);
637 LWP_WaitProcess(&dbase->flags);
642 if (!ubeacon_AmSyncSite()) {
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, 0);
678 /* we must abort the operation */
680 ContactQuorum_NoArguments(DISK_Abort, tt, 0); /* force aborts to the others */
696 ubik_BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
697 struct ubik_trans **transPtr)
699 return BeginTrans(dbase, transMode, transPtr, 0);
706 ubik_BeginTransReadAny(struct ubik_dbase *dbase, afs_int32 transMode,
707 struct ubik_trans **transPtr)
709 return BeginTrans(dbase, transMode, transPtr, 1);
716 ubik_BeginTransReadAnyWrite(struct ubik_dbase *dbase, afs_int32 transMode,
717 struct ubik_trans **transPtr)
719 return BeginTrans(dbase, transMode, transPtr, 2);
723 * \brief This routine ends a read or write transaction by aborting it.
726 ubik_AbortTrans(struct ubik_trans *transPtr)
730 struct ubik_dbase *dbase;
732 dbase = transPtr->dbase;
734 if (transPtr->flags & TRCACHELOCKED) {
735 ReleaseReadLock(&dbase->cache_lock);
736 transPtr->flags &= ~TRCACHELOCKED;
739 ObtainWriteLock(&dbase->cache_lock);
742 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
744 ReleaseWriteLock(&dbase->cache_lock);
746 /* see if we're still up-to-date */
747 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
748 udisk_abort(transPtr);
754 if (transPtr->type == UBIK_READTRANS) {
755 code = udisk_abort(transPtr);
761 /* below here, we know we're doing a write transaction */
762 if (!ubeacon_AmSyncSite()) {
763 udisk_abort(transPtr);
769 /* now it is safe to try remote abort */
770 code = ContactQuorum_NoArguments(DISK_Abort, transPtr, 0);
771 code2 = udisk_abort(transPtr);
774 return (code ? code : code2);
778 WritebackApplicationCache(struct ubik_dbase *dbase)
781 if (ubik_SyncWriterCacheProc) {
782 code = ubik_SyncWriterCacheProc();
785 /* we failed to sync the local cache, so just invalidate the cache;
786 * we'll try to read the cache in again on the next read */
787 memset(&dbase->cachedVersion, 0, sizeof(dbase->cachedVersion));
789 memcpy(&dbase->cachedVersion, &dbase->version,
790 sizeof(dbase->cachedVersion));
795 * \brief This routine ends a read or write transaction on the open transaction identified by transPtr.
796 * \return an error code.
799 ubik_EndTrans(struct ubik_trans *transPtr)
804 struct ubik_server *ts;
807 struct ubik_dbase *dbase;
809 if (transPtr->type == UBIK_WRITETRANS) {
810 code = ubik_Flush(transPtr);
812 ubik_AbortTrans(transPtr);
817 dbase = transPtr->dbase;
819 if (transPtr->flags & TRCACHELOCKED) {
820 ReleaseReadLock(&dbase->cache_lock);
821 transPtr->flags &= ~TRCACHELOCKED;
824 if (transPtr->type != UBIK_READTRANS) {
825 /* must hold cache_lock before DBHOLD'ing */
826 ObtainWriteLock(&dbase->cache_lock);
832 /* give up if no longer current */
833 if (!urecovery_AllBetter(dbase, transPtr->flags & TRREADANY)) {
834 udisk_abort(transPtr);
841 if (transPtr->type == UBIK_READTRANS) { /* reads are easy */
842 code = udisk_commit(transPtr);
844 goto success; /* update cachedVersion correctly */
850 if (!ubeacon_AmSyncSite()) { /* no longer sync site */
851 udisk_abort(transPtr);
858 /* now it is safe to do commit */
859 code = udisk_commit(transPtr);
861 /* db data has been committed locally; update the local cache so
862 * readers can get at it */
863 WritebackApplicationCache(dbase);
865 ReleaseWriteLock(&dbase->cache_lock);
867 code = ContactQuorum_NoArguments(DISK_Commit, transPtr, CStampVersion);
870 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
871 ReleaseWriteLock(&dbase->cache_lock);
875 /* failed to commit, so must return failure. Try to clear locks first, just for fun
876 * Note that we don't know if this transaction will eventually commit at this point.
877 * If it made it to a site that will be present in the next quorum, we win, otherwise
878 * we lose. If we contact a majority of sites, then we won't be here: contacting
879 * a majority guarantees commit, since it guarantees that one dude will be a
880 * member of the next quorum. */
881 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0);
886 /* before we can start sending unlock messages, we must wait until all servers
887 * that are possibly still functioning on the other side of a network partition
888 * have timed out. Check the server structures, compute how long to wait, then
889 * start the unlocks */
890 realStart = FT_ApproxTime();
892 /* wait for all servers to time out */
894 now = FT_ApproxTime();
895 /* check if we're still sync site, the guy should either come up
896 * to us, or timeout. Put safety check in anyway */
897 if (now - realStart > 10 * BIGTIME) {
898 ubik_stats.escapes++;
899 ubik_print("ubik escaping from commit wait\n");
902 for (ts = ubik_servers; ts; ts = ts->next) {
904 if (!ts->beaconSinceDown && now <= ts->lastBeaconSent + BIGTIME) {
907 /* this guy could have some damaged data, wait for him */
909 tv.tv_sec = 1; /* try again after a while (ha ha) */
912 #ifdef AFS_PTHREAD_ENV
913 /* we could release the dbase outside of the loop, but we do
914 * it here, in the loop, to avoid an unnecessary RELE/HOLD
915 * if all sites are up */
917 select(0, 0, 0, 0, &tv);
920 IOMGR_Select(0, 0, 0, 0, &tv); /* poll, should we wait on something? */
928 break; /* no down ones still pseudo-active */
931 /* finally, unlock all the dudes. We can return success independent of the number of servers
932 * that really unlock the dbase; the others will do it if/when they elect a new sync site.
933 * The transaction is committed anyway, since we succeeded in contacting a quorum
934 * at the start (when invoking the DiskCommit function).
936 ContactQuorum_NoArguments(DISK_ReleaseLocks, transPtr, 0);
940 /* don't update cachedVersion here; it should have been updated way back
941 * in ubik_CheckCache, and earlier in this function for writes */
944 ReleaseWriteLock(&dbase->cache_lock);
950 ObtainWriteLock(&dbase->cache_lock);
952 memset(&dbase->cachedVersion, 0, sizeof(struct ubik_version));
953 ReleaseWriteLock(&dbase->cache_lock);
958 * \brief This routine reads length bytes into buffer from the current position in the database.
960 * 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.
962 * \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.
965 ubik_Read(struct ubik_trans *transPtr, void *buffer,
970 /* reads are easy to do: handle locally */
971 DBHOLD(transPtr->dbase);
972 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
973 DBRELE(transPtr->dbase);
978 udisk_read(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
981 transPtr->seekPos += length;
983 DBRELE(transPtr->dbase);
988 * \brief This routine will flush the io data in the iovec structures.
990 * It first flushes to the local disk and then uses ContactQuorum to write it
991 * to the other servers.
994 ubik_Flush(struct ubik_trans *transPtr)
996 afs_int32 code, error = 0;
998 if (transPtr->type != UBIK_WRITETRANS)
1001 DBHOLD(transPtr->dbase);
1002 if (!transPtr->iovec_info.iovec_wrt_len
1003 || !transPtr->iovec_info.iovec_wrt_val) {
1004 DBRELE(transPtr->dbase);
1008 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1009 ERROR_EXIT(UNOQUORUM);
1010 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1011 ERROR_EXIT(UNOTSYNC);
1013 /* Update the rest of the servers in the quorum */
1015 ContactQuorum_DISK_WriteV(transPtr, 0, &transPtr->iovec_info,
1016 &transPtr->iovec_data);
1018 udisk_abort(transPtr);
1019 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0); /* force aborts to the others */
1020 transPtr->iovec_info.iovec_wrt_len = 0;
1021 transPtr->iovec_data.iovec_buf_len = 0;
1025 /* Wrote the buffers out, so start at scratch again */
1026 transPtr->iovec_info.iovec_wrt_len = 0;
1027 transPtr->iovec_data.iovec_buf_len = 0;
1030 DBRELE(transPtr->dbase);
1035 ubik_Write(struct ubik_trans *transPtr, void *vbuffer,
1038 struct ubik_iovec *iovec;
1039 afs_int32 code, error = 0;
1040 afs_int32 pos, len, size;
1041 char * buffer = (char *)vbuffer;
1043 if (transPtr->type != UBIK_WRITETRANS)
1048 if (length > IOVEC_MAXBUF) {
1049 for (pos = 0, len = length; len > 0; len -= size, pos += size) {
1050 size = ((len < IOVEC_MAXBUF) ? len : IOVEC_MAXBUF);
1051 code = ubik_Write(transPtr, buffer+pos, size);
1058 DBHOLD(transPtr->dbase);
1059 if (!transPtr->iovec_info.iovec_wrt_val) {
1060 transPtr->iovec_info.iovec_wrt_len = 0;
1061 transPtr->iovec_info.iovec_wrt_val =
1062 malloc(IOVEC_MAXWRT * sizeof(struct ubik_iovec));
1063 transPtr->iovec_data.iovec_buf_len = 0;
1064 transPtr->iovec_data.iovec_buf_val = malloc(IOVEC_MAXBUF);
1065 if (!transPtr->iovec_info.iovec_wrt_val
1066 || !transPtr->iovec_data.iovec_buf_val) {
1067 if (transPtr->iovec_info.iovec_wrt_val)
1068 free(transPtr->iovec_info.iovec_wrt_val);
1069 transPtr->iovec_info.iovec_wrt_val = 0;
1070 if (transPtr->iovec_data.iovec_buf_val)
1071 free(transPtr->iovec_data.iovec_buf_val);
1072 transPtr->iovec_data.iovec_buf_val = 0;
1073 DBRELE(transPtr->dbase);
1078 /* If this write won't fit in the structure, then flush it out and start anew */
1079 if ((transPtr->iovec_info.iovec_wrt_len >= IOVEC_MAXWRT)
1080 || ((length + transPtr->iovec_data.iovec_buf_len) > IOVEC_MAXBUF)) {
1081 /* Can't hold the DB lock over ubik_Flush */
1082 DBRELE(transPtr->dbase);
1083 code = ubik_Flush(transPtr);
1086 DBHOLD(transPtr->dbase);
1089 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1090 ERROR_EXIT(UNOQUORUM);
1091 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1092 ERROR_EXIT(UNOTSYNC);
1094 /* Write to the local disk */
1096 udisk_write(transPtr, transPtr->seekFile, buffer, transPtr->seekPos,
1099 udisk_abort(transPtr);
1100 transPtr->iovec_info.iovec_wrt_len = 0;
1101 transPtr->iovec_data.iovec_buf_len = 0;
1102 DBRELE(transPtr->dbase);
1106 /* Collect writes for the other ubik servers (to be done in bulk) */
1107 iovec = (struct ubik_iovec *)transPtr->iovec_info.iovec_wrt_val;
1108 iovec[transPtr->iovec_info.iovec_wrt_len].file = transPtr->seekFile;
1109 iovec[transPtr->iovec_info.iovec_wrt_len].position = transPtr->seekPos;
1110 iovec[transPtr->iovec_info.iovec_wrt_len].length = length;
1112 memcpy(&transPtr->iovec_data.
1113 iovec_buf_val[transPtr->iovec_data.iovec_buf_len], buffer, length);
1115 transPtr->iovec_info.iovec_wrt_len++;
1116 transPtr->iovec_data.iovec_buf_len += length;
1117 transPtr->seekPos += length;
1120 DBRELE(transPtr->dbase);
1125 * \brief This sets the file pointer associated with the current transaction
1126 * to the appropriate file and byte position.
1128 * Unlike Unix files, a transaction is labelled by both a file number \p fileid
1129 * and a byte position relative to the specified file \p position.
1132 ubik_Seek(struct ubik_trans *transPtr, afs_int32 fileid,
1137 DBHOLD(transPtr->dbase);
1138 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY)) {
1141 transPtr->seekFile = fileid;
1142 transPtr->seekPos = position;
1145 DBRELE(transPtr->dbase);
1150 * \brief This call returns the file pointer associated with the specified
1151 * transaction in \p fileid and \p position.
1154 ubik_Tell(struct ubik_trans *transPtr, afs_int32 * fileid,
1155 afs_int32 * position)
1157 DBHOLD(transPtr->dbase);
1158 *fileid = transPtr->seekFile;
1159 *position = transPtr->seekPos;
1160 DBRELE(transPtr->dbase);
1165 * \brief This sets the file size for the currently-selected file to \p length
1166 * bytes, if length is less than the file's current size.
1169 ubik_Truncate(struct ubik_trans *transPtr, afs_int32 length)
1171 afs_int32 code, error = 0;
1173 /* Will also catch if not UBIK_WRITETRANS */
1174 code = ubik_Flush(transPtr);
1178 DBHOLD(transPtr->dbase);
1179 /* first, check that quorum is still good, and that dbase is up-to-date */
1180 if (!urecovery_AllBetter(transPtr->dbase, transPtr->flags & TRREADANY))
1181 ERROR_EXIT(UNOQUORUM);
1182 if (!ubeacon_AmSyncSite())
1183 ERROR_EXIT(UNOTSYNC);
1185 /* now do the operation locally, and propagate it out */
1186 code = udisk_truncate(transPtr, transPtr->seekFile, length);
1189 ContactQuorum_DISK_Truncate(transPtr, 0, transPtr->seekFile,
1193 /* we must abort the operation */
1194 udisk_abort(transPtr);
1195 ContactQuorum_NoArguments(DISK_Abort, transPtr, 0); /* force aborts to the others */
1200 DBRELE(transPtr->dbase);
1205 * \brief set a lock; all locks are released on transaction end (commit/abort)
1208 ubik_SetLock(struct ubik_trans *atrans, afs_int32 apos, afs_int32 alen,
1211 afs_int32 code = 0, error = 0;
1213 if (atype == LOCKWRITE) {
1214 if (atrans->type == UBIK_READTRANS)
1216 code = ubik_Flush(atrans);
1221 DBHOLD(atrans->dbase);
1222 if (atype == LOCKREAD) {
1223 code = ulock_getLock(atrans, atype, 1);
1227 /* first, check that quorum is still good, and that dbase is up-to-date */
1228 if (!urecovery_AllBetter(atrans->dbase, atrans->flags & TRREADANY))
1229 ERROR_EXIT(UNOQUORUM);
1230 if (!ubeacon_AmSyncSite())
1231 ERROR_EXIT(UNOTSYNC);
1233 /* now do the operation locally, and propagate it out */
1234 code = ulock_getLock(atrans, atype, 1);
1236 code = ContactQuorum_DISK_Lock(atrans, 0, 0, 1 /*unused */ ,
1237 1 /*unused */ , LOCKWRITE);
1240 /* we must abort the operation */
1241 udisk_abort(atrans);
1242 ContactQuorum_NoArguments(DISK_Abort, atrans, 0); /* force aborts to the others */
1248 DBRELE(atrans->dbase);
1253 * \brief utility to wait for a version # to change
1256 ubik_WaitVersion(struct ubik_dbase *adatabase,
1257 struct ubik_version *aversion)
1261 /* wait until version # changes, and then return */
1262 if (vcmp(*aversion, adatabase->version) != 0) {
1266 #ifdef AFS_PTHREAD_ENV
1267 opr_cv_wait(&adatabase->version_cond, &adatabase->versionLock);
1270 LWP_WaitProcess(&adatabase->version); /* same vers, just wait */
1277 * \brief utility to get the version of the dbase a transaction is dealing with
1280 ubik_GetVersion(struct ubik_trans *atrans,
1281 struct ubik_version *avers)
1283 DBHOLD(atrans->dbase);
1284 *avers = atrans->dbase->version;
1285 DBRELE(atrans->dbase);
1290 * \brief Facility to simplify database caching.
1291 * \return zero if last trans was done on the local server and was successful.
1292 * \return -1 means bad (NULL) argument.
1294 * If return value is non-zero and the caller is a server caching part of the
1295 * Ubik database, it should invalidate that cache.
1298 ubik_CacheUpdate(struct ubik_trans *atrans)
1300 if (!(atrans && atrans->dbase))
1302 return vcmp(atrans->dbase->cachedVersion, atrans->dbase->version) != 0;
1306 * check and possibly update cache of ubik db.
1308 * If the version of the cached db data is out of date, this calls (*check) to
1309 * update the cache. If (*check) returns success, we update the version of the
1312 * Checking the version of the cached db data is done under a read lock;
1313 * updating the cache (and thus calling (*check)) is done under a write lock
1314 * so is guaranteed not to interfere with another thread's (*check). On
1315 * successful return, a read lock on the cached db data is obtained, which
1316 * will be released by ubik_EndTrans or ubik_AbortTrans.
1318 * @param[in] atrans ubik transaction
1319 * @param[in] check function to call to check/update cache
1320 * @param[in] rock rock to pass to *check
1322 * @return operation status
1324 * @retval nonzero error; cachedVersion not updated
1326 * @post On success, application cache is read-locked, and cache data is
1330 ubik_CheckCache(struct ubik_trans *atrans, ubik_updatecache_func cbf, void *rock)
1334 if (!(atrans && atrans->dbase))
1337 ObtainReadLock(&atrans->dbase->cache_lock);
1339 while (ubik_CacheUpdate(atrans) != 0) {
1341 ReleaseReadLock(&atrans->dbase->cache_lock);
1342 ObtainSharedLock(&atrans->dbase->cache_lock);
1344 if (ubik_CacheUpdate(atrans) != 0) {
1346 BoostSharedLock(&atrans->dbase->cache_lock);
1348 ret = (*cbf) (atrans, rock);
1350 memcpy(&atrans->dbase->cachedVersion, &atrans->dbase->version,
1351 sizeof(atrans->dbase->cachedVersion));
1355 /* It would be nice if we could convert from a shared lock to a read
1356 * lock... instead, just release the shared and acquire the read */
1357 ReleaseSharedLock(&atrans->dbase->cache_lock);
1360 /* if we have an error, don't retry, and don't hold any locks */
1364 ObtainReadLock(&atrans->dbase->cache_lock);
1367 atrans->flags |= TRCACHELOCKED;
1373 * "Who said anything about panicking?" snapped Arthur.
1374 * "This is still just the culture shock. You wait till I've settled down
1375 * into the situation and found my bearings. \em Then I'll start panicking!"
1378 * \returns There is no return from panic.
1381 panic(char *format, ...)
1385 va_start(ap, format);
1386 ubik_print("Ubik PANIC:\n");
1387 ubik_vprint(format, ap);
1391 ubik_print("BACK FROM ABORT\n"); /* shouldn't come back */
1392 exit(1); /* never know, though */
1396 * This function takes an IP addresses as its parameter. It returns the
1397 * the primary IP address that is on the host passed in, or 0 if not found.
1400 ubikGetPrimaryInterfaceAddr(afs_uint32 addr)
1402 struct ubik_server *ts;
1406 for (ts = ubik_servers; ts; ts = ts->next)
1407 for (j = 0; j < UBIK_MAX_INTERFACE_ADDR; j++)
1408 if (ts->addr[j] == addr) {
1410 return ts->addr[0]; /* net byte order */
1413 return 0; /* if not in server database, return error */
1417 ubik_CheckAuth(struct rx_call *acall)
1419 if (checkSecurityProc)
1420 return (*checkSecurityProc) (securityRock, acall);
1421 else if (ubik_CheckRXSecurityProc) {
1422 return (*ubik_CheckRXSecurityProc) (ubik_CheckRXSecurityRock, acall);
1428 ubik_SetServerSecurityProcs(void (*buildproc) (void *,
1429 struct rx_securityClass ***,
1431 int (*checkproc) (void *, struct rx_call *),
1434 buildSecClassesProc = buildproc;
1435 checkSecurityProc = checkproc;
1436 securityRock = rock;