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
);
264 ContactQuorum_DISK_Truncate(struct ubik_trans
*atrans
, int aflags
,
265 afs_int32 file
, afs_int32 length
)
267 struct ubik_server
*ts
= NULL
;
268 afs_int32 code
= 0, rcode
, okcalls
;
269 struct rx_connection
*conn
;
272 done
= ContactQuorum_iterate(atrans
, aflags
, &ts
, &conn
, &rcode
, &okcalls
, code
);
275 code
= DISK_Truncate(conn
, &atrans
->tid
, file
, length
);
276 done
= ContactQuorum_iterate(atrans
, aflags
, &ts
, &conn
, &rcode
, &okcalls
, code
);
278 return ContactQuorum_rcode(okcalls
, rcode
);
283 ContactQuorum_DISK_WriteV(struct ubik_trans
*atrans
, int aflags
,
284 iovec_wrt
* io_vector
, iovec_buf
*io_buffer
)
286 struct ubik_server
*ts
= NULL
;
287 afs_int32 code
= 0, rcode
, okcalls
;
288 struct rx_connection
*conn
;
291 done
= ContactQuorum_iterate(atrans
, aflags
, &ts
, &conn
, &rcode
, &okcalls
, code
);
294 code
= DISK_WriteV(conn
, &atrans
->tid
, io_vector
, io_buffer
);
295 if ((code
<= -450) && (code
> -500)) {
296 /* An RPC interface mismatch (as defined in comerr/error_msg.c).
297 * Un-bulk the entries and do individual DISK_Write calls
298 * instead of DISK_WriteV.
300 struct ubik_iovec
*iovec
=
301 (struct ubik_iovec
*)io_vector
->iovec_wrt_val
;
302 char *iobuf
= (char *)io_buffer
->iovec_buf_val
;
306 for (i
= 0, offset
= 0; i
< io_vector
->iovec_wrt_len
; i
++) {
307 /* Sanity check for going off end of buffer */
308 if ((offset
+ iovec
[i
].length
) > io_buffer
->iovec_buf_len
) {
312 tcbs
.bulkdata_len
= iovec
[i
].length
;
313 tcbs
.bulkdata_val
= &iobuf
[offset
];
314 code
= DISK_Write(conn
, &atrans
->tid
, iovec
[i
].file
,
315 iovec
[i
].position
, &tcbs
);
318 offset
+= iovec
[i
].length
;
322 done
= ContactQuorum_iterate(atrans
, aflags
, &ts
, &conn
, &rcode
, &okcalls
, code
);
324 return ContactQuorum_rcode(okcalls
, rcode
);
329 ContactQuorum_DISK_SetVersion(struct ubik_trans
*atrans
, int aflags
,
330 ubik_version
*OldVersion
,
331 ubik_version
*NewVersion
)
333 struct ubik_server
*ts
= NULL
;
334 afs_int32 code
= 0, rcode
, okcalls
;
335 struct rx_connection
*conn
;
338 done
= ContactQuorum_iterate(atrans
, aflags
, &ts
, &conn
, &rcode
, &okcalls
, code
);
341 code
= DISK_SetVersion(conn
, &atrans
->tid
, OldVersion
, NewVersion
);
342 done
= ContactQuorum_iterate(atrans
, aflags
, &ts
, &conn
, &rcode
, &okcalls
, code
);
344 return ContactQuorum_rcode(okcalls
, rcode
);
347 #if defined(AFS_PTHREAD_ENV)
349 ubik_thread_create(pthread_attr_t
*tattr
, pthread_t
*thread
, void *proc
) {
350 opr_Verify(pthread_attr_init(tattr
) == 0);
351 opr_Verify(pthread_attr_setdetachstate(tattr
,
352 PTHREAD_CREATE_DETACHED
) == 0);
353 opr_Verify(pthread_create(thread
, tattr
, proc
, NULL
) == 0);
359 * \brief This routine initializes the ubik system for a set of servers.
360 * \return 0 for success, or an error code on failure.
361 * \param serverList set of servers specified; nServers gives the number of entries in this array.
362 * \param pathName provides an initial prefix used for naming storage files used by this system.
363 * \param dbase the returned structure representing this instance of an ubik; it is passed to various calls below.
365 * \todo This routine should perhaps be generalized to a low-level disk interface providing read, write, file enumeration and sync operations.
367 * \warning The host named by myHost should not also be listed in serverList.
369 * \see ubik_ServerInit(), ubik_ServerInitByInfo()
372 ubik_ServerInitCommon(afs_uint32 myHost
, short myPort
,
373 struct afsconf_cell
*info
, char clones
[],
374 afs_uint32 serverList
[], const char *pathName
,
375 struct ubik_dbase
**dbase
)
377 struct ubik_dbase
*tdb
;
379 #ifdef AFS_PTHREAD_ENV
380 pthread_t rxServerThread
; /* pthread variables */
381 pthread_t ubeacon_InteractThread
;
382 pthread_t urecovery_InteractThread
;
383 pthread_attr_t rxServer_tattr
;
384 pthread_attr_t ubeacon_Interact_tattr
;
385 pthread_attr_t urecovery_Interact_tattr
;
388 extern int rx_stackSize
;
392 struct rx_securityClass
*secClass
;
395 struct rx_service
*tservice
;
397 initialize_U_error_table();
399 tdb
= malloc(sizeof(struct ubik_dbase
));
400 tdb
->pathName
= strdup(pathName
);
401 tdb
->activeTrans
= (struct ubik_trans
*)0;
402 memset(&tdb
->version
, 0, sizeof(struct ubik_version
));
403 memset(&tdb
->cachedVersion
, 0, sizeof(struct ubik_version
));
404 #ifdef AFS_PTHREAD_ENV
405 opr_mutex_init(&tdb
->versionLock
);
406 opr_mutex_init(&beacon_globals
.beacon_lock
);
407 opr_mutex_init(&vote_globals
.vote_lock
);
408 opr_mutex_init(&addr_globals
.addr_lock
);
409 opr_mutex_init(&version_globals
.version_lock
);
411 Lock_Init(&tdb
->versionLock
);
413 Lock_Init(&tdb
->cache_lock
);
415 tdb
->read
= uphys_read
;
416 tdb
->write
= uphys_write
;
417 tdb
->truncate
= uphys_truncate
;
418 tdb
->open
= uphys_invalidate
; /* this function isn't used any more */
419 tdb
->sync
= uphys_sync
;
420 tdb
->stat
= uphys_stat
;
421 tdb
->getlabel
= uphys_getlabel
;
422 tdb
->setlabel
= uphys_setlabel
;
423 tdb
->getnfiles
= uphys_getnfiles
;
424 tdb
->buffered_append
= uphys_buf_append
;
426 tdb
->tidCounter
= tdb
->writeTidCounter
= 0;
428 ubik_dbase
= tdb
; /* for now, only one db per server; can fix later when we have names for the other dbases */
430 #ifdef AFS_PTHREAD_ENV
431 opr_cv_init(&tdb
->version_cond
);
432 opr_cv_init(&tdb
->flags_cond
);
433 #endif /* AFS_PTHREAD_ENV */
437 /* the following call is idempotent so when/if it got called earlier,
438 * by whatever called us, it doesn't really matter -- klm */
439 code
= rx_Init(myPort
);
443 ubik_callPortal
= myPort
;
445 udisk_Init(ubik_nBuffers
);
451 code
= urecovery_Initialize(tdb
);
455 code
= ubeacon_InitServerListByInfo(myHost
, info
, clones
);
457 code
= ubeacon_InitServerList(myHost
, serverList
);
461 /* try to get an additional security object */
462 if (buildSecClassesProc
== NULL
) {
464 ubik_sc
= calloc(numClasses
, sizeof(struct rx_securityClass
*));
465 ubik_sc
[0] = rxnull_NewServerSecurityObject();
466 if (ubik_SRXSecurityProc
) {
467 code
= (*ubik_SRXSecurityProc
) (ubik_SRXSecurityRock
,
471 ubik_sc
[secIndex
] = secClass
;
475 (*buildSecClassesProc
) (securityRock
, &ubik_sc
, &numClasses
);
477 /* for backwards compat this should keep working as it does now
481 rx_NewService(0, VOTE_SERVICE_ID
, "VOTE", ubik_sc
, numClasses
,
482 VOTE_ExecuteRequest
);
483 if (tservice
== (struct rx_service
*)0) {
484 ViceLog(5, ("Could not create VOTE rx service!\n"));
487 rx_SetMinProcs(tservice
, 2);
488 rx_SetMaxProcs(tservice
, 3);
491 rx_NewService(0, DISK_SERVICE_ID
, "DISK", ubik_sc
, numClasses
,
492 DISK_ExecuteRequest
);
493 if (tservice
== (struct rx_service
*)0) {
494 ViceLog(5, ("Could not create DISK rx service!\n"));
497 rx_SetMinProcs(tservice
, 2);
498 rx_SetMaxProcs(tservice
, 3);
500 /* start an rx_ServerProc to handle incoming RPC's in particular the
501 * UpdateInterfaceAddr RPC that occurs in ubeacon_InitServerList. This avoids
502 * the "steplock" problem in ubik initialization. Defect 11037.
504 #ifdef AFS_PTHREAD_ENV
505 ubik_thread_create(&rxServer_tattr
, &rxServerThread
, (void *)rx_ServerProc
);
507 LWP_CreateProcess(rx_ServerProc
, rx_stackSize
, RX_PROCESS_PRIORITY
,
508 NULL
, "rx_ServerProc", &junk
);
511 /* send addrs to all other servers */
512 code
= ubeacon_updateUbikNetworkAddress(ubik_host
);
516 /* now start up async processes */
517 #ifdef AFS_PTHREAD_ENV
518 ubik_thread_create(&ubeacon_Interact_tattr
, &ubeacon_InteractThread
,
519 (void *)ubeacon_Interact
);
521 code
= LWP_CreateProcess(ubeacon_Interact
, 16384 /*8192 */ ,
522 LWP_MAX_PRIORITY
- 1, (void *)0, "beacon",
528 #ifdef AFS_PTHREAD_ENV
529 ubik_thread_create(&urecovery_Interact_tattr
, &urecovery_InteractThread
,
530 (void *)urecovery_Interact
);
531 return 0; /* is this correct? - klm */
533 code
= LWP_CreateProcess(urecovery_Interact
, 16384 /*8192 */ ,
534 LWP_MAX_PRIORITY
- 1, (void *)0, "recovery",
542 * \see ubik_ServerInitCommon()
545 ubik_ServerInitByInfo(afs_uint32 myHost
, short myPort
,
546 struct afsconf_cell
*info
, char clones
[],
547 const char *pathName
, struct ubik_dbase
**dbase
)
552 ubik_ServerInitCommon(myHost
, myPort
, info
, clones
, 0, pathName
,
558 * \see ubik_ServerInitCommon()
561 ubik_ServerInit(afs_uint32 myHost
, short myPort
, afs_uint32 serverList
[],
562 const char *pathName
, struct ubik_dbase
**dbase
)
567 ubik_ServerInitCommon(myHost
, myPort
, (struct afsconf_cell
*)0, 0,
568 serverList
, pathName
, dbase
);
573 * \brief This routine begins a read or write transaction on the transaction
574 * identified by transPtr, in the dbase named by dbase.
576 * An open mode of ubik_READTRANS identifies this as a read transaction,
577 * while a mode of ubik_WRITETRANS identifies this as a write transaction.
578 * transPtr is set to the returned transaction control block.
579 * The readAny flag is set to 0 or 1 or 2 by the wrapper functions
580 * ubik_BeginTrans() or ubik_BeginTransReadAny() or
581 * ubik_BeginTransReadAnyWrite() below.
583 * \note We can only begin transaction when we have an up-to-date database.
586 BeginTrans(struct ubik_dbase
*dbase
, afs_int32 transMode
,
587 struct ubik_trans
**transPtr
, int readAny
)
589 struct ubik_trans
*jt
;
590 struct ubik_trans
*tt
;
593 if (readAny
> 1 && ubik_SyncWriterCacheProc
== NULL
) {
594 /* it's not safe to use ubik_BeginTransReadAnyWrite without a
595 * cache-syncing function; fall back to ubik_BeginTransReadAny,
596 * which is safe but slower */
597 ViceLog(0, ("ubik_BeginTransReadAnyWrite called, but "
598 "ubik_SyncWriterCacheProc not set; pretending "
599 "ubik_BeginTransReadAny was called instead\n"));
603 if ((transMode
!= UBIK_READTRANS
) && readAny
)
606 if (urecovery_AllBetter(dbase
, readAny
) == 0) {
610 /* otherwise we have a quorum, use it */
612 /* make sure that at most one write transaction occurs at any one time. This
613 * has nothing to do with transaction locking; that's enforced by the lock package. However,
614 * we can't even handle two non-conflicting writes, since our log and recovery modules
615 * don't know how to restore one without possibly picking up some data from the other. */
616 if (transMode
== UBIK_WRITETRANS
) {
617 /* if we're writing already, wait */
618 while (dbase
->flags
& DBWRITING
) {
619 #ifdef AFS_PTHREAD_ENV
620 opr_cv_wait(&dbase
->flags_cond
, &dbase
->versionLock
);
623 LWP_WaitProcess(&dbase
->flags
);
628 if (!ubeacon_AmSyncSite()) {
632 if (!ubeacon_SyncSiteAdvertised()) {
633 /* i am the sync-site but the remotes are not aware yet */
639 /* create the transaction */
640 code
= udisk_begin(dbase
, transMode
, &jt
); /* can't take address of register var */
641 tt
= jt
; /* move to a register */
642 if (code
|| tt
== NULL
) {
648 tt
->flags
|= TRREADANY
;
650 tt
->flags
|= TRREADWRITE
;
653 /* label trans and dbase with new tid */
654 tt
->tid
.epoch
= version_globals
.ubik_epochTime
;
655 /* bump by two, since tidCounter+1 means trans id'd by tidCounter has finished */
656 tt
->tid
.counter
= (dbase
->tidCounter
+= 2);
658 if (transMode
== UBIK_WRITETRANS
) {
659 /* for a write trans, we have to keep track of the write tid counter too */
660 dbase
->writeTidCounter
= tt
->tid
.counter
;
665 if (transMode
== UBIK_WRITETRANS
) {
666 /* next try to start transaction on appropriate number of machines */
667 code
= ContactQuorum_NoArguments(DISK_Begin
, tt
, CCheckSyncAdvertised
);
669 /* we must abort the operation */
671 ContactQuorum_NoArguments(DISK_Abort
, tt
, 0); /* force aborts to the others */
687 ubik_BeginTrans(struct ubik_dbase
*dbase
, afs_int32 transMode
,
688 struct ubik_trans
**transPtr
)
690 return BeginTrans(dbase
, transMode
, transPtr
, 0);
697 ubik_BeginTransReadAny(struct ubik_dbase
*dbase
, afs_int32 transMode
,
698 struct ubik_trans
**transPtr
)
700 return BeginTrans(dbase
, transMode
, transPtr
, 1);
707 ubik_BeginTransReadAnyWrite(struct ubik_dbase
*dbase
, afs_int32 transMode
,
708 struct ubik_trans
**transPtr
)
710 return BeginTrans(dbase
, transMode
, transPtr
, 2);
714 * \brief This routine ends a read or write transaction by aborting it.
717 ubik_AbortTrans(struct ubik_trans
*transPtr
)
721 struct ubik_dbase
*dbase
;
723 dbase
= transPtr
->dbase
;
725 if (transPtr
->flags
& TRCACHELOCKED
) {
726 ReleaseReadLock(&dbase
->cache_lock
);
727 transPtr
->flags
&= ~TRCACHELOCKED
;
730 ObtainWriteLock(&dbase
->cache_lock
);
733 memset(&dbase
->cachedVersion
, 0, sizeof(struct ubik_version
));
735 ReleaseWriteLock(&dbase
->cache_lock
);
737 /* see if we're still up-to-date */
738 if (!urecovery_AllBetter(dbase
, transPtr
->flags
& TRREADANY
)) {
739 udisk_abort(transPtr
);
745 if (transPtr
->type
== UBIK_READTRANS
) {
746 code
= udisk_abort(transPtr
);
752 /* below here, we know we're doing a write transaction */
753 if (!ubeacon_AmSyncSite()) {
754 udisk_abort(transPtr
);
760 /* now it is safe to try remote abort */
761 code
= ContactQuorum_NoArguments(DISK_Abort
, transPtr
, 0);
762 code2
= udisk_abort(transPtr
);
765 return (code
? code
: code2
);
769 WritebackApplicationCache(struct ubik_dbase
*dbase
)
772 if (ubik_SyncWriterCacheProc
) {
773 code
= ubik_SyncWriterCacheProc();
776 /* we failed to sync the local cache, so just invalidate the cache;
777 * we'll try to read the cache in again on the next read */
778 memset(&dbase
->cachedVersion
, 0, sizeof(dbase
->cachedVersion
));
780 memcpy(&dbase
->cachedVersion
, &dbase
->version
,
781 sizeof(dbase
->cachedVersion
));
786 * \brief This routine ends a read or write transaction on the open transaction identified by transPtr.
787 * \return an error code.
790 ubik_EndTrans(struct ubik_trans
*transPtr
)
795 struct ubik_server
*ts
;
798 struct ubik_dbase
*dbase
;
800 if (transPtr
->type
== UBIK_WRITETRANS
) {
801 code
= ubik_Flush(transPtr
);
803 ubik_AbortTrans(transPtr
);
808 dbase
= transPtr
->dbase
;
810 if (transPtr
->flags
& TRCACHELOCKED
) {
811 ReleaseReadLock(&dbase
->cache_lock
);
812 transPtr
->flags
&= ~TRCACHELOCKED
;
815 if (transPtr
->type
!= UBIK_READTRANS
) {
816 /* must hold cache_lock before DBHOLD'ing */
817 ObtainWriteLock(&dbase
->cache_lock
);
823 /* give up if no longer current */
824 if (!urecovery_AllBetter(dbase
, transPtr
->flags
& TRREADANY
)) {
825 udisk_abort(transPtr
);
832 if (transPtr
->type
== UBIK_READTRANS
) { /* reads are easy */
833 code
= udisk_commit(transPtr
);
835 goto success
; /* update cachedVersion correctly */
841 if (!ubeacon_AmSyncSite()) { /* no longer sync site */
842 udisk_abort(transPtr
);
849 /* now it is safe to do commit */
850 code
= udisk_commit(transPtr
);
852 /* db data has been committed locally; update the local cache so
853 * readers can get at it */
854 WritebackApplicationCache(dbase
);
856 ReleaseWriteLock(&dbase
->cache_lock
);
858 code
= ContactQuorum_NoArguments(DISK_Commit
, transPtr
, CStampVersion
);
861 memset(&dbase
->cachedVersion
, 0, sizeof(struct ubik_version
));
862 ReleaseWriteLock(&dbase
->cache_lock
);
866 /* failed to commit, so must return failure. Try to clear locks first, just for fun
867 * Note that we don't know if this transaction will eventually commit at this point.
868 * If it made it to a site that will be present in the next quorum, we win, otherwise
869 * we lose. If we contact a majority of sites, then we won't be here: contacting
870 * a majority guarantees commit, since it guarantees that one dude will be a
871 * member of the next quorum. */
872 ContactQuorum_NoArguments(DISK_ReleaseLocks
, transPtr
, 0);
877 /* before we can start sending unlock messages, we must wait until all servers
878 * that are possibly still functioning on the other side of a network partition
879 * have timed out. Check the server structures, compute how long to wait, then
880 * start the unlocks */
881 realStart
= FT_ApproxTime();
883 /* wait for all servers to time out */
885 now
= FT_ApproxTime();
886 /* check if we're still sync site, the guy should either come up
887 * to us, or timeout. Put safety check in anyway */
888 if (now
- realStart
> 10 * BIGTIME
) {
889 ubik_stats
.escapes
++;
890 ViceLog(0, ("ubik escaping from commit wait\n"));
893 for (ts
= ubik_servers
; ts
; ts
= ts
->next
) {
895 if (!ts
->beaconSinceDown
&& now
<= ts
->lastBeaconSent
+ BIGTIME
) {
898 /* this guy could have some damaged data, wait for him */
900 tv
.tv_sec
= 1; /* try again after a while (ha ha) */
903 #ifdef AFS_PTHREAD_ENV
904 /* we could release the dbase outside of the loop, but we do
905 * it here, in the loop, to avoid an unnecessary RELE/HOLD
906 * if all sites are up */
908 select(0, 0, 0, 0, &tv
);
911 IOMGR_Select(0, 0, 0, 0, &tv
); /* poll, should we wait on something? */
919 break; /* no down ones still pseudo-active */
922 /* finally, unlock all the dudes. We can return success independent of the number of servers
923 * that really unlock the dbase; the others will do it if/when they elect a new sync site.
924 * The transaction is committed anyway, since we succeeded in contacting a quorum
925 * at the start (when invoking the DiskCommit function).
927 ContactQuorum_NoArguments(DISK_ReleaseLocks
, transPtr
, 0);
931 /* don't update cachedVersion here; it should have been updated way back
932 * in ubik_CheckCache, and earlier in this function for writes */
935 ReleaseWriteLock(&dbase
->cache_lock
);
941 ObtainWriteLock(&dbase
->cache_lock
);
943 memset(&dbase
->cachedVersion
, 0, sizeof(struct ubik_version
));
944 ReleaseWriteLock(&dbase
->cache_lock
);
949 * \brief This routine reads length bytes into buffer from the current position in the database.
951 * 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.
953 * \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.
956 ubik_Read(struct ubik_trans
*transPtr
, void *buffer
,
961 /* reads are easy to do: handle locally */
962 DBHOLD(transPtr
->dbase
);
963 if (!urecovery_AllBetter(transPtr
->dbase
, transPtr
->flags
& TRREADANY
)) {
964 DBRELE(transPtr
->dbase
);
969 udisk_read(transPtr
, transPtr
->seekFile
, buffer
, transPtr
->seekPos
,
972 transPtr
->seekPos
+= length
;
974 DBRELE(transPtr
->dbase
);
979 * \brief This routine will flush the io data in the iovec structures.
981 * It first flushes to the local disk and then uses ContactQuorum to write it
982 * to the other servers.
985 ubik_Flush(struct ubik_trans
*transPtr
)
987 afs_int32 code
, error
= 0;
989 if (transPtr
->type
!= UBIK_WRITETRANS
)
992 DBHOLD(transPtr
->dbase
);
993 if (!transPtr
->iovec_info
.iovec_wrt_len
994 || !transPtr
->iovec_info
.iovec_wrt_val
) {
995 DBRELE(transPtr
->dbase
);
999 if (!urecovery_AllBetter(transPtr
->dbase
, transPtr
->flags
& TRREADANY
))
1000 ERROR_EXIT(UNOQUORUM
);
1001 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1002 ERROR_EXIT(UNOTSYNC
);
1004 /* Update the rest of the servers in the quorum */
1006 ContactQuorum_DISK_WriteV(transPtr
, 0, &transPtr
->iovec_info
,
1007 &transPtr
->iovec_data
);
1009 udisk_abort(transPtr
);
1010 ContactQuorum_NoArguments(DISK_Abort
, transPtr
, 0); /* force aborts to the others */
1011 transPtr
->iovec_info
.iovec_wrt_len
= 0;
1012 transPtr
->iovec_data
.iovec_buf_len
= 0;
1016 /* Wrote the buffers out, so start at scratch again */
1017 transPtr
->iovec_info
.iovec_wrt_len
= 0;
1018 transPtr
->iovec_data
.iovec_buf_len
= 0;
1021 DBRELE(transPtr
->dbase
);
1026 ubik_Write(struct ubik_trans
*transPtr
, void *vbuffer
,
1029 struct ubik_iovec
*iovec
;
1030 afs_int32 code
, error
= 0;
1031 afs_int32 pos
, len
, size
;
1032 char * buffer
= (char *)vbuffer
;
1034 if (transPtr
->type
!= UBIK_WRITETRANS
)
1039 if (length
> IOVEC_MAXBUF
) {
1040 for (pos
= 0, len
= length
; len
> 0; len
-= size
, pos
+= size
) {
1041 size
= ((len
< IOVEC_MAXBUF
) ? len
: IOVEC_MAXBUF
);
1042 code
= ubik_Write(transPtr
, buffer
+pos
, size
);
1049 DBHOLD(transPtr
->dbase
);
1050 if (!transPtr
->iovec_info
.iovec_wrt_val
) {
1051 transPtr
->iovec_info
.iovec_wrt_len
= 0;
1052 transPtr
->iovec_info
.iovec_wrt_val
=
1053 malloc(IOVEC_MAXWRT
* sizeof(struct ubik_iovec
));
1054 transPtr
->iovec_data
.iovec_buf_len
= 0;
1055 transPtr
->iovec_data
.iovec_buf_val
= malloc(IOVEC_MAXBUF
);
1056 if (!transPtr
->iovec_info
.iovec_wrt_val
1057 || !transPtr
->iovec_data
.iovec_buf_val
) {
1058 if (transPtr
->iovec_info
.iovec_wrt_val
)
1059 free(transPtr
->iovec_info
.iovec_wrt_val
);
1060 transPtr
->iovec_info
.iovec_wrt_val
= 0;
1061 if (transPtr
->iovec_data
.iovec_buf_val
)
1062 free(transPtr
->iovec_data
.iovec_buf_val
);
1063 transPtr
->iovec_data
.iovec_buf_val
= 0;
1064 DBRELE(transPtr
->dbase
);
1069 /* If this write won't fit in the structure, then flush it out and start anew */
1070 if ((transPtr
->iovec_info
.iovec_wrt_len
>= IOVEC_MAXWRT
)
1071 || ((length
+ transPtr
->iovec_data
.iovec_buf_len
) > IOVEC_MAXBUF
)) {
1072 /* Can't hold the DB lock over ubik_Flush */
1073 DBRELE(transPtr
->dbase
);
1074 code
= ubik_Flush(transPtr
);
1077 DBHOLD(transPtr
->dbase
);
1080 if (!urecovery_AllBetter(transPtr
->dbase
, transPtr
->flags
& TRREADANY
))
1081 ERROR_EXIT(UNOQUORUM
);
1082 if (!ubeacon_AmSyncSite()) /* only sync site can write */
1083 ERROR_EXIT(UNOTSYNC
);
1085 /* Write to the local disk */
1087 udisk_write(transPtr
, transPtr
->seekFile
, buffer
, transPtr
->seekPos
,
1090 udisk_abort(transPtr
);
1091 transPtr
->iovec_info
.iovec_wrt_len
= 0;
1092 transPtr
->iovec_data
.iovec_buf_len
= 0;
1093 DBRELE(transPtr
->dbase
);
1097 /* Collect writes for the other ubik servers (to be done in bulk) */
1098 iovec
= (struct ubik_iovec
*)transPtr
->iovec_info
.iovec_wrt_val
;
1099 iovec
[transPtr
->iovec_info
.iovec_wrt_len
].file
= transPtr
->seekFile
;
1100 iovec
[transPtr
->iovec_info
.iovec_wrt_len
].position
= transPtr
->seekPos
;
1101 iovec
[transPtr
->iovec_info
.iovec_wrt_len
].length
= length
;
1103 memcpy(&transPtr
->iovec_data
.
1104 iovec_buf_val
[transPtr
->iovec_data
.iovec_buf_len
], buffer
, length
);
1106 transPtr
->iovec_info
.iovec_wrt_len
++;
1107 transPtr
->iovec_data
.iovec_buf_len
+= length
;
1108 transPtr
->seekPos
+= length
;
1111 DBRELE(transPtr
->dbase
);
1116 * \brief This sets the file pointer associated with the current transaction
1117 * to the appropriate file and byte position.
1119 * Unlike Unix files, a transaction is labelled by both a file number \p fileid
1120 * and a byte position relative to the specified file \p position.
1123 ubik_Seek(struct ubik_trans
*transPtr
, afs_int32 fileid
,
1128 DBHOLD(transPtr
->dbase
);
1129 if (!urecovery_AllBetter(transPtr
->dbase
, transPtr
->flags
& TRREADANY
)) {
1132 transPtr
->seekFile
= fileid
;
1133 transPtr
->seekPos
= position
;
1136 DBRELE(transPtr
->dbase
);
1141 * \brief This call returns the file pointer associated with the specified
1142 * transaction in \p fileid and \p position.
1145 ubik_Tell(struct ubik_trans
*transPtr
, afs_int32
* fileid
,
1146 afs_int32
* position
)
1148 DBHOLD(transPtr
->dbase
);
1149 *fileid
= transPtr
->seekFile
;
1150 *position
= transPtr
->seekPos
;
1151 DBRELE(transPtr
->dbase
);
1156 * \brief This sets the file size for the currently-selected file to \p length
1157 * bytes, if length is less than the file's current size.
1160 ubik_Truncate(struct ubik_trans
*transPtr
, afs_int32 length
)
1162 afs_int32 code
, error
= 0;
1164 /* Will also catch if not UBIK_WRITETRANS */
1165 code
= ubik_Flush(transPtr
);
1169 DBHOLD(transPtr
->dbase
);
1170 /* first, check that quorum is still good, and that dbase is up-to-date */
1171 if (!urecovery_AllBetter(transPtr
->dbase
, transPtr
->flags
& TRREADANY
))
1172 ERROR_EXIT(UNOQUORUM
);
1173 if (!ubeacon_AmSyncSite())
1174 ERROR_EXIT(UNOTSYNC
);
1176 /* now do the operation locally, and propagate it out */
1177 code
= udisk_truncate(transPtr
, transPtr
->seekFile
, length
);
1180 ContactQuorum_DISK_Truncate(transPtr
, 0, transPtr
->seekFile
,
1184 /* we must abort the operation */
1185 udisk_abort(transPtr
);
1186 ContactQuorum_NoArguments(DISK_Abort
, transPtr
, 0); /* force aborts to the others */
1191 DBRELE(transPtr
->dbase
);
1196 * \brief set a lock; all locks are released on transaction end (commit/abort)
1199 ubik_SetLock(struct ubik_trans
*atrans
, afs_int32 apos
, afs_int32 alen
,
1202 afs_int32 code
= 0, error
= 0;
1204 if (atype
== LOCKWRITE
) {
1205 if (atrans
->type
== UBIK_READTRANS
)
1207 code
= ubik_Flush(atrans
);
1212 DBHOLD(atrans
->dbase
);
1213 if (atype
== LOCKREAD
) {
1214 code
= ulock_getLock(atrans
, atype
, 1);
1218 /* first, check that quorum is still good, and that dbase is up-to-date */
1219 if (!urecovery_AllBetter(atrans
->dbase
, atrans
->flags
& TRREADANY
))
1220 ERROR_EXIT(UNOQUORUM
);
1221 if (!ubeacon_AmSyncSite())
1222 ERROR_EXIT(UNOTSYNC
);
1224 /* now do the operation locally, and propagate it out */
1225 code
= ulock_getLock(atrans
, atype
, 1);
1227 code
= ContactQuorum_DISK_Lock(atrans
, 0, 0, 1 /*unused */ ,
1228 1 /*unused */ , LOCKWRITE
);
1231 /* we must abort the operation */
1232 udisk_abort(atrans
);
1233 ContactQuorum_NoArguments(DISK_Abort
, atrans
, 0); /* force aborts to the others */
1239 DBRELE(atrans
->dbase
);
1244 * \brief utility to wait for a version # to change
1247 ubik_WaitVersion(struct ubik_dbase
*adatabase
,
1248 struct ubik_version
*aversion
)
1252 /* wait until version # changes, and then return */
1253 if (vcmp(*aversion
, adatabase
->version
) != 0) {
1257 #ifdef AFS_PTHREAD_ENV
1258 opr_cv_wait(&adatabase
->version_cond
, &adatabase
->versionLock
);
1261 LWP_WaitProcess(&adatabase
->version
); /* same vers, just wait */
1268 * \brief utility to get the version of the dbase a transaction is dealing with
1271 ubik_GetVersion(struct ubik_trans
*atrans
,
1272 struct ubik_version
*avers
)
1274 DBHOLD(atrans
->dbase
);
1275 *avers
= atrans
->dbase
->version
;
1276 DBRELE(atrans
->dbase
);
1281 * \brief Facility to simplify database caching.
1282 * \return zero if last trans was done on the local server and was successful.
1283 * \return -1 means bad (NULL) argument.
1285 * If return value is non-zero and the caller is a server caching part of the
1286 * Ubik database, it should invalidate that cache.
1289 ubik_CacheUpdate(struct ubik_trans
*atrans
)
1291 if (!(atrans
&& atrans
->dbase
))
1293 return vcmp(atrans
->dbase
->cachedVersion
, atrans
->dbase
->version
) != 0;
1297 * check and possibly update cache of ubik db.
1299 * If the version of the cached db data is out of date, this calls (*check) to
1300 * update the cache. If (*check) returns success, we update the version of the
1303 * Checking the version of the cached db data is done under a read lock;
1304 * updating the cache (and thus calling (*check)) is done under a write lock
1305 * so is guaranteed not to interfere with another thread's (*check). On
1306 * successful return, a read lock on the cached db data is obtained, which
1307 * will be released by ubik_EndTrans or ubik_AbortTrans.
1309 * @param[in] atrans ubik transaction
1310 * @param[in] check function to call to check/update cache
1311 * @param[in] rock rock to pass to *check
1313 * @return operation status
1315 * @retval nonzero error; cachedVersion not updated
1317 * @post On success, application cache is read-locked, and cache data is
1321 ubik_CheckCache(struct ubik_trans
*atrans
, ubik_updatecache_func cbf
, void *rock
)
1325 if (!(atrans
&& atrans
->dbase
))
1328 ObtainReadLock(&atrans
->dbase
->cache_lock
);
1330 while (ubik_CacheUpdate(atrans
) != 0) {
1332 ReleaseReadLock(&atrans
->dbase
->cache_lock
);
1333 ObtainSharedLock(&atrans
->dbase
->cache_lock
);
1335 if (ubik_CacheUpdate(atrans
) != 0) {
1337 BoostSharedLock(&atrans
->dbase
->cache_lock
);
1339 ret
= (*cbf
) (atrans
, rock
);
1341 memcpy(&atrans
->dbase
->cachedVersion
, &atrans
->dbase
->version
,
1342 sizeof(atrans
->dbase
->cachedVersion
));
1346 /* It would be nice if we could convert from a shared lock to a read
1347 * lock... instead, just release the shared and acquire the read */
1348 ReleaseSharedLock(&atrans
->dbase
->cache_lock
);
1351 /* if we have an error, don't retry, and don't hold any locks */
1355 ObtainReadLock(&atrans
->dbase
->cache_lock
);
1358 atrans
->flags
|= TRCACHELOCKED
;
1364 * "Who said anything about panicking?" snapped Arthur.
1365 * "This is still just the culture shock. You wait till I've settled down
1366 * into the situation and found my bearings. \em Then I'll start panicking!"
1369 * \returns There is no return from panic.
1372 panic(char *format
, ...)
1376 va_start(ap
, format
);
1377 ViceLog(0, ("Ubik PANIC:\n"));
1378 vViceLog(0, (format
, ap
));
1382 ViceLog(0, ("BACK FROM ABORT\n")); /* shouldn't come back */
1383 exit(1); /* never know, though */
1387 * This function takes an IP addresses as its parameter. It returns the
1388 * the primary IP address that is on the host passed in, or 0 if not found.
1391 ubikGetPrimaryInterfaceAddr(afs_uint32 addr
)
1393 struct ubik_server
*ts
;
1397 for (ts
= ubik_servers
; ts
; ts
= ts
->next
)
1398 for (j
= 0; j
< UBIK_MAX_INTERFACE_ADDR
; j
++)
1399 if (ts
->addr
[j
] == addr
) {
1401 return ts
->addr
[0]; /* net byte order */
1404 return 0; /* if not in server database, return error */
1408 ubik_CheckAuth(struct rx_call
*acall
)
1410 if (checkSecurityProc
)
1411 return (*checkSecurityProc
) (securityRock
, acall
);
1412 else if (ubik_CheckRXSecurityProc
) {
1413 return (*ubik_CheckRXSecurityProc
) (ubik_CheckRXSecurityRock
, acall
);
1419 ubik_SetServerSecurityProcs(void (*buildproc
) (void *,
1420 struct rx_securityClass
***,
1422 int (*checkproc
) (void *, struct rx_call
*),
1425 buildSecClassesProc
= buildproc
;
1426 checkSecurityProc
= checkproc
;
1427 securityRock
= rock
;