ubik: avoid DISK_Begin on sites that didn't vote for sync

[openafs.git] / src / ubik / ubik.c

diff --git a/src/ubik/ubik.c b/src/ubik/ubik.c
index b42bf4c..bc04f99 100644 (file)
--- a/src/ubik/ubik.c
+++ b/src/ubik/ubik.c
@@ -12,11 +12,19 @@
 
 #include <roken.h>
 
+
 #include <afs/opr.h>
+#ifdef AFS_PTHREAD_ENV
+# include <opr/lock.h>
+#else
+# include <opr/lockstub.h>
+#endif
+
 #include <lock.h>
 #include <rx/rx.h>
 #include <afs/cellconfig.h>
 
+
 #define UBIK_INTERNALS
 #include "ubik.h"
 #include "ubik_int.h"
@@ -70,7 +78,6 @@ afs_int32 ubik_quorum = 0;
 struct ubik_dbase *ubik_dbase = 0;
 struct ubik_stats ubik_stats;
 afs_uint32 ubik_host[UBIK_MAX_INTERFACE_ADDR];
-afs_int32 ubik_epochTime = 0;
 afs_int32 urecovery_state = 0;
 int (*ubik_SyncWriterCacheProc) (void);
 struct ubik_server *ubik_servers;
@@ -101,6 +108,7 @@ static void *securityRock = NULL;
 struct version_data version_globals;
 
 #define        CStampVersion       1   /* meaning set ts->version */
+#define        CCheckSyncAdvertised        2   /* check if the remote knows we are the sync-site */
 
 static_inline struct rx_connection *
 Quorum_StartIO(struct ubik_trans *atrans, struct ubik_server *as)
@@ -174,7 +182,10 @@ ContactQuorum_iterate(struct ubik_trans *atrans, int aflags, struct ubik_server
     if (!(*ts))
        return 1;
     UBIK_BEACON_LOCK;
-    if (!(*ts)->up || !(*ts)->currentDB) {
+    if (!(*ts)->up || !(*ts)->currentDB ||
+       /* do not call DISK_Begin until we know that lastYesState is set on the
+        * remote in question; otherwise, DISK_Begin will fail. */
+       ((aflags & CCheckSyncAdvertised) && !((*ts)->beaconSinceDown && (*ts)->lastVote))) {
        UBIK_BEACON_UNLOCK;
        (*ts)->currentDB = 0;   /* db is no longer current; we just missed an update */
        return 0;               /* not up-to-date, don't bother.  NULL conn will tell caller not to use */
@@ -194,7 +205,7 @@ ContactQuorum_rcode(int okcalls, afs_int32 rcode)
     if (okcalls + 1 >= ubik_quorum)
        return 0;
     else
-       return rcode;
+       return (rcode != 0) ? rcode : UNOQUORUM;
 }
 
 /*!
@@ -410,11 +421,11 @@ ubik_ServerInitCommon(afs_uint32 myHost, short myPort,
     memset(&tdb->version, 0, sizeof(struct ubik_version));
     memset(&tdb->cachedVersion, 0, sizeof(struct ubik_version));
 #ifdef AFS_PTHREAD_ENV
-    MUTEX_INIT(&tdb->versionLock, "version lock", MUTEX_DEFAULT, 0);
-    MUTEX_INIT(&beacon_globals.beacon_lock, "beacon lock", MUTEX_DEFAULT, 0);
-    MUTEX_INIT(&vote_globals.vote_lock, "vote lock", MUTEX_DEFAULT, 0);
-    MUTEX_INIT(&addr_globals.addr_lock, "address lock", MUTEX_DEFAULT, 0);
-    MUTEX_INIT(&version_globals.version_lock, "version lock", MUTEX_DEFAULT, 0);
+    opr_mutex_init(&tdb->versionLock);
+    opr_mutex_init(&beacon_globals.beacon_lock);
+    opr_mutex_init(&vote_globals.vote_lock);
+    opr_mutex_init(&addr_globals.addr_lock);
+    opr_mutex_init(&version_globals.version_lock);
 #else
     Lock_Init(&tdb->versionLock);
 #endif
@@ -435,8 +446,8 @@ ubik_ServerInitCommon(afs_uint32 myHost, short myPort,
     ubik_dbase = tdb;          /* for now, only one db per server; can fix later when we have names for the other dbases */
 
 #ifdef AFS_PTHREAD_ENV
-    CV_INIT(&tdb->version_cond, "version", CV_DEFAULT, 0);
-    CV_INIT(&tdb->flags_cond, "flags", CV_DEFAULT, 0);
+    opr_cv_init(&tdb->version_cond);
+    opr_cv_init(&tdb->flags_cond);
 #endif /* AFS_PTHREAD_ENV */
 
     /* initialize RX */
@@ -624,7 +635,7 @@ BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
        /* if we're writing already, wait */
        while (dbase->flags & DBWRITING) {
 #ifdef AFS_PTHREAD_ENV
-           CV_WAIT(&dbase->flags_cond, &dbase->versionLock);
+           opr_cv_wait(&dbase->flags_cond, &dbase->versionLock);
 #else
            DBRELE(dbase);
            LWP_WaitProcess(&dbase->flags);
@@ -636,6 +647,11 @@ BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
            DBRELE(dbase);
            return UNOTSYNC;
        }
+       if (!ubeacon_SyncSiteAdvertised()) {
+           /* i am the sync-site but the remotes are not aware yet */
+           DBRELE(dbase);
+           return UNOQUORUM;
+       }
     }
 
     /* create the transaction */
@@ -653,29 +669,31 @@ BeginTrans(struct ubik_dbase *dbase, afs_int32 transMode,
        }
     }
     /* label trans and dbase with new tid */
-    tt->tid.epoch = ubik_epochTime;
+    tt->tid.epoch = version_globals.ubik_epochTime;
     /* bump by two, since tidCounter+1 means trans id'd by tidCounter has finished */
     tt->tid.counter = (dbase->tidCounter += 2);
 
     if (transMode == UBIK_WRITETRANS) {
        /* for a write trans, we have to keep track of the write tid counter too */
        dbase->writeTidCounter = tt->tid.counter;
+    }
+
+    UBIK_VERSION_UNLOCK;
 
+    if (transMode == UBIK_WRITETRANS) {
        /* next try to start transaction on appropriate number of machines */
-       code = ContactQuorum_NoArguments(DISK_Begin, tt, 0);
+       code = ContactQuorum_NoArguments(DISK_Begin, tt, CCheckSyncAdvertised);
        if (code) {
            /* we must abort the operation */
            udisk_abort(tt);
            ContactQuorum_NoArguments(DISK_Abort, tt, 0); /* force aborts to the others */
            udisk_end(tt);
-           UBIK_VERSION_UNLOCK;
            DBRELE(dbase);
            return code;
        }
     }
 
     *transPtr = tt;
-    UBIK_VERSION_UNLOCK;
     DBRELE(dbase);
     return 0;
 }
@@ -1255,7 +1273,7 @@ ubik_WaitVersion(struct ubik_dbase *adatabase,
            return 0;
        }
 #ifdef AFS_PTHREAD_ENV
-       CV_WAIT(&adatabase->version_cond, &adatabase->versionLock);
+       opr_cv_wait(&adatabase->version_cond, &adatabase->versionLock);
 #else
        DBRELE(adatabase);
        LWP_WaitProcess(&adatabase->version);   /* same vers, just wait */
@@ -1374,7 +1392,7 @@ panic(char *format, ...)
     va_list ap;
 
     va_start(ap, format);
-    ubik_print("Ubik PANIC: ");
+    ubik_print("Ubik PANIC:\n");
     ubik_vprint(format, ap);
     va_end(ap);