rec-gen can be larger than cluster gen

Signed-off-by: mhannum <mhannum@bloomberg.net>

Author: markhannum

berkdb/txn/txn.c

@@ -1023,6 +1023,7 @@ void comdb2_cheapstack_sym(FILE *f, char *fmt, ...);
 
 extern int gbl_fullrecovery;
 int gbl_endianize_locklist = 1;
+int gbl_emit_gen_commits = 1;
 
 /*
  * __txn_commit --
@@ -1131,7 +1132,7 @@ __txn_commit_int(txnp, flags, ltranid, llid, last_commit_lsn, rlocks, inlks,
 	}
 
     /* don't let full recovery write a (higher) generation: it will force this newly-recovered node to be master on the next election */
-	elect_highest_committed_gen = !gbl_fullrecovery;
+	elect_highest_committed_gen = (!gbl_fullrecovery && gbl_emit_gen_commits);
 	db_rep = dbenv->rep_handle;
 	rep = db_rep->region;
 

db/db_tunables.c

@@ -149,6 +149,7 @@ extern int gbl_round_robin_stripes;
 extern int skip_clear_queue_extents;
 extern int gbl_rep_skip_recovery;
 extern int gbl_retrieve_gen_from_ckp;
+extern int gbl_emit_gen_commits;
 extern int gbl_recovery_ckp;
 extern int gbl_reproduce_ckp_bug;
 extern int gbl_sample_queries;

db/db_tunables.h

@@ -1057,6 +1057,8 @@ REGISTER_TUNABLE("rep_process_txn_trace",
                  NULL, NULL, NULL, NULL);
 REGISTER_TUNABLE("rep_skip_recovery", "Skip recovery if truncate won't unwind a transaction.  (Default: off)",
                  TUNABLE_BOOLEAN, &gbl_rep_skip_recovery, 0, NULL, NULL, NULL, NULL);
+REGISTER_TUNABLE("emit_gen_commits", "Emit commit-records which include cluster generation.  (Default: on)",
+                 TUNABLE_BOOLEAN, &gbl_emit_gen_commits, 0, NULL, NULL, NULL, NULL);
 /* 'retrieve_gen_from_ckp' / 'recovery_ckp' disabled under legacy_defaults until db moves */
 REGISTER_TUNABLE("retrieve_gen_from_ckp", "Retrieve generation from ckp records.  (Default: on)", TUNABLE_BOOLEAN,
                  &gbl_retrieve_gen_from_ckp, 0, NULL, NULL, NULL, NULL);

db/sqllogfill.c

@@ -326,9 +326,24 @@ static void print_record_info(const char *prefix, const char *lsn)
     }
 }
 
+extern int gbl_match_on_ckp;
+
 static inline int gen_okay(bdb_state_type *bdb_state, int64_t recgen, u_int32_t mygen)
 {
-    return recgen <= mygen;
+    if (recgen > mygen) {
+        static int64_t cnt = 0;
+        static int lastprint = 0;
+        cnt++;
+        if (gbl_debug_sql_logfill) {
+            int now = comdb2_time_epoch();
+            if (now - lastprint > 0) {
+                logmsg(LOGMSG_USER, "%s: recgen > mygen: recgen=%" PRId64 " mygen=%u cnt=%" PRId64 "\n",
+                       __func__, recgen, mygen, cnt);
+                lastprint = now;
+            }
+        }
+    }
+    return 1;
 }
 
 static int apply_record(bdb_state_type *bdb_state, const char *lsn, void *blob, int blob_len, int64_t genp,
@@ -698,7 +713,7 @@ static void *sql_apply_thread(void *arg)
             BDB_READLOCK(__func__);
             u_int32_t gen;
             bdb_state->dbenv->get_rep_gen(bdb_state->dbenv, &gen);
-            if (copy.gen == gen && (copy.recgen == 0 || gen >= copy.recgen)) {
+            if (copy.gen == gen && (copy.recgen == 0 || gen_okay(bdb_state, copy.recgen, gen))) {
                 rep_apply_caller = "sqllogfill-apply-thread";
                 bdb_state->dbenv->apply_log(bdb_state->dbenv, copy.file, copy.offset, copy.rectype, copy.blob,
                                             copy.blob_len);

tests/tunables.test/t00_all_tunables.expected

@@ -320,6 +320,7 @@
 (name='elect_on_mismatched_master', description='', type='BOOLEAN', value='ON', read_only='N')
 (name='elect_priority_bias', description='Bias this node's election priority by this amount.  (Default: 0)', type='INTEGER', value='0', read_only='N')
 (name='electtimebase', description='Master election timeout base value.', type='INTEGER', value='50', read_only='N')
+(name='emit_gen_commits', description='Emit commit-records which include cluster generation.  (Default: on)', type='BOOLEAN', value='ON', read_only='N')
 (name='emptystrnum', description='Empty strings don't convert to numbers', type='BOOLEAN', value='ON', read_only='N')
 (name='enable_2pc', description='Enable 2pc fdb transactions.  (Default: off)', type='BOOLEAN', value='OFF', read_only='Y')
 (name='enable_berkdb_retry_deadlock_bias', description='', type='BOOLEAN', value='OFF', read_only='Y')