{173778048} Count fix

bdb/bdb_cursor.h

@@ -199,6 +199,11 @@ int bdb_get_lsn_context_from_timestamp(bdb_state_type *bdb_state,
 int bdb_get_context_from_lsn(bdb_state_type *bdb_state, void *lsnp,
                              unsigned long long *ret_context, int *bdberr);
 
-int bdb_direct_count(bdb_cursor_ifn_t *, int ixnum, int64_t *count, int is_snapcur, uint32_t last_commit_lsn_file, uint32_t last_commit_lsn_offset, uint32_t last_checkpoint_lsn_file, uint32_t last_checkpoint_lsn_offset);
+int bdb_direct_count_int(bdb_state_type *state, int ixnum, int64_t *rcnt, int is_snapcur,
+                         uint32_t modsnap_start_lsn_file, uint32_t modsnap_start_lsn_offset,
+                         uint32_t last_checkpoint_lsn_file, uint32_t last_checkpoint_lsn_offset, int parallel_count);
+int bdb_direct_count(bdb_cursor_ifn_t *cur, int ixnum, int64_t *rcnt, int is_snapcur, uint32_t modsnap_start_lsn_file,
+                     uint32_t modsnap_start_lsn_offset, uint32_t last_checkpoint_lsn_file,
+                     uint32_t last_checkpoint_lsn_offset);
 
 #endif

bdb/count.c

@@ -208,6 +208,7 @@ int bdb_count_int(bdb_state_type *bdb_state, int *bdberr)
         return -1;
     }
 
+    // TODO: We are not using the smallest index here, why? We are using ix0
     rc = bdb_state->dbp_ix[0]->cursor(bdb_state->dbp_ix[0], 0, &dbcp, 0);
     if (rc != 0) {
         myfree(buffer);
@@ -300,7 +301,13 @@ int bdb_count(bdb_state_type *bdb_state, int *bdberr)
     int ret;
     BDB_READLOCK("bdb_count");
 
-    ret = bdb_count_int(bdb_state, bdberr);
+    // ret = bdb_count_int(bdb_state, bdberr);
+    int64_t count;
+    // bdb_count_int would use ix 0 if available, do the same thing here
+    ret = bdb_direct_count_int(bdb_state, bdb_state->numix > 0 ? 0 : -1, &count, 0, 0, 0, 0, 0, 0);
     BDB_RELLOCK();
-    return ret;
+    if (ret == 0)
+        return count;
+    *bdberr = (ret == BDBERR_DEADLOCK ? ret : BDBERR_MISC);
+    return -1;
 }

bdb/cursor.c

@@ -7454,19 +7454,92 @@ static void *db_count(void *varg)
     return NULL;
 }
 
-int gbl_parallel_count = 0;
-int bdb_direct_count(bdb_cursor_ifn_t *cur, int ixnum, int64_t *rcnt, int is_snapcur, uint32_t modsnap_start_lsn_file, uint32_t modsnap_start_lsn_offset, uint32_t last_checkpoint_lsn_file, uint32_t last_checkpoint_lsn_offset)
+static int find_recnum_first_last(DBC *dbcp, int *rnum, int last)
+{
+    DBT dbt_key, dbt_data;
+    int recnum;
+    int rc;
+    memset(&dbt_key, 0, sizeof(dbt_key));
+    memset(&dbt_data, 0, sizeof(dbt_data));
+    dbt_key.flags |= DB_DBT_MALLOC;
+    dbt_data.flags |= DB_DBT_MALLOC;
+    rc = dbcp->c_get(dbcp, &dbt_key, &dbt_data, last ? DB_LAST : DB_FIRST);
+    if (rc)
+        return rc;
+    if (dbt_key.data)
+        free(dbt_key.data);
+    if (dbt_data.data)
+        free(dbt_data.data);
+
+    memset(&dbt_data, 0, sizeof(dbt_data));
+    dbt_data.data = &recnum;
+    dbt_data.ulen = sizeof(int);
+    dbt_data.flags |= DB_DBT_USERMEM;
+    rc = dbcp->c_get(dbcp, &dbt_key, &dbt_data, DB_GET_RECNO);
+    if (rc)
+        return rc;
+    *rnum = recnum;
+    return 0;
+}
+
+static int find_count_recnums(bdb_state_type *state, int64_t *rcnt)
+{
+    int lowrecnum;
+    int highrecnum;
+    int rc;
+    DBC *dbcp;
+    if (!state->have_recnums)
+        return -1;
+    for (int i = 0; i < state->numix; i++) {
+        if (!state->ixrecnum[i])
+            continue;
+        rc = state->dbp_ix[i]->cursor(state->dbp_ix[i], 0, &dbcp, 0);
+        if (rc)
+            return rc;
+        if ((rc = find_recnum_first_last(dbcp, &lowrecnum, 0))) {
+            dbcp->c_close(dbcp);
+            return rc;
+        }
+        if ((rc = find_recnum_first_last(dbcp, &highrecnum, 1))) {
+            dbcp->c_close(dbcp);
+            return rc;
+        }
+        dbcp->c_close(dbcp);
+        *rcnt = highrecnum - lowrecnum + 1;
+        return 0;
+    }
+    return -1;
+}
+
+int bdb_direct_count_int(bdb_state_type *state, int ixnum, int64_t *rcnt, int is_snapcur,
+                         uint32_t modsnap_start_lsn_file, uint32_t modsnap_start_lsn_offset,
+                         uint32_t last_checkpoint_lsn_file, uint32_t last_checkpoint_lsn_offset, int parallel_count)
 {
     int64_t count = 0;
-    int parallel_count;
-    bdb_state_type *state = cur->impl->state;
+    int rc = 0;
     DB **db;
     int stripes;
     pthread_attr_t attr;
+
+    // first try recnums optimization
+    if (state->have_recnums) {
+        rc = find_count_recnums(state, &count);
+        if (rc == DB_LOCK_DEADLOCK) {
+            rc = BDBERR_DEADLOCK;
+        } else if (rc == DB_NOTFOUND) {
+            rc = 0;
+        } else if (rc != 0) {
+            rc = -1;
+        }
+
+        if (rc == 0)
+            *rcnt = count;
+        return rc;
+    }
+
     if (ixnum < 0) { // data
         db = state->dbp_data[0];
         stripes = state->attr->dtastripe;
-        parallel_count = gbl_parallel_count;
         Pthread_attr_init(&attr);
 #ifdef PTHREAD_STACK_MIN
         Pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN + 512 * 1024);
@@ -7493,7 +7566,6 @@ int bdb_direct_count(bdb_cursor_ifn_t *cur, int ixnum, int64_t *rcnt, int is_sna
             db_count(&args[i]);
         }
     }
-    int rc = 0;
     void *ret;
     for (int i = 0; i < stripes; ++i) {
         if (parallel_count) {
@@ -7516,3 +7588,13 @@ int bdb_direct_count(bdb_cursor_ifn_t *cur, int ixnum, int64_t *rcnt, int is_sna
     if (rc == 0) *rcnt = count;
     return rc;
 }
+
+int gbl_parallel_count = 0;
+int bdb_direct_count(bdb_cursor_ifn_t *cur, int ixnum, int64_t *rcnt, int is_snapcur, uint32_t modsnap_start_lsn_file,
+                     uint32_t modsnap_start_lsn_offset, uint32_t last_checkpoint_lsn_file,
+                     uint32_t last_checkpoint_lsn_offset)
+{
+    return bdb_direct_count_int(cur->impl->state, ixnum, rcnt, is_snapcur, modsnap_start_lsn_file,
+                                modsnap_start_lsn_offset, last_checkpoint_lsn_file, last_checkpoint_lsn_offset,
+                                gbl_parallel_count);
+}

tests/recnum.test/runit

@@ -36,4 +36,16 @@ if [[ $leaked_pages -lt 0 ]]; then
     cat out.res
     exit 1
 fi
+count=$(cdb2sql --tabs ${CDB2_OPTIONS} $dbnm default "select count(*) from t")
+if [[ $count != 10000 ]]; then
+    echo "Expected count 10000, got count $count"
+    exit 1
+fi
+
+cdb2sql --tabs ${CDB2_OPTIONS} $dbnm default "truncate t"
+count=$(cdb2sql --tabs ${CDB2_OPTIONS} $dbnm default "select count(*) from t")
+if [[ $count != 0 ]]; then
+    echo "Expected count 0, got count $count"
+    exit 1
+fi
 exit 0