Fixed bug in Aria when used with enterprise mariadb-backup

If the backup finished in the middle of a Aria bulk load insert,
which could happen with LOAD DATA INFILE, CREATE ... SELECT etc)
there was a chance that Aria recovery would fail on the backup.

Fixed by ensuring that bulk load operations for Aria are not allowed
under BACKUP LOCK.
I also changed so that the table TRN is updated just before truncate
which ensures that old redo's for the table are ignored.
I also enabled Aria redo for DDL's to be able to repeat REPAIR commands.
Without this change recovery would not work on repaired tables.

Notes:
- We take the backup lock protection at the end of bulk insert (as we
  don't want to keep the lock over a very long running insert).
  If mariadb-backup keeps the backup lock too long,  this may fail with
  a lock timeout. In this case the batch insert will fail and the table
  will be truncated (set to it's original state).

sql/handler.cc

@@ -1011,12 +1011,30 @@ void ha_end_backup()
                            PLUGIN_IS_DELETED|PLUGIN_IS_READY, 0);
 }
 
-void handler::log_not_redoable_operation(const char *operation)
+/*
+  Take a lock to block MDL_BACKUP_DDL (used by maria-backup) until
+  the DDL operation is taking place
+*/
+
+bool handler::log_not_redoable_operation(const char *operation)
 {
   DBUG_ENTER("log_not_redoable_operation");
   if (table->s->tmp_table == NO_TMP_TABLE)
   {
+    /*
+      Take a lock to ensure that mariadb-backup will notice the
+      new log entry (and re-copy the table if needed).
+    */
+    THD *thd= table->in_use;
+    MDL_request mdl_backup;
     backup_log_info ddl_log;
+
+    MDL_REQUEST_INIT(&mdl_backup, MDL_key::BACKUP, "", "", MDL_BACKUP_DDL,
+                     MDL_STATEMENT);
+    if (thd->mdl_context.acquire_lock(&mdl_backup,
+                                      thd->variables.lock_wait_timeout))
+      DBUG_RETURN(1);
+
     bzero(&ddl_log, sizeof(ddl_log));
     lex_string_set(&ddl_log.query, operation);
     /*
@@ -1032,7 +1050,7 @@ void handler::log_not_redoable_operation(const char *operation)
     ddl_log.org_table_id=     table->s->tabledef_version;
     backup_log_ddl(&ddl_log);
   }
-  DBUG_VOID_RETURN;
+  DBUG_RETURN(0);
 }
 
 /*

sql/handler.h

@@ -5198,7 +5198,7 @@ class handler :public Sql_alloc
     return (lower_case_table_names == 2 && !(ha_table_flags() & HA_FILE_BASED));
   }
 
-  void log_not_redoable_operation(const char *operation);
+  bool log_not_redoable_operation(const char *operation);
 protected:
   Handler_share *get_ha_share_ptr();
   void set_ha_share_ptr(Handler_share *arg_ha_share);

storage/maria/ha_maria.cc

@@ -872,7 +872,7 @@ extern "C" {
 
 int _ma_killed_ptr(HA_CHECK *param)
 {
-  if (likely(thd_killed((THD*)param->thd)) == 0)
+  if (!param->thd || likely(thd_killed((THD*)param->thd)) == 0)
     return 0;
   my_errno= HA_ERR_ABORTED_BY_USER;
   return 1;
@@ -901,9 +901,10 @@ int _ma_killed_ptr(HA_CHECK *param)
 void _ma_report_progress(HA_CHECK *param, ulonglong progress,
                          ulonglong max_progress)
 {
-  thd_progress_report((THD*)param->thd,
-                      progress + max_progress * param->stage,
-                      max_progress * param->max_stage);
+  if (param->thd)
+    thd_progress_report((THD*)param->thd,
+                        progress + max_progress * param->stage,
+                        max_progress * param->max_stage);
 }
 
 
@@ -2263,16 +2264,19 @@ void ha_maria::start_bulk_insert(ha_rows rows, uint flags)
       {
         bulk_insert_single_undo= BULK_INSERT_SINGLE_UNDO_AND_NO_REPAIR;
         write_log_record_for_bulk_insert(file);
-        _ma_tmp_disable_logging_for_table(file, TRUE);
         /*
           Pages currently in the page cache have type PAGECACHE_LSN_PAGE, we
           are not allowed to overwrite them with PAGECACHE_PLAIN_PAGE, so
           throw them away. It is not losing data, because we just wrote and
           forced an UNDO which will for sure empty the table if we crash. The
           upcoming unique-key insertions however need a proper index, so we
           cannot leave the corrupted on-disk index file, thus we truncate it.
+
+          The following call will log the truncate and update the lsn for the table
+          to ensure that all redo's before this will be ignored.
         */
         maria_delete_all_rows(file);
+        _ma_tmp_disable_logging_for_table(file, TRUE);
       }
     }
     else if (!file->bulk_insert &&
@@ -2303,23 +2307,58 @@ void ha_maria::start_bulk_insert(ha_rows rows, uint flags)
 
 int ha_maria::end_bulk_insert()
 {
-  int first_error, error;
-  my_bool abort= file->s->deleting;
+  int first_error, first_errno, error;
+  my_bool abort= file->s->deleting, empty_table= 0;
+  uint enable_index_mode= HA_KEY_SWITCH_NONUNIQ_SAVE;
   DBUG_ENTER("ha_maria::end_bulk_insert");
 
   if ((first_error= maria_end_bulk_insert(file, abort)))
+  {
+    first_errno= my_errno;
     abort= 1;
+  }
 
   if ((error= maria_extra(file, HA_EXTRA_NO_CACHE, 0)))
   {
-    first_error= first_error ? first_error : error;
+    if (!first_error)
+    {
+      first_error= error;
+      first_errno= my_errno;
+    }
     abort= 1;
   }
 
-  if (!abort && can_enable_indexes)
-    if ((error= enable_indexes(HA_KEY_SWITCH_NONUNIQ_SAVE)))
-      first_error= first_error ? first_error : error;
+  if (bulk_insert_single_undo != BULK_INSERT_NONE)
+  {
+    if (log_not_redoable_operation("BULK_INSERT"))
+    {
+      /* Got lock timeout. revert back to empty file and give error */
+      if (!first_error)
+      {
+        first_error= 1;
+        first_errno= my_errno;
+      }
+      enable_index_mode= HA_KEY_SWITCH_ALL;
+      empty_table= 1;
+      /*
+        Ignore all changed pages, required by _ma_renable_logging_for_table()
+      */
+      _ma_flush_table_files(file, MARIA_FLUSH_DATA|MARIA_FLUSH_INDEX,
+                            FLUSH_IGNORE_CHANGED, FLUSH_IGNORE_CHANGED);
+    }
+  }
 
+  if (!abort && can_enable_indexes)
+  {
+    if ((error= enable_indexes(enable_index_mode)))
+    {
+      if (!first_error)
+      {
+        first_error= 1;
+        first_errno= my_errno;
+      }
+    }
+  }
   if (bulk_insert_single_undo != BULK_INSERT_NONE)
   {
     /*
@@ -2328,12 +2367,23 @@ int ha_maria::end_bulk_insert()
     */
     if ((error= _ma_reenable_logging_for_table(file,
                                                bulk_insert_single_undo ==
-                                               BULK_INSERT_SINGLE_UNDO_AND_NO_REPAIR)))
-      first_error= first_error ? first_error : error;
-    bulk_insert_single_undo= BULK_INSERT_NONE;  // Safety
-    log_not_redoable_operation("BULK_INSERT");
+                                               BULK_INSERT_SINGLE_UNDO_AND_NO_REPAIR)) &&
+        !empty_table)
+    {
+      if (!first_error)
+      {
+        first_error= 1;
+        first_errno= my_errno;
+      }
+    }
+    bulk_insert_single_undo= BULK_INSERT_NONE;  // Safety if called again
   }
+  if (empty_table)
+    maria_delete_all_rows(file);
+
   can_enable_indexes= 0;
+  if (first_error)
+    my_errno= first_errno;
   DBUG_RETURN(first_error);
 }
 

storage/maria/ma_delete_all.c

@@ -64,6 +64,7 @@ int maria_delete_all_rows(MARIA_HA *info)
     */
     LEX_CUSTRING log_array[TRANSLOG_INTERNAL_PARTS + 1];
     uchar log_data[FILEID_STORE_SIZE];
+    my_bool error;
     log_array[TRANSLOG_INTERNAL_PARTS + 0].str=    log_data;
     log_array[TRANSLOG_INTERNAL_PARTS + 0].length= sizeof(log_data);
     if (unlikely(translog_write_record(&lsn, LOGREC_REDO_DELETE_ALL,
@@ -78,6 +79,32 @@ int maria_delete_all_rows(MARIA_HA *info)
     */
     if (_ma_mark_file_changed(share))
       goto err;
+
+    /*
+      Because LOGREC_REDO_DELETE_ALL does not operate on pages, it has the
+      following problem:
+      delete_all; inserts (redo_insert); all pages get flushed; checkpoint:
+      the dirty pages list will be empty. In recovery, delete_all is executed,
+      but redo_insert are skipped (dirty pages list is empty).
+      To avoid this, we need to set skip_redo_lsn now, and thus need to sync
+      files.
+      Also fixes the problem of:
+      bulk insert; insert; delete_all; crash:
+      "bulk insert" is skipped (no REDOs), so if "insert" would not be skipped
+      (if we didn't update skip_redo_lsn below) then "insert" would be tried
+      and fail, saying that it sees that the first page has to be created
+      though the inserted row has rownr>0.
+
+      We use lsn-1 below to ensure that the above redo will be executed
+    */
+     error= _ma_state_info_write(share,
+                                 MA_STATE_INFO_WRITE_DONT_MOVE_OFFSET |
+                                 MA_STATE_INFO_WRITE_LOCK) ||
+       _ma_update_state_lsns(share, lsn-1, info->trn->trid, FALSE, FALSE) ||
+       _ma_sync_table_files(info);
+     info->trn->rec_lsn= LSN_IMPOSSIBLE;
+     if (error)
+       goto err;
   }
   else
   {
@@ -113,28 +140,9 @@ int maria_delete_all_rows(MARIA_HA *info)
 
   if (log_record)
   {
-    /*
-      Because LOGREC_REDO_DELETE_ALL does not operate on pages, it has the
-      following problem:
-      delete_all; inserts (redo_insert); all pages get flushed; checkpoint:
-      the dirty pages list will be empty. In recovery, delete_all is executed,
-      but redo_insert are skipped (dirty pages list is empty).
-      To avoid this, we need to set skip_redo_lsn now, and thus need to sync
-      files.
-      Also fixes the problem of:
-      bulk insert; insert; delete_all; crash:
-      "bulk insert" is skipped (no REDOs), so if "insert" would not be skipped
-      (if we didn't update skip_redo_lsn below) then "insert" would be tried
-      and fail, saying that it sees that the first page has to be created
-      though the inserted row has rownr>0.
-    */
-    my_bool error= _ma_state_info_write(share,
-                                        MA_STATE_INFO_WRITE_DONT_MOVE_OFFSET |
-                                        MA_STATE_INFO_WRITE_LOCK) ||
-      _ma_update_state_lsns(share, lsn, info->trn->trid, FALSE, FALSE) ||
-      _ma_sync_table_files(info);
-    info->trn->rec_lsn= LSN_IMPOSSIBLE;
-    if (error)
+    /* Update lsn to signal that the above redo does not have to be executed anymore */
+    if ( _ma_update_state_lsns(share, lsn, info->trn->trid, FALSE, FALSE) ||
+         _ma_sync_table_files(info))
       goto err;
   }
 

storage/maria/ma_recovery.c

@@ -1175,16 +1175,6 @@ prototype_redo_exec_hook(REDO_REPAIR_TABLE)
   /* We try to get table first, so that we get the table in in the trace log */
   info= get_MARIA_HA_from_REDO_record(rec);
 
-  if (skip_DDLs)
-  {
-    /*
-      REPAIR is not exactly a DDL, but it manipulates files without logging
-      insertions into them.
-    */
-    tprint(tracef, "we skip DDLs\n");
-    DBUG_RETURN(0);
-  }
-
   if (!info)
   {
     /* no such table, don't need to warn */
@@ -1196,6 +1186,13 @@ prototype_redo_exec_hook(REDO_REPAIR_TABLE)
     tprint(tracef, "we skip repairing crashed table\n");
     DBUG_RETURN(0);
   }
+
+  if (rec->lsn <= info->s->state.is_of_horizon)
+  {
+    DBUG_PRINT("info", ("Table is up to date, skipping redo"));
+    DBUG_RETURN(0);
+  }
+
   /*
     Otherwise, the mapping is newer than the table, and our record is newer
     than the mapping, so we can repair.
@@ -1560,7 +1557,6 @@ prototype_redo_exec_hook(REDO_INSERT_ROW_HEAD)
   uchar *buff= NULL;
   MARIA_HA *info= get_MARIA_HA_from_REDO_record(rec);
   if (info == NULL || maria_is_crashed(info))
-
   {
     /*
       Table was skipped at open time (because later dropped/renamed, not