Fixed some bugs in the Maria storage engine

- Changed default recovery mode from OFF to NORMAL to get automatic repair of not properly closed tables.
- Fixed a rase condition when two threads calls external_lock and thr_lock() in different order. When this happend the transaction that called external lock first
  and thr_lock() last did not see see the rows from the other transaction, even if if it had to wait in thr_lock() for other to complete.
- Fixed that one can run maria_chk on an automatcally recovered tables without warnings about too small transaction id
- Don't give warning that crashed table could not be repaired if repair was disabled (and thus not run)
- Fixed a error result from flush_key_cache() which caused a DBUG_ASSERT() when one was using concurrent reads on non transactional tables that was updated.

client/mysqldump.c:
  Add "" around error message to make it more readable
client/mysqltest.cc:
  Free environment variables
mysql-test/r/mysqldump.result:
  Updated results
mysql-test/r/openssl_1.result:
  Updated results
mysql-test/suite/maria/r/maria-recover.result:
  Updated results
mysql-test/suite/maria/r/maria3.result:
  Updated results
mysql-test/suite/maria/t/maria3.test:
  Added more test of temporary tables
storage/maria/ha_maria.cc:
  Changed default recovery mode from OFF to NORMAL to get automatic repair of not properly closed tables.
  Start transaction in ma_block_get_status() instead of in ha_maria::external_lock().
  - This fixes a rase condition when two threads calls external lock and thr_lock() in different order. When this happend the transaction that called external lock first and thr_lock() last did not see see the rows from the other transaction, even if if it had to wait in thr_lock() for other to complete.
  Store latest transaction id in controll file if recovery was done.
  - This allows one to run maria_chk on an automatcally recovered tables without warnings about too small transaction id
storage/maria/ha_maria.h:
  Don't give warning that crashed table could not be repaired if repair was disabled (and thus not run)
storage/maria/ma_blockrec.h:
  Added new function "_ma_block_get_status_no_versioning()"
storage/maria/ma_init.c:
  Added hook to create trn in ma_block_get_status() if we are using MariaDB
storage/maria/ma_open.c:
  Ensure we call _ma_block_get_status_no_versioning() for transactional tables without versioning (like tables with fulltext)
storage/maria/ma_pagecache.c:
  Allow one to flush blocks that are pinned for read.
  This fixed a error result from flush_key_cache() which caused a DBUG_ASSERT() when one was using concurrent reads on non transactional tables that was updated.
storage/maria/ma_recovery.c:
  Set maria_recovery_changed_data to 1 if recover changed something.
  Set max_trid_in_control_file to max found trn if we found a bigger trn.
  The allows will ensure that the control file is up to date after recovery which allows one to run maria_chk on the tables without warnings about too big trn
storage/maria/ma_state.c:
  Call maria_create_trn_hook() in _ma_setup_live_state() instead of ha_maria::external_lock()
  This ensures that 'state' and trn are in sync and thus fixes the race condition mentioned for ha_maria.cc
storage/maria/ma_static.c:
  Added maria_create_trn_hook() and maria_recovery_changed_data
storage/maria/maria_def.h:
  Added MARIA_HANDLER->external_ptr, which is used to hold MariaDB thd.
  Added some new external variables
  Removed reference to non existing function: maria_concurrent_inserts()

1 files changed, 9 insertions, 8 deletions

diff --git a/storage/maria/ma_pagecache.c b/storage/maria/ma_pagecache.c
index dd676028f05..b95a6f53c06 100644
--- a/storage/maria/ma_pagecache.c
+++ b/storage/maria/ma_pagecache.c
@@ -4187,7 +4187,13 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
   {
     PAGECACHE_BLOCK_LINK *block= *cache;
 
-    if (block->pins)
+    /*
+      This code is only run for non transactional tables.
+      We may have other threads reading the block during flush,
+      as non transactional tables can have many readers while the
+      one writer is doing the flush.
+    */
+    if (block->wlocks)
     {
       KEYCACHE_DBUG_PRINT("flush_cached_blocks",
                           ("block: %u (0x%lx)  pinned",
@@ -4204,13 +4210,9 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
         *first_errno= HA_ERR_INTERNAL_ERROR;
       continue;
     }
-    /* if the block is not pinned then it is not write locked */
-    DBUG_ASSERT(block->wlocks == 0);
-    DBUG_ASSERT(block->pins == 0);
     if (make_lock_and_pin(pagecache, block,
-                          PAGECACHE_LOCK_WRITE, PAGECACHE_PIN, FALSE))
+                          PAGECACHE_LOCK_READ, PAGECACHE_PIN, FALSE))
       DBUG_ASSERT(0);
-    DBUG_ASSERT(block->pins == 1);
 
     KEYCACHE_DBUG_PRINT("flush_cached_blocks",
                         ("block: %u (0x%lx)  to be flushed",
@@ -4222,7 +4224,6 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
     DBUG_PRINT("info", ("block: %u (0x%lx)  pins: %u",
                         PCBLOCK_NUMBER(pagecache, block), (ulong)block,
                         block->pins));
-    DBUG_ASSERT(block->pins == 1);
     /**
        @todo IO If page is contiguous with next page to flush, group flushes
        in one single my_pwrite().
@@ -4241,7 +4242,7 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
     pagecache_pthread_mutex_lock(&pagecache->cache_lock);
 
     if (make_lock_and_pin(pagecache, block,
-                          PAGECACHE_LOCK_WRITE_UNLOCK,
+                          PAGECACHE_LOCK_READ_UNLOCK,
                           PAGECACHE_UNPIN, FALSE))
       DBUG_ASSERT(0);