bug#2831 - --extenral-locking does not fully work with --myisam-recover.

Changed the semantics of open_count so that it is decremented
at every unlock (if it was incremented due to data changes).
So it indicates a crash, if it is non-zero after a lock.
The table will then be repaired.


myisam/mi_close.c:
  bug#2831 - --extenral-locking does not fully work with --myisam-recover.
  To avoid flushing the open_count at every unlock,
  we need to do so at close at least.
myisam/mi_locking.c:
  bug#2831 - --extenral-locking does not fully work with --myisam-recover.
  open_count is now decremented at unlock (from a writelock) with
  mi_unlock_open_count(). After every new lock the state is read
  from the index file and the open_count checked. If not zero,
  another server must have crashed, so the table is marked as crashed.
  In certain situations the decremented open_count mut be flushed to
  the index file. I tried to keep these extra flushes as seldom as possible.
sql/ha_myisam.cc:
  bug#2831 - --extenral-locking does not fully work with --myisam-recover.
  Added code to repair the table, if it is marked crashed after
  successful locking and the --myisam-recover option is set.
sql/sql_table.cc:
  This does not really belong to the bugfix for #2831.
  But it was detected during fixing the external locking.

1 files changed, 21 insertions, 3 deletions

diff --git a/sql/ha_myisam.cc b/sql/ha_myisam.cc
index d5bbf9b5a92..2b7b8f436b1 100644
--- a/sql/ha_myisam.cc
+++ b/sql/ha_myisam.cc
@@ -1000,9 +1000,27 @@ int ha_myisam::delete_table(const char *name)
 
 int ha_myisam::external_lock(THD *thd, int lock_type)
 {
-  return mi_lock_database(file, !table->tmp_table ?
-			  lock_type : ((lock_type == F_UNLCK) ?
-				       F_UNLCK : F_EXTRA_LCK));
+  int rc;
+
+  while ((! (rc= mi_lock_database(file, !table->tmp_table ?
+                                  lock_type : ((lock_type == F_UNLCK) ?
+                                               F_UNLCK : F_EXTRA_LCK)))) &&
+         mi_is_crashed(file) && (myisam_recover_options != HA_RECOVER_NONE))
+  {
+    /*
+      check_and_repair() implicitly write locks the table, unless a
+      LOCK TABLES is in effect. It should be safer to always write lock here.
+      The implicit lock by check_and_repair() will then be a no-op.
+      check_and_repair() does not restore the original lock, but unlocks the
+      table. So we have to get the requested lock type again. And then to 
+      check, if the table has been crashed again meanwhile by another server.
+      If anything fails, we break.
+    */
+    if (((lock_type != F_WRLCK) && (rc= mi_lock_database(file, F_WRLCK))) ||
+        (rc= check_and_repair(thd)))
+      break;
+  }
+  return rc;
 }