Fix for:

Bug #4810 "deadlock with KILL when the victim was in a wait state"
(I included mutex unlock into exit_cond() for future safety)
and BUG#4827 "KILL while START SLAVE may lead to replication slave crash"


sql/lock.cc:
  we did exit_cond() before unlock(LOCK_open), which led to deadlocks with THD::awake(). Fixing this.
sql/log.cc:
  mutex unlock is now included in exit_cond()
sql/repl_failsafe.cc:
  we did exit_cond() before unlock(LOCK_rpl_status), which led to deadlocks with THD::awake(). Fixing this.
sql/slave.cc:
  we did exit_cond() before unlock(cond_lock), which led to deadlocks with THD::awake(). Fixing this.
  Fixing also that if killed while waiting for slave thread to start, we don't release the mutex
  (that caused a double release of the mutex => crash).
sql/sql_class.h:
  comments about exit_cond()/enter_cond().
  Mutex unlock is now included in exit_cond() so that it's always done in the good order.
sql/sql_table.cc:
  unlock is now included in exit_cond().

sql/lock.cc

@@ -692,15 +692,14 @@ bool lock_global_read_lock(THD *thd)
     while (protect_against_global_read_lock && !thd->killed)
       pthread_cond_wait(&COND_refresh, &LOCK_open);
     waiting_for_read_lock--;
-    thd->exit_cond(old_message);
     if (thd->killed)
     {
-      (void) pthread_mutex_unlock(&LOCK_open);
+      thd->exit_cond(old_message);
       DBUG_RETURN(1);
     }
     thd->global_read_lock=1;
     global_read_lock++;
-    (void) pthread_mutex_unlock(&LOCK_open);
+    thd->exit_cond(old_message);
   }
   DBUG_RETURN(0);
 }
@@ -721,11 +720,12 @@ void unlock_global_read_lock(THD *thd)
 bool wait_if_global_read_lock(THD *thd, bool abort_on_refresh)
 {
   const char *old_message;
-  bool result=0;
+  bool result= 0, need_exit_cond;
   DBUG_ENTER("wait_if_global_read_lock");
 
+  LINT_INIT(old_message);
   (void) pthread_mutex_lock(&LOCK_open);
-  if (global_read_lock)
+  if (need_exit_cond= (bool)global_read_lock)
   {
     if (thd->global_read_lock)		// This thread had the read locks
     {
@@ -740,11 +740,13 @@ bool wait_if_global_read_lock(THD *thd, bool abort_on_refresh)
       (void) pthread_cond_wait(&COND_refresh,&LOCK_open);
     if (thd->killed)
       result=1;
-    thd->exit_cond(old_message);
   }
   if (!abort_on_refresh && !result)
     protect_against_global_read_lock++;
-  pthread_mutex_unlock(&LOCK_open);
+  if (unlikely(need_exit_cond)) // global read locks are rare
+    thd->exit_cond(old_message);
+  else
+    pthread_mutex_unlock(&LOCK_open);
   DBUG_RETURN(result);
 }
 

sql/log.cc

@@ -1533,12 +1533,8 @@ bool MYSQL_LOG::write(THD *thd,const char *query, uint query_length,
 
   NOTES
     One must have a lock on LOCK_log before calling this function.
-    This lock will be freed before return!
-
-    The reason for the above is that for enter_cond() / exit_cond() to
-    work the mutex must be got before enter_cond() but releases before
-    exit_cond().
-    If you don't do it this way, you will get a deadlock in THD::awake()
+    This lock will be freed before return! That's required by
+    THD::enter_cond() (see NOTES in sql_class.h).
 */
 
 
@@ -1551,7 +1547,6 @@ the I/O slave thread to update it" :
                                         "Has sent all binlog to slave; \
 waiting for binlog to be updated"); 
   pthread_cond_wait(&update_cond, &LOCK_log);
-  pthread_mutex_unlock(&LOCK_log);		// See NOTES
   thd->exit_cond(old_msg);
 }
 

sql/repl_failsafe.cc

@@ -584,6 +584,8 @@ pthread_handler_decl(handle_failsafe_rpl,arg)
   THD *thd = new THD;
   thd->thread_stack = (char*)&thd;
   MYSQL* recovery_captain = 0;
+  const char* msg;
+
   pthread_detach_this_thread();
   if (init_failsafe_rpl_thread(thd) || !(recovery_captain=mc_mysql_init(0)))
   {
@@ -591,11 +593,11 @@ pthread_handler_decl(handle_failsafe_rpl,arg)
     goto err;
   }
   pthread_mutex_lock(&LOCK_rpl_status);
+  msg= thd->enter_cond(&COND_rpl_status,
+                       &LOCK_rpl_status, "Waiting for request");
   while (!thd->killed && !abort_loop)
   {
     bool break_req_chain = 0;
-    const char* msg = thd->enter_cond(&COND_rpl_status,
-				      &LOCK_rpl_status, "Waiting for request");
     pthread_cond_wait(&COND_rpl_status, &LOCK_rpl_status);
     thd->proc_info="Processing request";
     while (!break_req_chain)
@@ -613,9 +615,8 @@ pthread_handler_decl(handle_failsafe_rpl,arg)
 	break;
       }
     }
-    thd->exit_cond(msg);
   }
-  pthread_mutex_unlock(&LOCK_rpl_status);
+  thd->exit_cond(msg);
 err:
   if (recovery_captain)
     mc_mysql_close(recovery_captain);

sql/slave.cc

@@ -582,7 +582,7 @@ int start_slave_thread(pthread_handler h_func, pthread_mutex_t *start_lock,
       pthread_mutex_unlock(start_lock);
     DBUG_RETURN(ER_SLAVE_THREAD);
   }
-  if (start_cond && cond_lock)
+  if (start_cond && cond_lock) // caller has cond_lock
   {
     THD* thd = current_thd;
     while (start_id == *slave_run_id)
@@ -592,11 +592,9 @@ int start_slave_thread(pthread_handler h_func, pthread_mutex_t *start_lock,
 					    "Waiting for slave thread to start");
       pthread_cond_wait(start_cond,cond_lock);
       thd->exit_cond(old_msg);
+      pthread_mutex_lock(cond_lock); // re-acquire it as exit_cond() released
       if (thd->killed)
-      {
-	pthread_mutex_unlock(cond_lock);
 	DBUG_RETURN(ER_SERVER_SHUTDOWN);
-      }
     }
   }
   if (start_lock)
@@ -1561,7 +1559,6 @@ thread to free enough relay log space");
          !rli->ignore_log_space_limit)
     pthread_cond_wait(&rli->log_space_cond, &rli->log_space_lock);
   thd->exit_cond(save_proc_info);
-  pthread_mutex_unlock(&rli->log_space_lock);
   DBUG_RETURN(slave_killed);
 }
 
@@ -1965,6 +1962,9 @@ int st_relay_log_info::wait_for_pos(THD* thd, String* log_name,
                       (long) timeout));
 
   pthread_mutex_lock(&data_lock);
+  const char *msg= thd->enter_cond(&data_cond, &data_lock,
+                                   "Waiting for the SQL slave thread to "
+                                   "advance position");
   /* 
      This function will abort when it notices that some CHANGE MASTER or
      RESET MASTER has changed the master info.
@@ -2063,9 +2063,6 @@ int st_relay_log_info::wait_for_pos(THD* thd, String* log_name,
     //wait for master update, with optional timeout.
     
     DBUG_PRINT("info",("Waiting for master update"));
-    const char* msg = thd->enter_cond(&data_cond, &data_lock,
-                                      "Waiting for the SQL slave thread to \
-advance position");
     /*
       We are going to pthread_cond_(timed)wait(); if the SQL thread stops it
       will wake us up.
@@ -2087,8 +2084,7 @@ advance position");
     }
     else
       pthread_cond_wait(&data_cond, &data_lock);
-    DBUG_PRINT("info",("Got signal of master update"));
-    thd->exit_cond(msg);
+    DBUG_PRINT("info",("Got signal of master update or timed out"));
     if (error == ETIMEDOUT || error == ETIME)
     {
       error= -1;
@@ -2100,7 +2096,7 @@ advance position");
   }
 
 err:
-  pthread_mutex_unlock(&data_lock);
+  thd->exit_cond(msg);
   DBUG_PRINT("exit",("killed: %d  abort: %d  slave_running: %d \
 improper_arguments: %d  timed_out: %d",
                      (int) thd->killed,

sql/sql_class.h

@@ -537,12 +537,9 @@ public:
   void awake(bool prepare_to_die);
   /*
     For enter_cond() / exit_cond() to work the mutex must be got before
-    enter_cond() but released before exit_cond() (in 4.1, assertions will soon
-    ensure this). Use must be:
-    lock mutex; enter_cond(); ...; unlock mutex; exit_cond().
-    If you don't do it this way, you will get a deadlock if another thread is
-    doing a THD::awake() on you.
-    
+    enter_cond() (in 4.1 an assertion will soon ensure this); this mutex is
+    then released by exit_cond(). Use must be:
+    lock mutex; enter_cond(); your code; exit_cond().
   */
   inline const char* enter_cond(pthread_cond_t *cond, pthread_mutex_t* mutex,
 			  const char* msg)
@@ -555,6 +552,13 @@ public:
   }
   inline void exit_cond(const char* old_msg)
   {
+    /*
+      Putting the mutex unlock in exit_cond() ensures that
+      mysys_var->current_mutex is always unlocked _before_ mysys_var->mutex is
+      locked (if that would not be the case, you'll get a deadlock if someone
+      does a THD::awake() on you).
+    */
+    pthread_mutex_unlock(mysys_var->current_mutex);
     pthread_mutex_lock(&mysys_var->mutex);
     mysys_var->current_mutex = 0;
     mysys_var->current_cond = 0;

sql/sql_table.cc

@@ -1301,7 +1301,6 @@ static int mysql_admin_table(THD* thd, TABLE_LIST* tables,
 	dropping_tables--;
       }
       thd->exit_cond(old_message);
-      pthread_mutex_unlock(&LOCK_open);
       if (thd->killed)
 	goto err;
       open_for_modify=0;