Avoiding a theoretically possible crash (pthread_mutex_lock(0)) which could...

Avoiding a theoretically possible crash (pthread_mutex_lock(0)) which could (at least in POSIX Threads books)
happen on SMP machines, when a thread is going to wait on a condition and it is KILLed at the
same time.
Cleaning code a bit by adding a test in enter_cond() that we have the mutex (was already the case in all places
where it's called except one which is fixed here).


sql/log.cc:
  safe_mutex_assert_owner() is now in THD::enter_cond()
sql/slave.cc:
  lock mutex before waiting on condition.
sql/sql_class.cc:
  THD::awake(): before locking the mutex, let's test it's not zero;
  in theory indeed, the killer thread may see current_cond non-zero and current_mutex zero
  (order of assignments is not guaranteed by POSIX).
  A comment noting that there is still a small chance a KILL does not work and needs
  being re-issued.
sql/sql_class.h:
  Assert in enter_cond() that we have the mutex.
  It is already the case in all places where we call enter_cond(), so better ensure it there.

sql/log.cc

@@ -1544,7 +1544,6 @@ bool MYSQL_LOG::write(THD *thd,const char *query, uint query_length,
 
 void MYSQL_LOG:: wait_for_update(THD* thd, bool master_or_slave)
 {
-  safe_mutex_assert_owner(&LOCK_log);
   const char* old_msg = thd->enter_cond(&update_cond, &LOCK_log,
                                         master_or_slave ?
                                         "Has read all relay log; waiting for \

sql/slave.cc

@@ -588,6 +588,7 @@ int start_slave_thread(pthread_handler h_func, pthread_mutex_t *start_lock,
     while (start_id == *slave_run_id)
     {
       DBUG_PRINT("sleep",("Waiting for slave thread to start"));
+      pthread_mutex_lock(cond_lock);
       const char* old_msg = thd->enter_cond(start_cond,cond_lock,
 					    "Waiting for slave thread to start");
       pthread_cond_wait(start_cond,cond_lock);

sql/sql_class.cc

@@ -299,8 +299,18 @@ void THD::awake(bool prepare_to_die)
       exits the cond in the time between read and broadcast, but that is
       ok since all we want to do is to make the victim thread get out
       of waiting on current_cond.
+      If we see a non-zero current_cond: it cannot be an old value (because
+      then exit_cond() should have run and it can't because we have mutex); so
+      it is the true value but maybe current_mutex is not yet non-zero (we're
+      in the middle of enter_cond() and there is a "memory order
+      inversion"). So we test the mutex too to not lock 0.
+      Note that there is a small chance we fail to kill. If victim has locked
+      current_mutex, and hasn't entered enter_cond(), then we don't know it's
+      going to wait on cond. Then victim goes into its cond "forever" (until
+      we issue a second KILL). True we have set its thd->killed but it may not
+      see it immediately and so may have time to reach the cond_wait().
     */
-    if (mysys_var->current_cond)
+    if (mysys_var->current_cond && mysys_var->current_mutex)
     {
       pthread_mutex_lock(mysys_var->current_mutex);
       pthread_cond_broadcast(mysys_var->current_cond);

sql/sql_class.h

@@ -539,6 +539,7 @@ class THD :public ilink
 			  const char* msg)
   {
     const char* old_msg = proc_info;
+    safe_mutex_assert_owner(mutex);
     mysys_var->current_mutex = mutex;
     mysys_var->current_cond = cond;
     proc_info = msg;