Merge ibabaev@bk-internal.mysql.com:/home/bk/mysql-4.1

into rurik.mysql.com:/home/igor/mysql-4.1

sql/log.cc

@@ -966,14 +966,6 @@ void MYSQL_LOG::new_file(bool need_lock)
 	THD* thd = current_thd;
 	Rotate_log_event r(thd,new_name+dirname_length(new_name));
 	r.set_log_pos(this);
-
-	/*
-	  Because this log rotation could have been initiated by a master of
-	  the slave running with log-bin, we set the flag on rotate
-	  event to prevent infinite log rotation loop
-	*/
-	if (thd->slave_thread)
-	  r.flags|= LOG_EVENT_FORCED_ROTATE_F;
 	r.write(&log_file);
 	bytes_written += r.get_event_len();
       }

sql/log_event.cc

@@ -1057,7 +1057,8 @@ int Start_log_event::write_data(IO_CACHE* file)
   The master started
 
   IMPLEMENTATION
-    - To handle the case where the master died without a stop event,
+    - To handle the case where the master died without having time to write DROP
+      TEMPORARY TABLE, DO RELEASE_LOCK (prepared statements' deletion is TODO),
       we clean up all temporary tables + locks that we got.
       However, we don't clean temporary tables if the master was 3.23
       (this is because a 3.23 master writes a Start_log_event at every
@@ -1065,11 +1066,20 @@ int Start_log_event::write_data(IO_CACHE* file)
       on the slave when FLUSH LOGS is issued on the master). 
 
   TODO
-    - Remove all active user locks
+    - Remove all active user locks.
+      Guilhem 2003-06: this is true but not urgent: the worst it can cause is
+      the use of a bit of memory for a user lock which will not be used
+      anymore. If the user lock is later used, the old one will be released. In
+      other words, no deadlock problem.
     - If we have an active transaction at this point, the master died
       in the middle while writing the transaction to the binary log.
       In this case we should stop the slave.
-
+      Guilhem 2003-06: I don't think we should. As the binlog is written before
+      the table changes are committed, rollback has occured on the master; we
+      should rather rollback on the slave and go on. If we don't rollback, and
+      the next query is not BEGIN, then it will be considered as part of the
+      unfinished transaction, and so will be rolled back at next BEGIN, which is
+      a bug.
 */
 
 #if defined(HAVE_REPLICATION) && !defined(MYSQL_CLIENT)
@@ -1079,6 +1089,11 @@ int Start_log_event::exec_event(struct st_relay_log_info* rli)
 
   if (!rli->mi->old_format)
   {
+    /*
+      If the master died before writing the COMMIT to the binlog, rollback;
+      otherwise it does not hurt to rollback.
+    */
+    ha_rollback(thd);
     /* 
        If 4.0 master, all temporary tables have been deleted on the master; 
        if 3.23 master, this is far from sure.
@@ -1703,8 +1718,6 @@ void Rotate_log_event::pack_info(Protocol *protocol)
   b_pos+= ident_len;
   b_pos= strmov(b_pos, ";pos=");
   b_pos=longlong10_to_str(pos, b_pos, 10);
-  if (flags & LOG_EVENT_FORCED_ROTATE_F)
-    b_pos= strmov(b_pos ,"; forced by master");
   protocol->store(buf, b_pos-buf, &my_charset_bin);
   my_free(buf, MYF(MY_ALLOW_ZERO_PTR));
 }
@@ -1728,8 +1741,6 @@ void Rotate_log_event::print(FILE* file, bool short_form, char* last_db)
     my_fwrite(file, (byte*) new_log_ident, (uint)ident_len, 
 	      MYF(MY_NABP | MY_WME));
   fprintf(file, "  pos: %s", llstr(pos, buf));
-  if (flags & LOG_EVENT_FORCED_ROTATE_F)
-    fprintf(file,"  forced by master");
   fputc('\n', file);
   fflush(file);
 }
@@ -2399,28 +2410,20 @@ void Stop_log_event::print(FILE* file, bool short_form, char* last_db)
 /*
   Stop_log_event::exec_event()
 
-  The master stopped. Clean up all temporary tables + locks that the
-  master may have set.
-
-  TODO
-    - Remove all active user locks
+  The master stopped. 
+  We used to clean up all temporary tables but this is useless as, as the master
+  has shut down properly, it has written all DROP TEMPORARY TABLE and DO
+  RELEASE_LOCK (prepared statements' deletion is TODO).
+  We used to clean up slave_load_tmpdir, but this is useless as it has been
+  cleared at the end of LOAD DATA INFILE.
+  So we have nothing to do here.
+  The place were we must do this cleaning is in Start_log_event::exec_event(),
+  not here. Because if we come here, the master was sane.
 */
 
 #ifndef MYSQL_CLIENT
 int Stop_log_event::exec_event(struct st_relay_log_info* rli)
 {
-  /*
-    do not clean up immediately after rotate event;
-    QQ: this should be a useless test: the only case when it is false is when
-    shutdown occurred just after FLUSH LOGS. It has nothing to do with Rotate?
-    By the way, immediately after a Rotate the I/O thread does not write
-    the Stop to the relay log, so we won't come here in that case.
-  */
-  if (rli->group_master_log_pos > BIN_LOG_HEADER_SIZE) 
-  {
-    close_temporary_tables(thd);
-    cleanup_load_tmpdir();
-  }
   /*
     We do not want to update master_log pos because we get a rotate event
     before stop, so by now group_master_log_name is set to the next log.
@@ -2965,10 +2968,10 @@ int Execute_load_log_event::exec_event(struct st_relay_log_info* rli)
     goto err;
   }
   /*
-    We want to disable binary logging in slave thread because we need the file
-    events to appear in the same order as they do on the master relative to
-    other events, so that we can preserve ascending order of log sequence
-    numbers - needed to handle failover .
+    We are going to create a Load_log_event to finally load into the table.
+    This event should not go into the binlog: in the binlog we only want the
+    Create_file, Append_blocks and Execute_load. We disable binary logging and
+    restore the thread's options just after finishing the load.
   */
   save_options = thd->options;
   thd->options &= ~ (ulong) (OPTION_BIN_LOG);

sql/log_event.h

@@ -34,15 +34,21 @@
 #define LOG_READ_TOO_LARGE -7
 
 #define LOG_EVENT_OFFSET 4
+
 #define BINLOG_VERSION    3
 
 /*
  We could have used SERVER_VERSION_LENGTH, but this introduces an
  obscure dependency - if somebody decided to change SERVER_VERSION_LENGTH
- this would have broke the replication protocol
+ this would have broken the replication protocol
 */
 #define ST_SERVER_VER_LEN 50
 
+/*
+  These are flags and structs to handle all the LOAD DATA INFILE options (LINES
+  TERMINATED etc).
+*/
+
 #define DUMPFILE_FLAG		0x1
 #define OPT_ENCLOSED_FLAG	0x2
 #define REPLACE_FLAG		0x4
@@ -121,11 +127,17 @@ struct sql_ex_info
 
   See the #defines below for the format specifics.
 
+  The events which really update data are Query_log_event and
+  Load_log_event/Create_file_log_event/Execute_load_log_event (these 3 act
+  together to replicate LOAD DATA INFILE, with the help of
+  Append_block_log_event which prepares temporary files to load into the table).
+
  ****************************************************************************/
 
+#define LOG_EVENT_HEADER_LEN 19     /* the fixed header length */
+#define OLD_HEADER_LEN       13     /* the fixed header length in 3.23 */
+
 /* event-specific post-header sizes */
-#define LOG_EVENT_HEADER_LEN 19
-#define OLD_HEADER_LEN       13
 #define QUERY_HEADER_LEN     (4 + 4 + 1 + 2)
 #define LOAD_HEADER_LEN      (4 + 4 + 4 + 1 +1 + 4)
 #define START_HEADER_LEN     (2 + ST_SERVER_VER_LEN + 4)
@@ -135,7 +147,10 @@ struct sql_ex_info
 #define EXEC_LOAD_HEADER_LEN   4
 #define DELETE_FILE_HEADER_LEN 4
 
-/* event header offsets */
+/* 
+   Event header offsets; 
+   these point to places inside the fixed header.
+*/
 
 #define EVENT_TYPE_OFFSET    4
 #define SERVER_ID_OFFSET     5
@@ -149,7 +164,7 @@ struct sql_ex_info
 #define ST_SERVER_VER_OFFSET  2
 #define ST_CREATED_OFFSET     (ST_SERVER_VER_OFFSET + ST_SERVER_VER_LEN)
 
-/* slave event post-header */
+/* slave event post-header (this event is never written) */
 
 #define SL_MASTER_PORT_OFFSET   8
 #define SL_MASTER_POS_OFFSET    0
@@ -197,14 +212,20 @@ struct sql_ex_info
 #define R_POS_OFFSET       0
 #define R_IDENT_OFFSET     8
 
+/* CF to DF handle LOAD DATA INFILE */
+
+/* CF = "Create File" */
 #define CF_FILE_ID_OFFSET  0
 #define CF_DATA_OFFSET     CREATE_FILE_HEADER_LEN
 
+/* AB = "Append Block" */
 #define AB_FILE_ID_OFFSET  0
 #define AB_DATA_OFFSET     APPEND_BLOCK_HEADER_LEN
 
+/* EL = "Execute Load" */
 #define EL_FILE_ID_OFFSET  0
 
+/* DF = "Delete File" */
 #define DF_FILE_ID_OFFSET  0
 
 #define QUERY_EVENT_OVERHEAD	(LOG_EVENT_HEADER_LEN+QUERY_HEADER_LEN)
@@ -217,13 +238,31 @@ struct sql_ex_info
 #define EXEC_LOAD_EVENT_OVERHEAD (LOG_EVENT_HEADER_LEN+EXEC_LOAD_HEADER_LEN)
 #define APPEND_BLOCK_EVENT_OVERHEAD (LOG_EVENT_HEADER_LEN+APPEND_BLOCK_HEADER_LEN)
 
-
+/* 4 bytes which all binlogs should begin with */
 #define BINLOG_MAGIC        "\xfe\x62\x69\x6e"
 
+/*
+  The 2 flags below were useless :
+  - the first one was never set
+  - the second one was set in all Rotate events on the master, but not used for
+  anything useful.
+  So they are now removed and their place may later be reused for other
+  flags. Then one must remember that Rotate events in 4.x have
+  LOG_EVENT_FORCED_ROTATE_F set, so one should not rely on the value of the
+  replacing flag when reading a Rotate event. 
+  I keep the defines here just to remember what they were.
+*/
+#ifdef TO_BE_REMOVED
 #define LOG_EVENT_TIME_F            0x1
 #define LOG_EVENT_FORCED_ROTATE_F   0x2 
-#define LOG_EVENT_THREAD_SPECIFIC_F 0x4 /* query depends on thread  
-                                           (for example: TEMPORARY TABLE) */
+#endif
+/* 
+   If the query depends on the thread (for example: TEMPORARY TABLE).
+   Currently this is used by mysqlbinlog to know it must print
+   SET @@PSEUDO_THREAD_ID=xx; before the query (it would not hurt to print it
+   for every query but this would be slow).
+*/
+#define LOG_EVENT_THREAD_SPECIFIC_F 0x4 
 
 enum Log_event_type
 {
@@ -258,30 +297,81 @@ struct st_relay_log_info;
 class Log_event
 {
 public:
+  /* 
+     The offset in the log where this event originally appeared (it is preserved
+     in relay logs, making SHOW SLAVE STATUS able to print coordinates of the
+     event in the master's binlog). Note: when a transaction is written by the
+     master to its binlog (wrapped in BEGIN/COMMIT) the log_pos of all the
+     queries it contains is the one of the BEGIN (this way, when one does SHOW
+     SLAVE STATUS it sees the offset of the BEGIN, which is logical as rollback
+     may occur), except the COMMIT query which has its real offset.
+  */
   my_off_t log_pos;
+  /* 
+     A temp buffer for read_log_event; it is later analysed according to the
+     event's type, and its content is distributed in the event-specific fields.
+  */
   char *temp_buf; 
+  /*
+    Timestamp on the master(for debugging and replication of NOW()/TIMESTAMP). 
+    It is important for queries and LOAD DATA INFILE. This is set at the event's
+    creation time, except for Query and Load (et al.) events where this is set
+    at the query's execution time, which guarantees good replication (otherwise,
+    we could have a query and its event with different timestamps).
+  */
   time_t when;
+  /* The number of seconds the query took to run on the master. */
   ulong exec_time;
+  /* 
+     The master's server id (is preserved in the relay log; used to prevent from
+     infinite loops in circular replication). 
+  */
   uint32 server_id;
   uint cached_event_len;
+
+  /*
+    Some 16 flags. Only one is really used now; look above for
+    LOG_EVENT_TIME_F, LOG_EVENT_FORCED_ROTATE_F, LOG_EVENT_THREAD_SPECIFIC_F
+    for notes.
+  */
   uint16 flags;
+
   bool cache_stmt;
 #ifndef MYSQL_CLIENT
   THD* thd;
 
   Log_event(THD* thd_arg, uint16 flags_arg, bool cache_stmt);
   Log_event();
+  /*
+    read_log_event() functions read an event from a binlog or relay log; used by
+    SHOW BINLOG EVENTS, the binlog_dump thread on the master (reads master's
+    binlog), the slave IO thread (reads the event sent by binlog_dump), the
+    slave SQL thread (reads the event from the relay log).
+  */
   // if mutex is 0, the read will proceed without mutex
   static Log_event* read_log_event(IO_CACHE* file,
 				   pthread_mutex_t* log_lock,
 				   bool old_format);
   static int read_log_event(IO_CACHE* file, String* packet,
 			    pthread_mutex_t* log_lock);
+  /* set_log_pos() is used to fill log_pos with tell(log). */
   void set_log_pos(MYSQL_LOG* log);
+  /*
+    init_show_field_list() prepares the column names and types for the output of
+    SHOW BINLOG EVENTS; it is used only by SHOW BINLOG EVENTS.
+  */
   static void init_show_field_list(List<Item>* field_list);
 #ifdef HAVE_REPLICATION
   int net_send(Protocol *protocol, const char* log_name, my_off_t pos);
+  /*
+    pack_info() is used by SHOW BINLOG EVENTS; as print() it prepares and sends
+    a string to display to the user, so it resembles print().
+  */
   virtual void pack_info(Protocol *protocol);
+  /*
+    The SQL slave thread calls exec_event() to execute the event; this is where
+    the slave's data is modified.
+  */
   virtual int exec_event(struct st_relay_log_info* rli);
 #endif /* HAVE_REPLICATION */
   virtual const char* get_db()
@@ -291,6 +381,7 @@ public:
 #else
  // avoid having to link mysqlbinlog against libpthread
   static Log_event* read_log_event(IO_CACHE* file, bool old_format);
+  /* print*() functions are used by mysqlbinlog */
   virtual void print(FILE* file, bool short_form = 0, char* last_db = 0) = 0;
   void print_timestamp(FILE* file, time_t *ts = 0);
   void print_header(FILE* file);
@@ -336,6 +427,7 @@ public:
   }
   static Log_event* read_log_event(const char* buf, int event_len,
 				   const char **error, bool old_format);
+  /* returns the human readable name of the event's type */
   const char* get_type_str();
 };
 
@@ -403,6 +495,8 @@ public:
 /*****************************************************************************
 
   Slave Log Event class
+  Note that this class is currently not used at all; no code writes a
+  Slave_log_event (though some code in repl_failsafe.cc reads Slave_log_event).
 
  ****************************************************************************/
 class Slave_log_event: public Log_event
@@ -593,7 +687,7 @@ public:
 
   Rand Log Event class
 
-  Logs random seed used by the next RAND()
+  Logs random seed used by the next RAND(), and by PASSWORD() in 4.1.
 
  ****************************************************************************/
 class Rand_log_event: public Log_event
@@ -626,6 +720,9 @@ class Rand_log_event: public Log_event
 
   User var Log Event class
 
+  Every time a query uses the value of a user variable, a User_var_log_event is
+  written before the Query_log_event, to set the user variable.
+
  ****************************************************************************/
 class User_var_log_event: public Log_event
 {

sql/slave.cc

@@ -1426,7 +1426,6 @@ int init_master_info(MASTER_INFO* mi, const char* master_info_fname,
     DBUG_RETURN(0);
   mi->mysql=0;
   mi->file_id=1;
-  mi->ignore_stop_event=0;
   fn_format(fname, master_info_fname, mysql_data_home, "", 4+32);
 
   /*
@@ -2746,6 +2745,8 @@ static int process_io_rotate(MASTER_INFO *mi, Rotate_log_event *rev)
 /*
   queue_old_event()
 
+  Writes a 3.23 event to the relay log.
+
   TODO: 
     Test this code before release - it has to be tested on a separate
     setup with 3.23 master 
@@ -2790,8 +2791,7 @@ static int queue_old_event(MASTER_INFO *mi, const char *buf,
   ev->log_pos = mi->master_log_pos;
   switch (ev->get_type_code()) {
   case STOP_EVENT:
-    ignore_event= mi->ignore_stop_event;
-    mi->ignore_stop_event=0;
+    ignore_event= 1;
     inc_pos= event_len;
     break;
   case ROTATE_EVENT:
@@ -2801,7 +2801,6 @@ static int queue_old_event(MASTER_INFO *mi, const char *buf,
       pthread_mutex_unlock(&mi->data_lock);
       DBUG_RETURN(1);
     }
-    mi->ignore_stop_event=1;
     inc_pos= 0;
     break;
   case CREATE_FILE_EVENT:
@@ -2817,7 +2816,6 @@ static int queue_old_event(MASTER_INFO *mi, const char *buf,
     DBUG_RETURN(error);
   }
   default:
-    mi->ignore_stop_event=0;
     inc_pos= event_len;
     break;
   }
@@ -2842,15 +2840,12 @@ static int queue_old_event(MASTER_INFO *mi, const char *buf,
 /*
   queue_event()
 
-  TODO: verify the issue with stop events, see if we need them at all
-  in the relay log
 */
 
 int queue_event(MASTER_INFO* mi,const char* buf, ulong event_len)
 {
   int error= 0;
   ulong inc_pos;
-  bool ignore_event= 0;
   RELAY_LOG_INFO *rli= &mi->rli;
   DBUG_ENTER("queue_event");
 
@@ -2861,39 +2856,77 @@ int queue_event(MASTER_INFO* mi,const char* buf, ulong event_len)
 
   /*
     TODO: figure out if other events in addition to Rotate
-    require special processing
+    require special processing.
+    Guilhem 2003-06 : I don't think so.
   */
   switch (buf[EVENT_TYPE_OFFSET]) {
   case STOP_EVENT:
-    ignore_event= mi->ignore_stop_event;
-    mi->ignore_stop_event= 0;
-    inc_pos= event_len;
-    break;
+    /*
+      We needn't write this event to the relay log. Indeed, it just indicates a
+      master server shutdown. The only thing this does is cleaning. But cleaning
+      is already done on a per-master-thread basis (as the master server is
+      shutting down cleanly, it has written all DROP TEMPORARY TABLE and DO
+      RELEASE_LOCK; prepared statements' deletion are TODO).
+      
+      We don't even increment mi->master_log_pos, because we may be just after a
+      Rotate event. Btw, in a few milliseconds we are going to have a Start
+      event from the next binlog (unless the master is presently running without
+      --log-bin).
+    */
+    goto err;
   case ROTATE_EVENT:
   {
     Rotate_log_event rev(buf,event_len,0);
     if (unlikely(process_io_rotate(mi,&rev)))
     {
-      pthread_mutex_unlock(&mi->data_lock);
-      DBUG_RETURN(1);
+      error= 1;
+      goto err;
     }
-    mi->ignore_stop_event= 1;
+    /*
+      Now the I/O thread has just changed its mi->master_log_name, so
+      incrementing mi->master_log_pos is nonsense.
+    */
     inc_pos= 0;
     break;
   }
   default:
-    mi->ignore_stop_event= 0;
     inc_pos= event_len;
     break;
   }
 
-  if (likely(!ignore_event &&
-	     !(error= rli->relay_log.appendv(buf,event_len,0))))
+  /* 
+     If this event is originating from this server, don't queue it. 
+     We don't check this for 3.23 events because it's simpler like this; 3.23
+     will be filtered anyway by the SQL slave thread which also tests the server
+     id (we must also keep this test in the SQL thread, in case somebody
+     upgrades a 4.0 slave which has a not-filtered relay log).
+
+     ANY event coming from ourselves can be ignored: it is obvious for queries;
+     for STOP_EVENT/ROTATE_EVENT/START_EVENT: these cannot come from ourselves
+     (--log-slave-updates would not log that) unless this slave is also its
+     direct master (an unsupported, useless setup!).
+  */
+
+  if (uint4korr(buf + SERVER_ID_OFFSET) == ::server_id)
+  {
+    /*
+      Do not write it to the relay log.
+      We still want to increment, so that we won't re-read this event from the
+      master if the slave IO thread is now stopped/restarted (more efficient if
+      the events we are ignoring are big LOAD DATA INFILE).
+    */
+    mi->master_log_pos+= inc_pos;
+    DBUG_PRINT("info", ("master_log_pos: %d, event originating from the same server, ignored", (ulong) mi->master_log_pos));
+  }  
+  else /* write the event to the relay log */
+    if (likely(!(error= rli->relay_log.appendv(buf,event_len,0))))
     {
       mi->master_log_pos+= inc_pos;
       DBUG_PRINT("info", ("master_log_pos: %d", (ulong) mi->master_log_pos));
       rli->relay_log.harvest_bytes_written(&rli->log_space_total);
     }
+
+err:
   pthread_mutex_unlock(&mi->data_lock);
   DBUG_RETURN(error);
 }

sql/slave.h

@@ -308,8 +308,6 @@ typedef struct st_master_info
   bool old_format;			/* master binlog is in 3.23 format */
   volatile bool abort_slave, slave_running;
   volatile ulong slave_run_id;
-  bool ignore_stop_event;
-  
   
   st_master_info()
     :fd(-1), io_thd(0), inited(0), old_format(0),abort_slave(0),