Fix for bug in LOAD DATA INFILE and replication

Fix for SHOW VARIABLES in embedded server

Docs/internals.texi

@@ -96,13 +96,84 @@ cached for each user/database combination.
 Many use of @code{GROUP BY} or @code{DISTINCT} caches all found rows in 
 a @code{HEAP} table. (This is a very quick in-memory table with hash index.)
 
-@item Join Row Cache
+@item Join buffer Cache
 For every full join in a @code{SELECT} statement (a full join here means
 there were no keys that one could use to find the next table in a list), 
 the found rows are cached in a join cache.  One @code{SELECT} query can 
 use many join caches in the worst case.
 @end table
 
+@node join_buffer_size, flush tables, caching, Top
+@subchapter How MySQL uses the join_buffer cache
+
+Basic information about @code{join_buffer_size}:
+
+@itemize @bullet
+@item
+It's only used in the case when join type is of type @code{ALL} or
+@code{index}; In other words: no possible keys can be used.
+@item
+A join buffer is never allocated for the first not-const table,
+even it it would be of type @code{ALL}/@code{index}.
+@item
+The buffer is allocated when we need to do a each full join between two
+tables and freed after the query is done.
+@item
+Accepted row combinations of tables before the @code{ALL}/@code{index}
+able is stored in the cache and is used to compare against each read
+row in the @code{ALL} table.
+@item
+We only store the used fields in the join_buffer cache, not the
+whole rows.
+@end itemize
+
+Assume you have the following join:
+
+@example
+Table name	Type
+t1		range
+t2		ref
+t3		@code{ALL}
+@end example
+
+The join is then done as follows:
+
+@example
+- While rows in t1 matching range
+ - Read through all rows in t2 according to reference key
+  - Store used fields form t1,t2 in cache
+  - If cache is full
+    - Read through all rows in t3
+      - Compare t3 row against all t1,t2 combination in cache
+        - If rows satisfying join condition, send it to client
+    - Empty cache
+
+- Read through all rows in t3
+ - Compare t3 row against all stored t1,t2 combinations in cache
+   - If rows satisfying join condition, send it to client
+@end example
+
+The above means that table t3 is scanned
+
+@example
+(size-of-stored-row(t1,t2) * accepted-row-cominations(t1,t2))/
+join_buffer_size+1
+@end example
+times.
+
+Some conclusions:
+
+@itemize @bullet
+@item
+The larger the join_buff_size, the fewer scans of t3.
+If @code{join_buff_size} is already large enough to hold all previous row
+combinations then there is no speed to gain by making it bigger.
+@item
+If there is several tables of @code{ALL}/@code{index} then the we
+allocate one @code{join_buffer_size buffer} for each of them and use the
+same algorithm described above to handle it.  (In other words, we store
+the same row combination several times into different buffers)
+@end itemize
 
 @node flush tables, filesort, caching, Top
 @chapter How MySQL Handles @code{FLUSH TABLES}

configure.in

@@ -4,7 +4,7 @@ dnl Process this file with autoconf to produce a configure script.
 AC_INIT(sql/mysqld.cc)
 AC_CANONICAL_SYSTEM
 # The Docs Makefile.am parses this line!
-AM_INIT_AUTOMAKE(mysql, 4.0.8-gamma)
+AM_INIT_AUTOMAKE(mysql, 4.0.9-gamma)
 AM_CONFIG_HEADER(config.h)
 
 PROTOCOL_VERSION=10

sql/log_event.cc

@@ -206,7 +206,19 @@ Log_event::Log_event(const char* buf, bool old_format)
 
 int Log_event::exec_event(struct st_relay_log_info* rli)
 {
-  if (rli)					// QQ When is this not true ?
+  /*
+    rli is null when (as far as I (Guilhem) know)
+    the caller is
+    Load_log_event::exec_event *and* that one is called from
+    Execute_load_log_event::exec_event. 
+    In this case, we don't do anything here ;
+    Execute_load_log_event::exec_event will call Log_event::exec_event
+    again later with the proper rli.
+    Strictly speaking, if we were sure that rli is null
+    only in the case discussed above, 'if (rli)' is useless here.
+    But as we are not 100% sure, keep it for now.
+  */
+  if (rli)  
   {
     if (rli->inside_transaction)
       rli->inc_pending(get_event_len());
@@ -1773,8 +1785,34 @@ int Query_log_event::exec_event(struct st_relay_log_info* rli)
   return Log_event::exec_event(rli); 
 }
 
+/*
+  Does the data loading job when executing a LOAD DATA on the slave
+
+  SYNOPSIS
+    Load_log_event::exec_event
+      net  
+      rli                             
+      use_rli_only_for_errors	  - if set to 1, rli is provided to 
+                                  Load_log_event::exec_event only for this 
+				  function to have RPL_LOG_NAME and 
+				  rli->last_slave_error, both being used by 
+				  error reports. rli's position advancing
+				  is skipped (done by the caller which is
+				  Execute_load_log_event::exec_event).
+				  - if set to 0, rli is provided for full use,
+				  i.e. for error reports and position
+				  advancing.
+
+  DESCRIPTION
+    Does the data loading job when executing a LOAD DATA on the slave
+ 
+  RETURN VALUE
+    0           Success                                                 
+    1    	Failure
+*/
 
-int Load_log_event::exec_event(NET* net, struct st_relay_log_info* rli)
+int Load_log_event::exec_event(NET* net, struct st_relay_log_info* rli, 
+			       bool use_rli_only_for_errors)
 {
   init_sql_alloc(&thd->mem_root, 8192,0);
   thd->db = rewrite_db((char*)db);
@@ -1836,8 +1874,12 @@ int Load_log_event::exec_event(NET* net, struct st_relay_log_info* rli)
 		     TL_WRITE))
 	thd->query_error = 1;
       if (thd->cuted_fields)
+	/* 
+	   log_pos is the position of the LOAD
+	   event in the master log
+	*/
 	sql_print_error("Slave: load data infile at position %s in log \
-'%s' produced %d warning(s)", llstr(rli->master_log_pos,llbuff), RPL_LOG_NAME,
+'%s' produced %d warning(s)", llstr(log_pos,llbuff), RPL_LOG_NAME, 
 			thd->cuted_fields );
       if (net)
         net->pkt_nr= thd->net.pkt_nr;
@@ -1877,7 +1919,7 @@ int Load_log_event::exec_event(NET* net, struct st_relay_log_info* rli)
     return 1;
   }
 
-  return Log_event::exec_event(rli); 
+  return ( use_rli_only_for_errors ? 0 : Log_event::exec_event(rli) ); 
 }
 
 
@@ -2132,7 +2174,11 @@ int Execute_load_log_event::exec_event(struct st_relay_log_info* rli)
   save_options = thd->options;
   thd->options &= ~ (ulong) (OPTION_BIN_LOG);
   lev->thd = thd;
-  if (lev->exec_event(0,0))
+  /*
+    lev->exec_event should use rli only for errors
+    i.e. should not advance rli's position
+  */
+  if (lev->exec_event(0,rli,1)) 
   {
     slave_print_error(rli,my_errno, "Failed executing load from '%s'", fname);
     thd->options = save_options;

sql/log_event.h

@@ -417,9 +417,10 @@ public:
   const char* get_db() { return db; }
   int exec_event(struct st_relay_log_info* rli)
   {
-    return exec_event(thd->slave_net,rli);
+    return exec_event(thd->slave_net,rli,0);
   }
-  int exec_event(NET* net, struct st_relay_log_info* rli);
+  int exec_event(NET* net, struct st_relay_log_info* rli, 
+		 bool use_rli_only_for_errors);
 #else
   void print(FILE* file, bool short_form = 0, char* last_db = 0);
 #endif

sql/slave.cc

@@ -2343,12 +2343,6 @@ static int process_io_rotate(MASTER_INFO *mi, Rotate_log_event *rev)
 
   memcpy(mi->master_log_name, rev->new_log_ident, rev->ident_len+1);
   mi->master_log_pos= rev->pos;
-
-  pthread_mutex_lock(&mi->rli.data_lock);
-  memcpy(mi->rli.master_log_name, rev->new_log_ident, rev->ident_len+1);
-  mi->rli.master_log_pos= rev->pos;
-  pthread_mutex_unlock(&mi->rli.data_lock);
-
   DBUG_PRINT("info", ("master_log_pos: '%s' %d",
 		      mi->master_log_name, (ulong) mi->master_log_pos));
 #ifndef DBUG_OFF

sql/sql_show.cc

@@ -1217,6 +1217,7 @@ int mysqld_show(THD *thd, const char *wild, show_var_st *variables,
       case SHOW_RPL_STATUS:
 	net_store_data(&packet2, rpl_status_type[(int)rpl_status]);
 	break;
+#ifndef EMBEDDED_LIBRARY	
       case SHOW_SLAVE_RUNNING:
       {
 	LOCK_ACTIVE_MI;
@@ -1226,6 +1227,7 @@ int mysqld_show(THD *thd, const char *wild, show_var_st *variables,
 	UNLOCK_ACTIVE_MI;
 	break;
       }
+#endif
       case SHOW_OPENTABLES:
         net_store_data(&packet2,(uint32) cached_tables());
         break;