Skip to content

M
mariadb
Commit 477f42b1 authored by igor@rurik.mysql.com's avatar igor@rurik.mysql.com
Browse files
Options

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

into rurik.mysql.com:/home/igor/mysql-4.1
parents f928a0c4 9786806c
Show whitespace changes
Inline Side-by-side
Showing with 195 additions and 72 deletions
sql/log.cc
View file @ 477f42b1
......@@ -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
View file @ 477f42b1
......@@ -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
View file @ 477f42b1
......@@ -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
View file @ 477f42b1
......@@ -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
View file @ 477f42b1
......@@ -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),
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7