BUT REPLICATION FAILS

Problem:
When one enables binary logging without setting a server id, letting it
default to 0, the server changes it to 1, but replication does not work.
According to the documentation, a server id > 0, should be a legal configuration
to allow slave to connect and replication to work.

Analysis:
Apparently, this happens since the dawn of time as a design decision, meaning
that binlog => server_id.

In mysqld.cc::main()

[...]

  if (opt_bin_log && server_id == 0)
  {
    server_id= 1;
#ifdef EXTRA_DEBUG
    sql_print_warning("You have enabled the binary log, but you haven't set "
                      "server-id to a non-zero value: we force server id to
1; "
                      "updates will be logged to the binary log, but "
                      "connections from slaves will not be accepted.");
#endif
  }
[...]

What happened is that, when a slave tried to connect, it would get an error
stating that the master was misconfigured, even with server-id = 1.

After some testing we concluded that not even the documentation was accurate
enough. The test conducted were:

1) Start master with no options. Make sure that slave cannot connect (control
experience)
2) Start master with only binlog active. Make sure that slave cannot connect
(the bug situation)
3) Start master with binlog active and server-id == 0 in the command line.
Make sure that slave cannot connect. (as documented)
4) Start master with binlog active and server-id == 1 in the command line.
Change server_id to 0 using SQL before connecting slave. Make sure that slave
cannot connect.(as documented)

Situations 1) and 2) are OK and according with expected and analyzed
behavior. With 3) and 4) things are not as documented.

In Situation 3) the server starts and implicitly sets server id to 1. When
one starts a slave after configuring the master, there is no error in "SHOW
SLAVE STATUS". It connects and works. Bin and Relay log have a server id of
1.

In Situation 4) in which one explicitly change the server_id in runtime, the
behavior is the same as in Situation 3). Contents of the relay log and bin
log have a server id == 0

Fix:
For trunk only the decision was:
- Server_id must be explicitly set if binlog is active. If not, the server
will not start.
- if you set server_id=0, no slaves can connect, but your statements will be
binlogged with the provided server_id.
- if no server_id is set, it will have the default configured value and have
the same behaviour as the above, if the default is 0.

This patch will bring the code closer to which is documented. The dynamic change
of server-id was not addressed here.

For the previous version, one thinks that the documentation should be augmented
according to the information detailed in the bug page.