DISK-FULL CONDITION OCCURS

Description:
============
Master Server crashes if binary log is rotated and disk-full
condition occurs on binlog stored partition.

Analysis:
=========
In Master Slave setup when slave requests for data from the
master, a dump thread is created and the dump tread
initially checks for the existence of binary log. If the
binary log is enabled it continuously reads data from the
binary log and sends it to slave.

If due to some critical error, as mentioned in bug scenario,
rotation of binary log fails due to disk-full scenario then
the action specified in binlog_error_action is taken.  If
users choose to ignore the error, then the binary log will
be disabled.Both the binary log file/index file are closed.

Their corresponding IO_CACHEs are cleared. The server will
continue to do its work without writing transactions into
the binary log. In such case dump thread will get EOF on
current binary log and it will try to move to next binary
log. At this point dump thread assumes that binary log is
still active and tries to read the next binary log file name
from index file IO_CACHE.  Since the IO_CACHE is in cleared
state it causes server to exit abnormally.


Fix:
===
Added a check for the existence of binary log when dump
thread is switching to next binary log. If binary log is in
disabled state, all binary logs up to the current active log
are transmitted to slave and an error is returned to the
receiver thread.

During testing valgrind issues were found in open_binlog
code. At the time of disabling binary log only the IO_CACHE
of binary log index file is cleared without closing the
actual binary log index file. This results in memory leak.
Replaced individual cleanup steps with appropriate cleanup
function call.