This can occur because prior to this, there is a code path that looks
like this:
primaryConn = myLocalConn
CloseConnections will subsequently try to free both with PQFinish.
I'm not sure if this is the right fix -- it's more of hack -- without
more information about design intention.
One reasonable alternative would be to have CloseConnections perform
this check itself. As-is I am pretty sure that this fix leaves a
signal-race (when CloseConnections is called, without the check, in
the interrupt handler) unfixed.
Signed-off-by: Dan Farina <drfarina@acm.org>
Signed-off-by: Peter van Hardenberg <pvh@heroku.com>
A result is being cleared while there are still pointers that
refer to datums in it.
Signed-off-by: Dan Farina <drfarina@acm.org>
Signed-off-by: Peter van Hardenberg <pvh@heroku.com>
This is done in a separate patch to try and reduce the sound and fury
of the patch that actually did the conversion from sprintf to
snprintf-alikes.
Signed-off-by: Dan Farina <drfarina@acm.org>
Signed-off-by: Peter van Hardenberg <pvh@heroku.com>
Move out string operations to another file, and introduce a frontend
to snprintf for various situations. This change is important for
catching and eliminating sprintf overflows, which are as of now many
times silently corrupting memory.
Signed-off-by: Dan Farina <drfarina@acm.org>
Signed-off-by: Peter van Hardenberg <pvh@heroku.com>
pgsql conventions (tabs, four-spaces-wide, etc) applied all around.
Also tried to fix some very tiny capitalization errors, auto-fill
problems, and some inter-block vertical whitespacing issues.
Long strings in repmgr.c were left intact, though. They are rather
numerous and are less of a problem than tiny bits of function calls
and comments wrapping over a line; the latter kind of problem has been
mostly fixed.
Signed-off-by: Dan Farina <drfarina@acm.org>
Signed-off-by: Peter van Hardenberg <pvh@heroku.com>
PostgreSQL's GUC wal_keep_segments should have for "standby clone" to work
reliable. We default to 5000 but it could be changed for the user if it
knows what he is doing.
then try to get a new master 30 times (every 10 min) before exit.
It's not retrying forever because after some attempts it gives an
error of multiple files opened.
the main difference is that this new function doesn't return a bool but
a string representing the current major version of the postgresql server
or an empty string if it's earlier than 9.0.
Teach all commands that could connect to master and standby that they
both should have the same major version.
