Thread: pg_upgrade -j broken on Windows

pg_upgrade -j broken on Windows

From
Bruce Momjian
Date:
Everyone should be aware that the 9.3 pg_upgrade -j/--jobs option on
Windows is currently broken, and hopefully will be fixed by the next
beta.

Someone at PGDay UK told me they were getting pg_upgrade -j crashes on
Windows.  Andrew Dunstan was able to reproduce the crash, and that has
been fixed, but there is still a race condition that I am trying to
diagnose.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + It's impossible for everything to be true. +



Re: pg_upgrade -j broken on Windows

From
Bruce Momjian
Date:
On Thu, Jul 25, 2013 at 10:57:28AM -0400, Bruce Momjian wrote:
> Everyone should be aware that the 9.3 pg_upgrade -j/--jobs option on
> Windows is currently broken, and hopefully will be fixed by the next
> beta.
> 
> Someone at PGDay UK told me they were getting pg_upgrade -j crashes on
> Windows.  Andrew Dunstan was able to reproduce the crash, and that has
> been fixed, but there is still a race condition that I am trying to
> diagnose.

After three days of testing and creating a Windows MinGW build
environment (with Andrew's help), I have come up with the attached patch
which fixes the pg_upgrade -j race condition on Windows.  In summary,
creating a file with fopen() from a non-primary thread and then calling
system() soon after can result in a file-in-use error.  The solution was
to issue the fopen() after system() in such cases.

This would be applied to head and 9.3.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + It's impossible for everything to be true. +



Re: pg_upgrade -j broken on Windows

From
Bruce Momjian
Date:
On Fri, Jul 26, 2013 at 01:27:34PM -0400, Bruce Momjian wrote:
> On Thu, Jul 25, 2013 at 10:57:28AM -0400, Bruce Momjian wrote:
> > Everyone should be aware that the 9.3 pg_upgrade -j/--jobs option on
> > Windows is currently broken, and hopefully will be fixed by the next
> > beta.
> >
> > Someone at PGDay UK told me they were getting pg_upgrade -j crashes on
> > Windows.  Andrew Dunstan was able to reproduce the crash, and that has
> > been fixed, but there is still a race condition that I am trying to
> > diagnose.
>
> After three days of testing and creating a Windows MinGW build
> environment (with Andrew's help), I have come up with the attached patch
> which fixes the pg_upgrade -j race condition on Windows.  In summary,
> creating a file with fopen() from a non-primary thread and then calling
> system() soon after can result in a file-in-use error.  The solution was
> to issue the fopen() after system() in such cases.
>
> This would be applied to head and 9.3.

Sorry, patch attached.

--
  Bruce Momjian  <bruce@momjian.us>        http://momjian.us
  EnterpriseDB                             http://enterprisedb.com

  + It's impossible for everything to be true. +

Attachment

Re: pg_upgrade -j broken on Windows

From
Bruce Momjian
Date:
On Fri, Jul 26, 2013 at 01:31:45PM -0400, Bruce Momjian wrote:
> On Fri, Jul 26, 2013 at 01:27:34PM -0400, Bruce Momjian wrote:
> > On Thu, Jul 25, 2013 at 10:57:28AM -0400, Bruce Momjian wrote:
> > > Everyone should be aware that the 9.3 pg_upgrade -j/--jobs option on
> > > Windows is currently broken, and hopefully will be fixed by the next
> > > beta.
> > > 
> > > Someone at PGDay UK told me they were getting pg_upgrade -j crashes on
> > > Windows.  Andrew Dunstan was able to reproduce the crash, and that has
> > > been fixed, but there is still a race condition that I am trying to
> > > diagnose.
> > 
> > After three days of testing and creating a Windows MinGW build
> > environment (with Andrew's help), I have come up with the attached patch
> > which fixes the pg_upgrade -j race condition on Windows.  In summary,
> > creating a file with fopen() from a non-primary thread and then calling
> > system() soon after can result in a file-in-use error.  The solution was
> > to issue the fopen() after system() in such cases.
> > 
> > This would be applied to head and 9.3.
> 
> Sorry, patch attached.

Applied, and backpatched to 9.3.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + It's impossible for everything to be true. +