[Patch] PR 31228, Fix close-on-exec race.

[Andrew Haley]

David Daney writes:
 > Andrew Haley wrote:
 > > David Daney writes:
 
 > >  > How about this version?
 > >
 > > This is impressively heroic, but I really think you should measure the
 > > performance improvement before committing this.  :-)
 >
 > Hard data takes the fun out of things :-(

Haha!  I am *such* a scientist!  :-)

 > What does all this mean?  Well on my x86_64-pc-linux-gnu notebook,
 > it takes about 1 second to close 10,000,000 times an invalid file
 > descriptor.  The savings of only closing open files would appear to
 > make sense only if the default ulimit -n value is raised to a value
 > greater than about 3000.
 > 
 > The only reason to increase ulimit -n, from its default value of
 > 1024, is if you expect to have a lot of open files.  In this case
 > This is all moot as the time to close the open files would be much
 > greater than the overhead of closing invalid descriptors.
 > 
 > So if we want to fix the bug, I would recommend my original patch.

Yeah, we want to fix the bug.  This is OK.

Andrew.