can't create two temporary files

btw, this is not a bug in tmpfile() from php, as the following test works:

  function testTmpFile() {
    $eval = '
$tmpfile0 = tmpfile();
$tmpfile1 = tmpfile();
$tmpnam0 = stream_get_meta_data($tmpfile0);
$tmpnam1 = stream_get_meta_data($tmpfile1);
print json_encode(file_exists($tmpnam0["uri"]) && file_exists($tmpnam1["uri"]));
';
    $this->drush('php-eval', array($eval));
    $exist = json_decode($this->getOutput());
    $this->assertTrue($exist);
  }

Even more confusing and marvelous: when commenting out the "static" declaration, no file is created at *all*. Huge wtf.

Simpler test code also shows that the first tempfile is actually removed when the second tmp call is issued:

$tmpfile0 = drush_save_data_to_temp_file("foo");
print "file 0: $tmpfile0 exists? " . file_exists($tmpfile0) . "\n";
$tmpfile1 = drush_save_data_to_temp_file("bar");

print "file 0: $tmpfile0 exists? " . file_exists($tmpfile0) . "\n";
print "file 1: $tmpfile1 exists? " . file_exists($tmpfile1) . "\n";

yields:

file 0: /tmp/php0sbcYN exists? 1
file 0: /tmp/php0sbcYN exists?
file 1: /tmp/phpNOXNE2 exists? 1

So I think the problem resides in the "magic" of tmpfile(), which "magically" removes the file descriptor when it is closed (or in our case, garbage collected). Lots of fun. The following patch works, for example:

index 8dfaf35..830437b 100644
--- a/includes/drush.inc
+++ b/includes/drush.inc
@@ -1798,9 +1798,10 @@ function drush_mkdir($path) {
  *   A path to the file.
  */
 function drush_save_data_to_temp_file($data) {
-  static $fp;
+  static $fplist;
 
   $fp = tmpfile();
+  $fplist[] = $fp;
   fwrite($fp, $data);
   $meta_data = stream_get_meta_data($fp);
   $file = $meta_data['uri'];

...simply because we keep a reference to all the temporary files we create.

Drush 5 is not affected, because it doesn't use tmpfile() but tmpname(), which has the disadvantage of being vulnerable to a symlink attack, as we do not get an open file descriptor but a pathname which may be replaced during the very very short window during which we try to open the file. It is unclear to me whether we should use Drush 5's version of the above hack. Both alternatives seem problematic.

Anyways, with the above patch, the provided test passes and the whole test suite is in the green again.

From my latest comments, it seemed that Drush 5 wasn't affected by this bug, so Drush 6 is likely unaffected too.

However, both are probably vulnerably to a symlink attack, so there's a security issue here, but that's a separate issue.

I am closing this one as fixed in 5.x, according to my latest comments.