The Ruby Tempfile class has a finalizer that removes the file when the
GC runs. It's not predictible when the GC will run, so you have to
ensure that the instance of the class stays in scope for as long as you
need it.
Unfortunately the tempfile method is returning just the filename of the
temporary file, which means it goes out of scope when that method
returns. This allows the GC to reap it at any time after return.
In both CI and production environments we've seen this race fail,
causing apt-key add to fail a small (2-3%) amount of the time.
This changes the tempfile and source_to_file methods to return the
underlying Tempfile object, pushing it up into the caller's scope. Both
of the callers immediately use the object to get its filename and then
open the file, eliminating the race.
Tested this by adding 'GC.start; sleep(1)' immediately before the
command is run, to give the GC plenty of time to remove the tempfile if
it was going to.
If one should happen to have redirected /usr/bin/gpg to run GnuPG 2.x
rather than the more usual GnuPG 1.x, the apt_key provider fails with
the following error:
Could not prefetch apt_key provider 'apt_key': invalid date
This is because the output of "--with-colons" defaults to using
"fixed-list-mode" in 2.x but did not do so for 1.x. This new format
gives much more information about keys and also uses timestamps in
seconds from 1970-01-01 (UNIX epoch) rather than dates in the format
YYYY-MM-DD.
This patch adds "--fixed-list-mode" when calling apt-key, and adjusts
the code to parse the timestamps instead. This actually has several
advantages:
- Works the same with GnuPG 1.x and 2.x.
- More accurate expiry time tracking, not just entire days.
- No need to require 'date' any longer.
- Will allow the provider to expose more key information in future.
Tested on:
- Debian Wheezy (Puppet 2.7.23, Ruby 1.8.7p358)
- Debian Jessie (Puppet 3.7.2, Ruby 2.1.5p273)
better attempt at gpg version checking
adding in key length warning
removing version check, adding key check
adding tests
clean up the code
small changes
use commands
documentation updates
For some reason, even though our own tests pass with `require
puppet_x/apt_key/patch_openuri` every other module that depends on
puppetlabs-apt is now breaking in tests claiming it can't load that
module.
Somewhere along the way something is probably messing with LOADPATH
causing this to trip up. This should fix the issues for everyone.
This commit affects:
* puppetlabs/puppetlabs-apt#229
* puppetlabs/puppetlabs-postgresql#391
This commits introduces:
* The apt_key type;
* The apt_key provider;
* Unit tests for the type;
* Beaker/acceptance tests for the type/provider.
The idea behind apt_key is that apt::key will simply become a wrapper
that uses apt_key. Being a native type/provider apt_key is a lot less
error prone than the current exec behaviour of apt::key and adds a few
nice bonuses like inventory capabilities for mcollective users.
