The issue here is that when a middleware failed and a recovery sequence
started, it would halt at the "call" step because the "Call" didn't
properly recover the child sequence.
An additional issue was that a Warden had no "recover" method, meaning
embedded Wardens wouldn't recover their stacks properly.
This works by registering a `config` with `:provider => true` with the
same name as your provider. Vagrant will then automatically configure
the provider when `config.vm.provider` is used.
Boxes are provider-specific, and we don't know the provider until
Environment#machine is called, so we need to build up the machine
configuration during this time.
This will eventually replace the Environment#vms method. Because of the
introduction of providers, the environment doesn't know what the backing
of the machines will be (and they're _machines_ now, not _vms_).
Instead, users of Environment will now call `#machine` on the
environment to retrieve a machine with the given backing provider as it
needs it.
This is done by calling the `upgrade` method on the _old_ configuration
classes. The old configuration classes are given the complete new
configuration and can set whatever settings they need to on it.
These are classes that use NO core classes of Vagrant, and are therefore
safe to load for upgrades. i.e. a V2 core can load a V1 config class
that is deemed upgrade safe without crashing Vagrant.
Before we were manually going over every plugin and getting each piece,
all over the place. Now we have a central manager that will give us all
the pieces we want. There is still some cleanup to do here but this is
much better overall.
The built-in middleware sequences will now be hardcoded onto
Vagrant::Action. Other plugins can hook into these sequences to provide
verification and so on. So the VirtualBox plugin will hook into that
action sequence and add verification.
This works by now calling the `:ssh` action on the provider. This action
is allowed to do whatever it pleases, but should at some point probably
call the `SSHExec` built-in middleware.
The `SSHExec` built-in middleware was added. This uses the information
returned by `Machine#ssh_info` and uses the `Vagrant::Util::SSH` helper
to exec into the remote machine. The provider should do any work upfront
in verifying that the machine is ready to be SSHed into.
This starts the transition of replacing VM with Machine. Machine still
isn't ready to fully replace VM but by moving it now, I'm able to find
the spots that need to be fixed. At this point `vagrant status` works
with the new provider interface.
This is the class that will represent a machine that is managed by
Vagrant. The class has a number of attributes associated with it and is
meant to be a single API for managing the machines.
In the case that not provider is given then whatever provider the box
represents will be added to the system. Ideally, a provider will be
given, but if not, Vagrant still does a "best effort" to install the
box.
This involved defaulting all box searching at the moment to VirtualBox.
Additionally, box upgrading is not yet handled. This needs to be done at
some point.
The box collection can now find new-style boxes with providers and
return proper Box objects. In the future, we'll also have to implement
upgrading old style ones as well.
This is the beginning of the new box internals. The basic idea is that
the new box has a new field: provider. The provider will describe what
provider that box was built with and what provider it is made to work
with.
The future of subclassing things like configuration bases and so on will
be to use `Vagrant.plugin(version, component)`. For example:
`Vagrant.plugin("1", :provisioner)`.
Before, the tempfile "f" could be GC'd before the path was used,
resulting in failed tests because when it is GC'd the tempfile is
removed. We now store the tempfile in an instance variable so that it
isn't even available for GC until after the test is finished running.
Tests before were picking up a Vagrantfile in the Vagrant source
directory, which can cause some funny failures. This ensures that each
test run will actually establish a new temporary CWD so that a
Vagrantfile is hopefully never found.
The basic process for this is to:
1. Load the configuration using the proper loader for that version. i.e.
if you're loading V1 config, then use the V1 loader.
2. If we just loaded a version that isn't current (imagine we're
currently at V3), then we need to upgrade that config. So we first
ask the V2 loader to upgrade the V1 config to V2, then we ask the V3
loader to upgrade the V2 config to V3. We keep track of warnings and
errors throughout this process.
3. Finally, we have a current config, so we merge it into the in-process
configuration that is being loaded.
This moves out the concept of a "default VM" from the Environment class
and makes it the responsibility of the V1 configuration that at least
one VM is defined on it. This lets the configuration ultimately decide
what a "default" implementation is.
This is useful so that it can take a look at the final loaded
configuration object and possibly make some tweaks to the configuration
object. The use case this was built for was so that config V1 can verify
that there is always at least a single VM defined as a sub-VM, the
"default" VM.
Previously, all procs were assumed to just be the current version. This
is certainly not going to be true always so now the version number of
the configuration must be explicit if you're assigning a proc to the
configuration loader.
This means that the Config::Loader now only knows how to load
configuration for versions used to initialize the class. This lets
things like the tests be completely isolated from what the actual
configuration is for Vagrant. This will be immensely useful to verify
that the loader functionality works for non-trivial bits (like
upgrading) without depending on Vagrant's upgrading functionality.
Vagrant is only guaranteeing that the plugin definition superclass (the
Vagrant.plugin("1") part) is backwards compatible. Anything else, such
as Vagrant::Command::Base and so on, will likely change in future
versions. Beacuse of this, plugins should only immediately expose their
definition.
In order to support loading the other classes, plugins should defer
loading to the "activation" phase of a plugin. This can be done using
the `activated` block:
class MyPlugin < Vagrant.plugin("1")
name "my plugin"
activated do
require "myplugin/my_command"
end
command("foo") { MyCommand }
end
Plugin activation is done at two specific times:
* Right when a Vagrant::Environment is created and the global plugins
(such as from ~.vagrantrc) are loaded.
* Right before loading configuration, but after the Vagrantfiles have
been evaluated. This allows plugins to be defined within these files
as well.
Vagrant.configure is now how configuration is done in Vagrantfiles
(previously it was Vagrant::Config.run). This function takes a single
argument which is the version of configuration to use.
Various internals were updated for this new versioned configuration.
Note that multiple versions of configuration aren't yet used so aren't
fully supported by Vagrant, but the foundation is being set here.
Easy commands are well... easy! They don't offer the full power of
creating a completely custom command class, but they let you do the
basics (what almost everyone needs) with minimal fuss. Example:
class MyPlugin < Vagrant.plugin("1")
name "my-plugin"
easy_command "foo" do |action|
puts "HELLO!"
end
end
NOTE: The "action" stuff isn't done yet, but will be soon!
Config::Loader will be the new class responsible for loading configuration
and replaces the previous dual-role "Vagrant::Config" played. While this
commit is very early-stage, once this new architecture is flushed out, it
will make loading, using, and extending configuration much easier and cleaner.
Additionally, I believe this will help post Vagrant 1.0 if multi-language
configuration is implemented.
Mitchell Hashimoto
skim
Nate Smith
Ryan LeCompte
Sean Wolfe
Bob Van Zant