This PR refactors all of the methods on `Vfs` from accepting `&mut self` to
accepting `&self` and keeping data wrapped in a mutex. This builds on previous
changes to make reference count file contents and cleans up the last places
where we're returning borrowed data out of the VFS interface.
Once this change lands, there are two possible directions we can go that I see:
* Conservative: Refactor all remaining `&mut Vfs` handles to `&Vfs`
* Interesting: Embrace ref counting by changing `Vfs` methods to accept `self:
Arc<Self>`, which makes the `VfsEntry` API no longer need an explicit `Vfs`
argument for its operations.
* Change VfsFetcher to be immutable with internal locking
* Refactor Vfs::would_be_resident
* Refactor Vfs::read_if_not_exists
* Refactor Vfs::raise_file_removed
* Refactor Vfs::raise_file_changed
* Add Vfs::get_internal as bits of Vfs::get
* Switch Vfs to use internal locking
* Migrate all Vfs methods from &mut self to &self
* Make VfsEntry access Vfs immutably
* Remove outer VFS locking (#260)
* Refactor all snapshot middleware to accept &Vfs instead of &mut Vfs
* Remove outer VFS Mutex across the board
Initialization logic needed for serve, build, and upload is now
much more clear than it was when these functions were written.
This commit refactors all of them to use a new common_setup
module for all of their initialization that's the same.
Starts work on #55.
This is similar to the previous work in #125. It's gated behind a new Cargo
feature, `user-plugins`. This time, the config gate is much smaller. The
`plugins` member of projects is still accessible when plugins aren't enabled,
but is always empty. Additionally, user plugins are only enabled if there's a
Lua state present in the snapshot context when the `SnapshotUserPlugins`
snapshot middleware runs. This not ever the case currently.
This code has very little possibility of rotting while we focus on other work,
since it'll be guaranteed to still compile and can be tested in isolation
without the feature being turned on.
