The daemon creates a baseclass of either FuUsbDevice or FuUdevDevice when the
devices are added or coldplugged to match the quirk database and to find out
what plugin to run.
This is proxied to plugins, but they are given the GUsbDevice or GUdevDevice and
the FuDevice is just thrown away. Most plugins either use a FuUsbDevice or
superclassed version like FuNvmeDevice and so we re-create the FuDevice, re-probe
the hardware, re-query the quirk database and then return this to the daemon.
In some cases, plugins actually probe the hardware three times (!) by creating
a FuUsbDevice to get the quirks, so that the plugin knows what kind of
superclass to create, which then itself probes the hardware again.
Passing the temporary FuDevice to the plugins means that the simplest ones can
just fu_plugin_device_add() the passed in object, or create a superclass and
incorporate the actual GUsbDevice and all the GUIDs.
This breaks internal plugin API but speeds up startup substantially and deletes
a lot of code.
GLib creates two static inline functions for paramaters that may
not be used that set off warnings in clang but not gcc.
Ignore these on clang builds everywhere that
G_DEFINE_AUTOPTR_CLEANUP_FUNC is used.
This makes more sense; we're updating the device, not the plugin itself.
This also means we don't need to funnel everything through callbacks like
GFileProgressCallback and we can also update the state without adding an
explicit callback to each derived device type.
It's actually less scary to see a SHA1 hash than it is to see a path like
/sys/devices/pci0000:00/0000:00:1d.0/usb1/1-1/1-1.2. It's also way easier to
copy and paste into the various fwupdmgr command that require a device ID and
also means we can match a partial prefix much like git allows.
If we also move to a model where plugins can be changed during different stages
of the update (e.g. during detach) then the device might change connection type
and then the sysfs path not only becomes difficult to paste, but incorrect.
Session software doesn't care about the format of the device ID (it is supposed
to be an implementation detail) and so there's no API or ABI break here. A few
plugins also needed to be ported, but nothing too worrying.