This conversion is unfortunate in the sense that Unix local domain sockets are relatively lightweight. LocalHost UDP sockets are much heavier weight since they rely on the full UDP stack. If anyone is up for a complete redesign, then using some shared memory and a POSIX message queue would be lightweight again.
This commit also fixes several bugs that were not testable before the inode tree deadlock. I cannot say that the logic is 100% stable but it does not have basic functionality.
Squashed commit of the following:
fs/userfs: Order locking so that access to the shared I/O buffer is also locked.
fs/userfs: Converts to use LocalHost UDP loopback for IPC.
fs/userfs: This completes coding of the UserFS client and of the UserFS feature in general. This feature is being merged to main now because I believe it is innocuous. It is, however, untesed. The next step will be to develop a test case to verify the feature.
fs/userfs: Completes the request logic for the UserFS client. Still need the logic that receives the responses.
fs/userfs: Completes coding for most of the server side of the user filesystem logic.
fs/userfs: Big design changes, simplications. Use Unix domain local sockets instead of message queues. Easier to transfer big data in local sockets than message queues. Remove character drvier 'factory' it is not necessary.
fs/userfs: Minor reparitioning; volume private info does not need to be held on the OS client side.
libc/userfs: Add some of the server side logic.
fs/userfs: Add some UserFS initialization logic.
fs/userfs: Add frame work for the UserFS proxy. Remove all references to a block driver. There is no block dricer... what was I thinking?
fs/userfs: Add some initialization of the character driver, 'factory' device.
fs/userfs: Rename from fusefs to userfs to that we don't stomp on someone else's cool name.
Add a header file describing the fusefs interface.
Xiang Xiao