Writing a multithreaded daemon process

[Zhihui Zhang]

After reading part of the Arla daemon code available at
http://www.stacken.kth.se/projekt/arla/, I feel it is a real good design. 
The daemon code is a multi-threaded user process that does most of the job
and the kernel is an interface to VFS. They communicate with each other
via a character device.

But I am wondering if the performance is good since you sometimes have to
cross user-system boundary up to four times for a single read operations:

 read() --> Arla daemon --> network  --> install data into kernel

If the daemon can somehow reside entirely inside the kernel, like NFS
daemon, we can save those crossings.  But the daemon is a multi-threaded
process and we have no kernel thread yet, so I do not know how to do
better if possible.  Maybe all user filesystems have to live with this
limitation.

Any help is appreciated.

-Zhihui