• Contact: tech-kern
• Duration estimate: 3 months

IMPORTANT: This project was completed by Tomohiro Kusumi <tkusumi@netbsd.org> (a port from DragonFlyBSD which was a port from FreeBSD). You may still contact the people above for details, but please do not submit an application for this project.

NetBSD currently uses amd for automatically mounting (network) file systems. This software package implements an automounter file system as a userland NFS daemon. While this generally works it has major drawbacks:

• File systems are not mounted directly on the desired mount point. As a result applications frequently use incorrect pathnames (e.g. /amd/server/home/user instead of /home/user) for automatically mounted directories or files beneath them. This is especially problematic in heterogeneous enviroments where not all machines use the same automounter.
• The automounter daemon cannot handle high I/O load very well; file access occasionally fails with intermittent errors.
• Userland NFS daemons are prone to deadlocking.

The milestones of this project are:

• implement a new automounter solution which has configurable mount points
• improve behavior under load I/O
• show benchmarks and implement automated tests

There are at least two possible approaches: one is to port FreeBSD's autofs(4), which is an in-kernel file system. This has the advantage of already existing, and being at least loosely Linux and Solaris compatible. (One could also write a new autofs but that doesn't seem like a dominant strategy.)

Another possible approach is a new userland daemon using puffs.

Note that automount configuration is a complex and largely unsolved problem. While it's desirable to be compatible with existing automounter configuration (because often organizations like to distribute automounter maps through NIS or other facilities) it's also desirable to have a simple, comprehensible, and cogent configuration scheme and these desires are rather at odds.