On Jul 14, 2009, at 6:08 AM, roger peppe <rogpe...@gmail.com> wrote:
this is at a bit of a tangent from the previous discussion, but something i've always wondered: why does the linux 9p mount syscall bother with IP addresses at all? isn't it sufficient just to provide a facility for mounting a file descriptor (like the plan 9 syscall) and have an auxiliary command do the actual dial, authentication, etc?
This has always been an option. The tcp transport is there (and is default) because it is closest to the expected behavior for the naive Linux user. It is also there because at one time we used 9p as root (ala nfsroot direct option in the kernel).
It's worth noting there are performance/efficency implications for using the fd transport (particularly for the non-socket fd case like sshsrv) as well as the unix named pipes transport -- but I don't have quantitative analysis of the differences as it's hard to get an apples to apples comparison. Anthony ligouri has shown some really promising numbers for a hybrid transport based on the Linux splice command.
wouldn't that be simpler
The code in question is mostly common, maybe 20 lines of code for each tcp and unix, so I think it's worth having them as options so that user space mount tools arent required.
