Re: [9fans] Plan 9

[Alexander Viro <viro@math.psu.edu>]

On Mon, 5 Nov 2001, Boyd Roberts wrote:

> Most unix VFS's expect (if not insist) unix like semantics and are
> generally a horrible mess.  One good thing ULTRIX did was that you

... that being the main reason why I gave up on *BSD.  Linux VFS
was much cleaner and these days it's actually pretty decent.

> could have user mode NFS file-servers.  The down side was that NFS

You can do the same on a _lot_ of Unices, but most of them suffer
from fairly nasty deadlocks around the code handling mmap() over
NFS when server runs on the same box.  They may be harder or
easier to trigger, but if you manage to get the box low on memory
and create a situation when the only way to reclaim something is
to write out a dirty page and throw it away - expect a trouble
(page contents is sent to server, server needs memory to write it
out, memory pressure only leads to more requests of the same kind).

> was far too limiting.  I implemented a version of ftpfs on ULTRIX
> when I was at Digital.  Without going into NFS protocol problems
> the basic lesson is to _never_ implement a file-server protocol
> over an unreliable, connectionless transport.

<nod>

CODA is cleaner in that respect - exactly for these reasons.  For
ftpfs it's downright nice - especially since it has commit-on-close
semantics (open creates a copy on local fs and passes it back to
kernel, then all IO happens on local fs and only metadata-related
requests are passed to server; sync and close tell server to commit
the changes; there is a mechanism for handling invalidation).  For
RPC-style filesystems it's a pure hell, obviously - for that stuff
9P fits much better.