Re: [Qemu-devel] Breaking out virtfs as a standalone server?

[Rob Landley <rob@landley.net>]

On 04/11/2011 03:28 PM, Venkateswararao Jujjuri wrote:
> On 04/11/2011 06:42 AM, Rob Landley wrote:
>> Right now, there's no decent userspace server for the 9p filesystem that
>> I can find.  (In part because the 9P2000.L spec is an undocumented work
>> in progress.)
> This statement is true for 9P2000.L protocol;

According to my research on the topic, anyway:

  http://landley.livejournal.com/48698.html

> But for older protocols we have standalone servers like spfs/npfs.
> http://sourceforge.net/projects/npfs/

This would be the one that has no documentation or web page, builds
without ./configure, and has no way to specify which directory to export
but apparently can ONLY export / on the host?

> http://9p.cat-v.org/implementations

I looked at those.  Several are unfinished libraries, some also only
export / and treat the whole "restrict what you're exporting to a
subdir" problem as inherently unsolvable (readlink -f), the main one
everybody seems to test with is the Inferno equivalent of usermode
Linux, there's a python server that refuses to run without some third
party encryption/authentication library that it never OCCURRED to their
developers you might want t disable...

I also tracked down more like http://code.google.com/p/diod/ (and
emailed its author for a while: that project is stalled due to his
desire to rewrite large chunks of it, and a lack of time).

I also subscribed to v9fs-users, which is not the world's highest
traffic list:

  http://sourceforge.net/mailarchive/forum.php?forum_name=v9fs-users

>> The only up-to-date server seems to be virtfs in qemu, which has no TCP
>> transport layer.
>>
>> Are there any plans to:
>>
>> A) Add a TCP transport layer so we can test with something we can
>> intercept/examine/log/redirect with netcat and such?
>
> No plans as of now; I know folks in the Latchesar Ionkov attempted char
> dev transport.

NFS works over TCP.  Samba works over TCP.  But not p9.  Is there some
reason to go out of the way to avoid it?

> Not sure the latest though.
> 
> http://sourceforge.net/mailarchive/forum.php?thread_name=AANLkTim4eZttAmaNQfOuM1h7cmLvO-osckHNunMvG7o%2B%40mail.gmail.com&forum_name=v9fs-developer

Ah, if I want information about userspace servers I should subscribe to
v9fs-developer instead of v9fs-user, because it has no users yet.  Got it.

>> B) Break the 9p server out so it could be built as a standalone
>> userspace program?
> 
> No plans yet..and I think this is a bigger discussion.

If this filesystem is to become more than an academic exercise, it needs
a server that can export a specific directory.

> Being part of QEMU brings few implicit advantages like simplicity in
> sharing, security
> and performance advantage. I think taking it out can have its own merits.

I wasn't suggesting removing it from QEMU.  Having it in QEMU is great,
the code is written and works, it's useful as-is, and half the _idea_ of
this is that it's simpler than samba or NFS (which is damning with faint
praise, I know).

In theory, all the actual protocol encoding and decoding (and doing the
read/write/stat stuff on the host) is a single C file, correct?  Right
now there isn't any reference implementation of that server-side code,
but there is a working example of it.  Virtfs is a working example, just
not hooked up to a particularly useful transport.  (If it had a TCP
transport I could route it back out through a tap interface or -redir
port and use qemu as a test server... but it doesn't.  As far as I can
tell, virtio is intentionally the _least_ flexible mechanism for that
sort of thing.  Although maybe there's docs on this and I've just missed
them...)

Rob