[9fans] linux il/ip

[9fans] linux il/ip

[erik quanstrom]

i want to mount my fs from linux.  

has anyone implemented -- or started an implementation -- of
il for linux?  i found some thoughts of it in the archives, but
no actual information.

i'm sure it would be easier to put tcp in the fs code,
but that seemed too easy. ;-)

- erik

Re: [9fans] linux il/ip

[Sergey Zhilkin]

[-- Attachment #1: Type: text/plain, Size: 417 bytes --]

Hi !
Look at http://plan9.aichi-u.ac.jp/netlib/il/

2006/8/28, erik quanstrom <quanstro@quanstro.net>:
>
> i want to mount my fs from linux.
>
> has anyone implemented -- or started an implementation -- of
> il for linux?  i found some thoughts of it in the archives, but
> no actual information.
>
> i'm sure it would be easier to put tcp in the fs code,
> but that seemed too easy. ;-)
>
> - erik
>

[-- Attachment #2: Type: text/html, Size: 729 bytes --]

Re: [9fans] linux il/ip

[erik quanstrom]

that's not quite what i was looking for.  in order to use v9fs to mount
my fs directly (which running the fs kernel and therefore does not do tcp), 
i need the plan 9 kernel to directly recognize il packets.

- erik

Re: [9fans] linux il/ip

[Sergey Zhilkin]

[-- Attachment #1: Type: text/plain, Size: 316 bytes --]

Why not to use tcp ???

2006/8/28, erik quanstrom <quanstro@quanstro.net>:
>
> that's not quite what i was looking for.  in order to use v9fs to mount
> my fs directly (which running the fs kernel and therefore does not do
> tcp),
> i need the plan 9 kernel to directly recognize il packets.
>
> - erik
>

[-- Attachment #2: Type: text/html, Size: 565 bytes --]

Re: [9fans] linux il/ip

[erik quanstrom]

because the fileserver kernel does not offer tcp.

- erik

On Mon Aug 28 11:57:27 CDT 2006, szhilkin@gmail.com wrote:

> Why not to use tcp ???
> 

Re: [9fans] linux il/ip

[geoff]

The usual solution is to export the file server contents via a cpu
server.

Re: [9fans] linux il/ip

[erik quanstrom]

unfortunately, i do not currently have a machine that can fill that role.

- erik

Re: [9fans] linux il/ip

[lucio]

> The usual solution is to export the file server contents via a cpu
> server.

Well, I could have sworn FS had TCP built-in.  Imagine being wrong all
these years!  I guess I'm lucky I never had to test it :-)

++L

Re: [9fans] linux il/ip

[Brantley Coile]

IL works great in a local network.  I would guess that the change in
9P to better support a stream protocol like TCP was to get better
performance to system over the Internet.  One would tcp to a cpu
server serving 9P2000 which was IL'ing the fs.

My guess on the outside.  Jim?  Geoff?

 bc

void
tcprecv(Msgbuf *mb, Ifc*)
{
	mbfree(mb);
}



>> The usual solution is to export the file server contents via a cpu
>> server.
> 
> Well, I could have sworn FS had TCP built-in.  Imagine being wrong all
> these years!  I guess I'm lucky I never had to test it :-)
> 
> ++L

Re: [9fans] linux il/ip

[Skip Tavakkolian]

yes. in 2002 rob said:

	* We are phasing out the IL protocol since it doesn't handle 
	long-distance connections well (and long-distance networks 
	don't handle it well, either).  IL is still used by fs(4)
	(in time, that too will change) but TCP has become the stan-
	dard protocol for all other services.

http://tinyurl.com/o2lry

> IL works great in a local network.  I would guess that the change in
> 9P to better support a stream protocol like TCP was to get better
> performance to system over the Internet.  One would tcp to a cpu
> server serving 9P2000 which was IL'ing the fs.

Re: [9fans] linux il/ip

[erik quanstrom]

i'm not sure il's completely moot.  has anybody done any tcp vs il performance 
measurements on gigabit or 10 gigibit networks, including cpu usage?

the fact that it doesn't route well over the internet shouldn't be too much of a
bother most of the time.  my only problem is that of all the protocols that linux
choose to implement, il wasn't one of them. ;-)

- erik

Re: Re: [9fans] linux il/ip

[David Leimbach]

On 8/29/06, erik quanstrom <quanstro@quanstro.net> wrote:
> i'm not sure il's completely moot.  has anybody done any tcp vs il performance
> measurements on gigabit or 10 gigibit networks, including cpu usage?
>
> the fact that it doesn't route well over the internet shouldn't be too much of a
> bother most of the time.  my only problem is that of all the protocols that linux
> choose to implement, il wasn't one of them. ;-)

Tell me you don't use the ham radio support every day....

>
> - erik
>

Re: Re: [9fans] linux il/ip

[erik quanstrom]

okay, i'm a dunce.  please stop killing me for asking.

- erik

Re: [9fans] linux il/ip

[jmk]

il is dead, deal with it. if you like i'll remove it
from the distributed code. ken tried to put tcp into the
old fileserver and gave up in disgust.

On Tue Aug 29 15:53:57 EDT 2006, quanstro@quanstro.net wrote:
> i'm not sure il's completely moot.  has anybody done any tcp vs il performance 
> measurements on gigabit or 10 gigibit networks, including cpu usage?
> 
> the fact that it doesn't route well over the internet shouldn't be too much of a
> bother most of the time.  my only problem is that of all the protocols that linux
> choose to implement, il wasn't one of them. ;-)
> 
> - erik

Re: [9fans] linux il/ip

[Brantley Coile]

> il is dead, deal with it. if you like i'll remove it
> from the distributed code. ken tried to put tcp into the
> old fileserver and gave up in disgust.

Please don't.  We still depend on it.

Re: [9fans] linux il/ip

[Skip Tavakkolian]

>> il is dead, deal with it. if you like i'll remove it
>> from the distributed code. ken tried to put tcp into the
>> old fileserver and gave up in disgust.
> 
> Please don't.  We still depend on it.

i have the same request. thanks.

Re: [9fans] linux il/ip

[Dave Lukes]

</lurk>

> il is dead,

Empirically this is currently  untrue.

Also, a while ago, the perceptive Mr. Forsyth pointed out
that keeping >1 protocol around makes people more honest
(i.e. less inclined to make things more-or-less subtly 
protocol-dependent).

I know that's not an argument in favour of il,
but there ain't anything else is there?

Also, on a personal note,
one of the main strengths of plan9 from my pot of view is didacticism:
having a clean(er) protocol around to look at may in itself be useful.


> deal with it.

OK: I'll deal with it.  Here goes ...

	BoooHoooo!!! Bwaaaaahhhh!!!
	You BASTARD!!!!!!! YOU KILLED IL!!!!
	I will TRAVEL back in TIME and MUTATE your ANCESTORS!!!!!!!!!

There: I've dealt with it.

Just in case it causes offence: the above was a JOKE.




>  if you like i'll remove it from the distributed code.

Please don't do that:
you'll only encourage the kids to create mutant underground versions.

>  ken tried to put tcp into the old fileserver and gave up in disgust.

and you wonder why people have a lingering fondness for il?

I know nothing about il other than it's leaner than TCP,
both in implementation and on the wire,
but even that may be useful as an example of how to do it.

When in future I decide to hurl bits around a LAN at high speed,
I don't want TCP dreck vomited over my wires;
nor do I want to do the damned TCP ritual
where I have to have the damned manuals in front of
me to look at the packets because TCP is like pain:
the human mind can't recall it accurately
if it is to stay sane.

I'm sure that il ain't The Definitive Answer
but neither is TCP (definitely!)

   DaveL


<lurk reason="company has been bought">:-)

Re: [9fans] linux il/ip

[erik quanstrom]

> I know that's not an argument in favour of il,
> but there ain't anything else is there?

without doing a survey, linux has added three new ip protocols since the beginning of the
year.  dccp, sctp and tipc.  the only one that is smaller than tcp is dccp and it doesn't
provide for retransmission.  (it's intended for streaming content, i believe.)

linux 2.6.17
	proto	ip4 linecount	features
	udp	1594
	tcp	14628

	dccp	8224		congestion controlled, unreliable
	sctp	29139		reliable, mtu-aware, in-order + packet bundling, multipath
	tipc	18175		"transparent" ipc.

oh, for comparison's sake

cpu kernel:
	udp	647
	il	1408
	tcp	3177

- erik

Re: Re: [9fans] linux il/ip

[David Leimbach]

On 9/1/06, erik quanstrom <quanstro@quanstro.net> wrote:
> > I know that's not an argument in favour of il,
> > but there ain't anything else is there?
>
> without doing a survey, linux has added three new ip protocols since the beginning of the
> year.  dccp, sctp and tipc.  the only one that is smaller than tcp is dccp and it doesn't
> provide for retransmission.  (it's intended for streaming content, i believe.)
>
> linux 2.6.17
>         proto   ip4 linecount   features
>         udp     1594
>         tcp     14628
>
>         dccp    8224            congestion controlled, unreliable
>         sctp    29139           reliable, mtu-aware, in-order + packet bundling, multipath
>         tipc    18175           "transparent" ipc.
>
> oh, for comparison's sake
>
> cpu kernel:
>         udp     647
>         il      1408
>         tcp     3177
>
> - erik
>

So what's the performance of il vs tcp like on Plan 9?  Is it because
TCP could be done better?  Also what are the chances of adding TCP to
the FS kernel?

I just don't see that il support in linux is as likely to be
supportable as adding TCP to our more controlled code base.

Dave

Re: Re: [9fans] linux il/ip

[erik quanstrom]

good points.  but i wasn't trying to address that question.

i was trying to point out that there are a few alternative
protocols, but they are (at least according to the linux
implementation) huge and with the possible exception of sctp,
not very useful for plan 9.

- erik

Re: Re: Re: [9fans] linux il/ip

[David Leimbach]

On 9/1/06, erik quanstrom <quanstro@quanstro.net> wrote:
> good points.  but i wasn't trying to address that question.
>
> i was trying to point out that there are a few alternative
> protocols, but they are (at least according to the linux
> implementation) huge and with the possible exception of sctp,
> not very useful for plan 9.
>
> - erik
>

Well in that case I agree :-).

SCTP could be interesting to add to Plan 9.  Would it require DARPA
funding or an NSF grant? :-)

Re: Re: Re: [9fans] linux il/ip

[Christopher Nielsen]

On 9/1/06, David Leimbach <leimy2k@gmail.com> wrote:
> SCTP could be interesting to add to Plan 9.  Would it require DARPA
> funding or an NSF grant? :-)

haven't had internet access in a while, so i am just getting around to
responding to this.

i started an SCTP implementation a while back; it looked like an
interesting protocol. i didn't get too far before i gave up. the
complexity of the protocol gave me the willies.

i am considering going back to what i started, but currently, i think
my limited coding time will be better spent elsewhere. we'll see.

-- 
Christopher Nielsen
"They who can give up essential liberty for temporary
safety, deserve neither liberty nor safety." --Benjamin Franklin

Re: Re: Re: [9fans] linux il/ip

[erik quanstrom]

from what i can tell, SCTP combines the limitations of il with the complexity of tcp.
i think porting il to linux would be more useful.

- erik

On Mon Sep 25 18:18:00 CDT 2006, cnielsen@pobox.com wrote:
> On 9/1/06, David Leimbach <leimy2k@gmail.com> wrote:
> > SCTP could be interesting to add to Plan 9.  Would it require DARPA
> > funding or an NSF grant? :-)
> 
> haven't had internet access in a while, so i am just getting around to
> responding to this.
> 
> i started an SCTP implementation a while back; it looked like an
> interesting protocol. i didn't get too far before i gave up. the
> complexity of the protocol gave me the willies.
> 
> i am considering going back to what i started, but currently, i think
> my limited coding time will be better spent elsewhere. we'll see.

Re: [9fans] linux il/ip

[Martin Neubauer]

* erik quanstrom (quanstro@quanstro.net) wrote:
> from what i can tell, SCTP combines the limitations of il with the complexity of tcp.
> i think porting il to linux would be more useful.
> 
> - erik

Is there then a chance to get il extended ti ipv6?

Martin

Re: [9fans] linux il/ip

[erik quanstrom]

i don't think it would be very hard -- or very useful.
since many routers dislike any data that's not udp/tcp

if you'd like to extend il, i think that il/ethernet
would be a better way to go.  

- erik
On Tue Sep 26 03:20:53 CDT 2006, m.ne@gmx.net wrote:
> * erik quanstrom (quanstro@quanstro.net) wrote:
> > from what i can tell, SCTP combines the limitations of il with the complexity of tcp.
> > i think porting il to linux would be more useful.
> > 
> > - erik
> 
> Is there then a chance to get il extended ti ipv6?
> 
> Martin

Re: [9fans] linux il/ip

[Artem Letko]

On 9/26/06, erik quanstrom <quanstro@quanstro.net> wrote:
> i don't think it would be very hard -- or very useful.
> since many routers dislike any data that's not udp/tcp
>

that's not true - ip routers don't care about layers 4 and up most of the time

> if you'd like to extend il, i think that il/ethernet
> would be a better way to go.
>
> - erik

-art

Re: [9fans] linux il/ip

[Charles Forsyth]

> since many routers dislike any data that's not udp/tcp

>> that's not true - ip routers don't care about layers 4 and up most of the time

it's mainly nat boxes and other protocol transformers that are confused.
to assist that and similar transformations, the ipv6 designers
carefully put port numbers, lengths and checksums in standard places
in all v6 variants of the existing protocols (since the headers were changing anyway).
how's that???  they didn't!!??!?!

Re: [9fans] linux il/ip

[Russ Cox]

IL is long dead.

I used IL over long-distance connections.  It was awful.
TCP is a necessity once you move beyond the local ethernet.
There are no good reasons to try to port IL to other systems
or to try to update it for IPv6.

The only even plausible reason is to connect to old
Plan 9 file servers, but your effort would be better spent
writing some glue so that the old file server code could
run in user space on standard kernels.

If, as was mentioned at the start of the thread, you simply
want to mount your old file server from Linux, by far the
easiest solution is to find a Plan 9 box to proxy between
TCP and IL:

  cat >/bin/service/tcp1234 <<!
  #!/bin/rc
  exec /bin/aux/trampoline il!yourfileserver!9fs
  !
  chmod +x /bin/service/tcp1234

If you don't have an extra machine to serve that capacity,
the next easiest solution is to find one.

Russ

Re: [9fans] linux il/ip

[Brantley Coile]

> IL is long dead.

I don't want to seem cantankerous, but who decided
it was dead?  If you folks think of it as cast off
doesn't mean that someone else has an application
where IL would be a good solution.

 Brantley
 

Re: [9fans] linux il/ip

[jmk]

On Tue Sep 26 13:45:58 EDT 2006, brantley@coraid.com wrote:
> > IL is long dead.
> 
> I don't want to seem cantankerous, but who decided
> it was dead?  If you folks think of it as cast off
> doesn't mean that someone else has an application
> where IL would be a good solution.
> 
>  Brantley
>  

That would be me.

On Tue Aug 29 16:46:15 EDT 2006, jmk@plan9.bell-labs.com wrote:
> il is dead, deal with it.
> ...

I'll repeat what I said in another thread:

On Thu Sep 21 21:26:38 EDT 2006, jmk@plan9.bell-labs.com wrote:
> If you have energy to burn, use it to do something
> interesting. Leave obsessing with the past to the monkeys.

Re: [9fans] linux il/ip

[geoff]

Once we get the last of the data off emelie, we'll be able to compile
kernels without IL and stop using IL.  So for our purposes, IL will be
dead.

People with ken file servers (including me) will undoubtedly keep
using IL for a while, which seems appropriate.

Re: [9fans] linux il/ip

[erik quanstrom]

in my limited experience, the file server is much easier to set up,
makes fewer demands on hardware and exhibits lower lantency
 than venti+fossil running on a cpu kernel.  

but if you're going to run a cpu kernel, i see no point in the file
server code. 

- erik

On Tue Sep 26 12:19:23 CDT 2006, rsc@swtch.com wrote:
> The only even plausible reason is to connect to old
> Plan 9 file servers, but your effort would be better spent
> writing some glue so that the old file server code could
> run in user space on standard kernels.

Re: [9fans] linux il/ip

[Christopher Nielsen]

the problem is firewalls and nat do care, and they are everywhere.

back when i was purchasing large volumes of cisco equipment each year
for an employer, i tried to get cisco to add IL support to IOS. no
dice. apparently, we weren't purchasing enough.

i agree with russ and jmk. IL is long dead. there are plenty of other
projects that would be a better use of your time.

On 9/26/06, Artem Letko <aletko@gmail.com> wrote:
> On 9/26/06, erik quanstrom <quanstro@quanstro.net> wrote:
> > i don't think it would be very hard -- or very useful.
> > since many routers dislike any data that's not udp/tcp
> >
>
> that's not true - ip routers don't care about layers 4 and up most of the time
>
> > if you'd like to extend il, i think that il/ethernet
> > would be a better way to go.
> >
> > - erik
>
> -art
>
>


-- 
Christopher Nielsen
"They who can give up essential liberty for temporary
safety, deserve neither liberty nor safety." --Benjamin Franklin

Re: [9fans] linux il/ip

[erik quanstrom]

in defence of il, and more generally the idea that tcp is not the answer to every
question, i have three points and one question:

1.
i think this is one-size-fits-all thinking.  tcp may be all things to all people, but i can't
imagine that it's always the best solution.  two new protocols were added to linux this
year.  if they thought "cisco doesn't support it, lets give up" they wouldn't have written
27kloc for sctp.  (btw, there are a couple of commercial routers based on linux these
days so sctp is probablly fairly routable.)

2.
who said every project has to be the best use of one's time?  and if that were really
the measure, wouldn't we all be using windows?

3.
and in this case, all protocols don't need to be routed. (unless somebody made a new rule.)
perhaps that's an advantage if you want to keep your venti store to yourself.

okay.  so it's not routable over the internet, which may or may not be moot.
are there any other reasons that il is no good?  on my 1GHz pIV linux box, the
tcp throughput test to one of my plan 9 machines (just to confirm some previously 
observed wierd performance numbers), linux was gasping for breath at 315Mbit with tcp. 
i'd really love to know how linux would fair with other protocols.

- erik

On Tue Sep 26 20:09:00 CDT 2006, cnielsen@pobox.com wrote:
> the problem is firewalls and nat do care, and they are everywhere.
> 
> back when i was purchasing large volumes of cisco equipment each year
> for an employer, i tried to get cisco to add IL support to IOS. no
> dice. apparently, we weren't purchasing enough.
> 
> i agree with russ and jmk. IL is long dead. there are plenty of other
> projects that would be a better use of your time.

Re: Re: Re: [9fans] linux il/ip

[Sergey Zhilkin]

[-- Attachment #1: Type: text/plain, Size: 968 bytes --]

Maybe /sys/src/cmd/unix/9pfreebsd will be good startpoint ?

2006/9/26, erik quanstrom <quanstro@quanstro.net>:
>
> from what i can tell, SCTP combines the limitations of il with the
> complexity of tcp.
> i think porting il to linux would be more useful.
>
> - erik
>
> On Mon Sep 25 18:18:00 CDT 2006, cnielsen@pobox.com wrote:
> > On 9/1/06, David Leimbach <leimy2k@gmail.com> wrote:
> > > SCTP could be interesting to add to Plan 9.  Would it require DARPA
> > > funding or an NSF grant? :-)
> >
> > haven't had internet access in a while, so i am just getting around to
> > responding to this.
> >
> > i started an SCTP implementation a while back; it looked like an
> > interesting protocol. i didn't get too far before i gave up. the
> > complexity of the protocol gave me the willies.
> >
> > i am considering going back to what i started, but currently, i think
> > my limited coding time will be better spent elsewhere. we'll see.
>

[-- Attachment #2: Type: text/html, Size: 1354 bytes --]

Re: [9fans] linux il/ip

[geoff]

I don't think I was around when the decision to drop IL was made.
NAPT devices in particular didn't like it; among other things, its
port numbers are in different places in packet headers than TCP's and
UDP's.  I don't think an IPv6 version of IL was ever created either.

I'm not exactly sure what Brantley's question is.  We've still got
emelie and choline running Ken's file server and serving only 9P over
IL, but I'm moving their data into fossils and ventis as quickly as
possible.  The other file servers (e.g., edith, martha/sources) are
all fossil/venti or vacfs/venti servers serving 9P only over TCP.  So
some of the data seen externally is ultimately fetched via IL, but
that's a small fraction and getting smaller.

Re: [9fans] linux il/ip

[Brantley Coile]

> I'm not exactly sure what Brantley's question is.

Didn't have one.  Was just trying to shine a dim light on things.
(Not that I WANT the light to be dim.)

Re: [9fans] linux il/ip

[lucio]

> (which running the fs kernel and therefore does not do tcp)

Surely you are mistaken?  Geoff?

++L

Re: [9fans] linux il/ip

["Nils O. Selåsdal"]

lucio@proxima.alt.za wrote:
>> (which running the fs kernel and therefore does not do tcp)
> 
> Surely you are mistaken?  Geoff?
No TCP:
http://cm.bell-labs.com/sources/plan9/sys/src/fs/ip/

Re: [9fans] linux il/ip

[Sergey Zhilkin]

[-- Attachment #1: Type: text/plain, Size: 369 bytes --]

case Tcpproto:
		tcprecv(mb, ifc);
		break;


>From /sys/src/fs/ip/ip.c ....

And where it is now ?

2006/8/29, "Nils O. Selåsdal" <NOS@utel.no>:
>
> lucio@proxima.alt.za wrote:
> >> (which running the fs kernel and therefore does not do tcp)
> >
> > Surely you are mistaken?  Geoff?
> No TCP:
> http://cm.bell-labs.com/sources/plan9/sys/src/fs/ip/
>

[-- Attachment #2: Type: text/html, Size: 767 bytes --]

Re: [9fans] linux il/ip

["Nils O. Selåsdal"]

Sergey Zhilkin wrote:
> case Tcpproto:
> 		tcprecv(mb, ifc);
> 		break;
> 
> 
>  >From /sys/src/fs/ip/ip.c ....
> 
> And where it is now ?
It's in icmp.c

void
tcprecv(Msgbuf *mb, Ifc*)
{
	mbfree(mb);
}
Not very useful for doing tcp. It ought to respond with
an icmp packet saying protocol not supported btw :^)

Re: [9fans] linux il/ip

["Nils O. Selåsdal"]

erik quanstrom wrote:
> from what i can tell, SCTP combines the limitations of il with the complexity of tcp.
> i think porting il to linux would be more useful.
> 

It does have congestion and flow control, which is some of the 
limitations of IL.

Re: [9fans] linux il/ip

[erik quanstrom]

also it has extra cruft like mtu discovery and multihoming
of both source and destination.  this means that sctp is ~29kloc
and tcp is 14kloc of code in the linux kernel.  but sctp doesn't
have connections.

il in plan 9 is ~1kloc. so i think my parody is still apt.

- erik

On Tue Sep 26 04:08:52 CDT 2006, NOS@Utel.no wrote:
> erik quanstrom wrote:
> > from what i can tell, SCTP combines the limitations of il with the complexity of tcp.
> > i think porting il to linux would be more useful.
> > 
> 
> It does have congestion and flow control, which is some of the 
> limitations of IL.

Re: [9fans] linux il/ip

[Artem Letko]

On 9/26/06, erik quanstrom <quanstro@quanstro.net> wrote:
> also it has extra cruft like mtu discovery and multihoming
> of both source and destination.  this means that sctp is ~29kloc
> and tcp is 14kloc of code in the linux kernel.  but sctp doesn't
> have connections.
>

not true again: mtu discovery is a standard practice. is traceroute a
cruft? than mtu discovery is too.

> il in plan 9 is ~1kloc. so i think my parody is still apt.
>
> - erik
>

-art