[9fans] boot via aan from fileserver?

[9fans] boot via aan from fileserver?

[Axel Belinfante]

I'm currently booting diskless machines at home
(using boot flop with plan9.ini and kernel(s))
over cable modem from fs in the office
(via trampoline tcp <-> il on the cpu server)
Unfortunately, the connection occasionally breaks
giving the i/o on hangup channel error.
Could something like aan be used to make it more reliable?
I know the (non-fossil fake worm) fs itself does not speak aan,
but the trampoline is there already anyway.

I did reread nemo's message of 7 May 2003 ("update4d redirfs on sources")
in which he describes the use of redirfs to fall back from fs to kfs.
I assume I could try to use something like that,
but I would prefer to keep the machines diskless
to reduce the number of file servers to keep up to date.

Any ideas?
Axel.

Re: [9fans] boot via aan from fileserver?

[David Presotto]

[-- Attachment #1: Type: text/plain, Size: 286 bytes --]

When Vic Zandy gets to the lab in a month, we're going to stick his
Rocks stuff (an aan like thing) under all our server connections.
That way we'll be able to ride out network failures and perhaps
even server failures.  The latter will take a lot more work so
don't expect it soon.

[-- Attachment #2: Type: message/rfc822, Size: 3729 bytes --]

From: Axel Belinfante <Axel.Belinfante@cs.utwente.nl>
To: 9fans@cse.psu.edu
Subject: [9fans] boot via aan from fileserver?
Date: Sun, 25 Jan 2004 14:49:44 +0100
Message-ID: <200401251349.i0PDni804749@zamenhof.cs.utwente.nl>

I'm currently booting diskless machines at home
(using boot flop with plan9.ini and kernel(s))
over cable modem from fs in the office
(via trampoline tcp <-> il on the cpu server)
Unfortunately, the connection occasionally breaks
giving the i/o on hangup channel error.
Could something like aan be used to make it more reliable?
I know the (non-fossil fake worm) fs itself does not speak aan,
but the trampoline is there already anyway.

I did reread nemo's message of 7 May 2003 ("update4d redirfs on sources")
in which he describes the use of redirfs to fall back from fs to kfs.
I assume I could try to use something like that,
but I would prefer to keep the machines diskless
to reduce the number of file servers to keep up to date.

Any ideas?
Axel.

Re: [9fans] boot via aan from fileserver?

[Fco.J.Ballesteros]

[-- Attachment #1: Type: text/plain, Size: 221 bytes --]

In the mean time you can perhaps reexport your
worm fs form a cpu server, and use aan on that.
But I'd say you better switch to fossil. You could
have used aan directly then.

If I can help you anyhow, let me know.

[-- Attachment #2: Type: message/rfc822, Size: 5984 bytes --]

[-- Attachment #2.1.1: Type: text/plain, Size: 286 bytes --]

When Vic Zandy gets to the lab in a month, we're going to stick his
Rocks stuff (an aan like thing) under all our server connections.
That way we'll be able to ride out network failures and perhaps
even server failures.  The latter will take a lot more work so
don't expect it soon.

[-- Attachment #2.1.2: Type: message/rfc822, Size: 3729 bytes --]

From: Axel Belinfante <Axel.Belinfante@cs.utwente.nl>
To: 9fans@cse.psu.edu
Subject: [9fans] boot via aan from fileserver?
Date: Sun, 25 Jan 2004 14:49:44 +0100
Message-ID: <200401251349.i0PDni804749@zamenhof.cs.utwente.nl>

I'm currently booting diskless machines at home
(using boot flop with plan9.ini and kernel(s))
over cable modem from fs in the office
(via trampoline tcp <-> il on the cpu server)
Unfortunately, the connection occasionally breaks
giving the i/o on hangup channel error.
Could something like aan be used to make it more reliable?
I know the (non-fossil fake worm) fs itself does not speak aan,
but the trampoline is there already anyway.

I did reread nemo's message of 7 May 2003 ("update4d redirfs on sources")
in which he describes the use of redirfs to fall back from fs to kfs.
I assume I could try to use something like that,
but I would prefer to keep the machines diskless
to reduce the number of file servers to keep up to date.

Any ideas?
Axel.

Re: [9fans] boot via aan from fileserver?

[Axel Belinfante]

nemo replied to my question about using aan to boot from fileserver:
> In the mean time you can perhaps reexport your
> worm fs form a cpu server, and use aan on that.
> But I'd say you better switch to fossil. You could
> have used aan directly then.

I'm now trying the reexport approach.
The simplest approach seemed to use import to get /root from
the cpu server, and let import use aan.
However, import -p uses /bin/aan, whereas in the boot trick
I will have /boot/aan.  As work around I added a -P aanprog flag
to import, and hacked the boot code to have a new bootmethod 'tcpaan'
that run's this import to get /root from the cpu server,
in a way similar to fossil is started.

I at least succeeded to boot, and indeed aan is 'in' the connection,
as far as I can see.
I'll just let it run for some days and see what happens.\

Just curious: what network outage would I 'need' to benefit
from aan improved reliability?
(So far, I got the root fs via tcp over cable modem to cpu
 which run trampoline to get to worm fs via il.)

Out of curiousity:
would there have been a simpler way than the on I described above?

And just one remark: once I had been browsing the boot code
enough to be a bit comfortable with its structure,
adding the new boot method turned out to be surprisingly simple
(the fact that all pieces I needed similar code already appeared
 here and there (tcp, fossil boot) of course helped immensely)

Axel.

Re: [9fans] boot via aan from fileserver?

[Fco.J.Ballesteros]

> Out of curiousity:
> would there have been a simpler way than the on I described above?

I think I once did something similar but did not have to modify boot code.
Sadly, I don't remember exactly what I did. Probably I did not used aan
in the cpu server and thus didn't have to change things. It was when
we had (well, we still have) a flaky network and had to import the venerable
fs(8) into a cpu server to reexport it through tcp instead of il.

Re: [9fans] boot via aan from fileserver?

[Axel Belinfante]

> > Out of curiousity:
> > would there have been a simpler way than the on I described above?
>
> I think I once did something similar but did not have to modify boot code.
> Sadly, I don't remember exactly what I did. Probably I did not used aan
> in the cpu server and thus didn't have to change things. It was when
> we had (well, we still have) a flaky network and had to import the venerable
> fs(8) into a cpu server to reexport it through tcp instead of il.

hmm. interesting.

In the mean time, after being away for a day, I find at home that
my connection has been down long enough such that a machine
booted 'normally' from the work fs gives the 'i/o on hungup channel'
whereas the 'tcpaan' booted one does not complain but simply
hangs and waits -- seems what I'm doing does not (yet?) give
me what I'm looking for...

The aan log on the cpu server only shows a last 'waiting for
connection message.
The result of ^t^tp on the client scrolls too fast to be able
to read the start, and ends up partially garbled on another
machine via serial cable
(fix this by playing with meessages to eia0 ctl?
 or change baud rate in plan9.ini console entry?)

Got similar result after rebooting the machine and
(once up and having run for a while) unplugging its network cable
for a while and then plugging it back in.
Somehow it looks like the client doesn't notice the connection
is broken and something has to be done?

Too late to experiment more now. I'll try to find some time later
to play a bit with debugging on the client aan enabled.
(and a running snoopy)

Axel.

Re: [9fans] boot via aan from fileserver?

[Axel Belinfante]

Just for those interested:
> In the mean time, after being away for a day, I find at home that
> my connection has been down long enough such that a machine
> booted 'normally' from the work fs gives the 'i/o on hungup channel'
> whereas the 'tcpaan' booted one does not complain but simply
> hangs and waits -- seems what I'm doing does not (yet?) give
> me what I'm looking for...

After thinking about the code of aan.c I supected that the syslog(0, ...)
calls in the if(dialstring) branch of reconnect() were
causing the problems:
network problem -> pending connection to fs -> trying to consult
fs to see if we can write to a log file not possible ->
deadlock while waiting for the result of the log file existence test?!?

Experiments seem to confirm this -- replacing those calls to
syslog(0, Logname, ...) by calls to dmessage(0, ...)
seems to have solved the problem.

Now my only remaining problem is to make a number of connections to u9fs
reliable as well. Since there is no unix aan, easiest trick seems
to mount them all on the cpu server (where I mount them anyway)
and reexport from there.
If there were something like aan for unix, I could maybe even mount
those u9fs things on the cpu server via /lib/namespace.$sysname
(and then on the terminal at home in the same way)

Axel.