Re: [9fans] killing rogue processes

["Bruce Ellis" <brucee@chunder.com>]

----- Original Message -----
From: "Russ Cox" <rsc@swtch.com>
To: <9fans@cse.psu.edu>
Sent: Friday, December 19, 2003 3:57 AM
Subject: Re: [9fans] killing rogue processes


...

> Kill the server it is talking to, which is clearly misbehaving

no, you have to try harder...  sometimes a lot harder.
it is not a problem with misbehaving servers.  it's a flaw
in 9p(2000), or devmnt, take your pick.

when a process exits it closes its fds.
if the close is the final reference to a channel then
the chan is clunked and if the chan is being served
the poor old process sends off a Tclunk and sits
around waiting for the Rclunk.  if the server is also
exiting (which happens when they are in the same
killed pgrp) they might be lucky enough to *not* be
stuck waiting for each other.

a very common scenario is a deadly embrace between
rio upas/fs and exportfs.  it happens all the time to me.
rob's suggestion was to "rattle the cage enough" to get
things to go away.

things were even worse in inferno where fds are never
closed explicitly.  i fixed it in inferno because any
system of any complexity demanded a huge stick to
rattle the cage.

as for "who is using inferno", well i use it for almost
everything - almost all code i've written this year has
been in limbo.  i run emu on plan9 for development,
emu on win2k to give me access to all the devices
that i can't get data sheets for (particularly usb midi
and non-standard usb audio).  of course it also runs
in my phone and router (wrt54g) and on $50 386
machines from china, various arm/xscale boards etc.
my inferno web server (with active scripting) is so
much smaller and faster than apache that it runs
without pain on all of the above platforms.

i've had informal sporadic discussions with VN about
how to share the wealth (the inferno i run is so divergent
from theirs that i wouldn't attempt a merge).  maybe
something will happen - when the weather gets worse.

just my threepence worth.

brucee