inetd fd leak?

[Denys Vlasenko]

On Thu, Sep 3, 2009 at 8:24 PM, Cathey, Jim<jcathey at ciena.com> wrote:
>>Please try 1.15.0
>
> After back-porting the changes to the 1.13.2 environment
> so that it will compile, it still exhibits the slow leakage
> of socket fd's.  (The changes to 1.15.0 are mostly #definery
> to compile, and putting back the non-parser_t version of the
> config file logic.  I don't think this is significant to the
> problem at hand.)  It's been running less than an hour, and
> there are already six sockets open, up from one in the beginning.
>
> (Our target system is expected to run for years at a time with
> no problems and no reboots, and it will soon be in a noticeable
> part of the cell phone infrastructure of the USA, in thousands
> of field sites that nobody wants to have to drive to to
> power-cycle because they can't telnet in anymore!  We're
> motivated to fix this.)
>
> Stimulus behavior is a mix of the telnet firehose I gave,
> along with disabling and enabling the telnet service a
> few times via commenting in and out the (lone) service
> line and HUPping inetd.

Let's try to simplify it. Try some (or all) of these:

* Run inetd with the same config on your own x?86 desktop.
  Try to make it leak fds. Does it still happen?
  (iow: does architecture, CPU speed, or network latency matter?)

* Does it happen if you start inetd, *don't* SIGHUP it,
  and repeatedly run ./bury against it?
  (is SIGHUPing a factor in reproducing it?)

* Does it happen if you replace
  telnet stream tcp nowait.12 root /usr/sbin/telnetd telnetd -i
  with
  telnet stream tcp nowait.12 root /bin/true true
  (iow, if server prog just exits at once)?

--
vda