Hi there,

I'm having an issue with a process dying (being killed by OpenVZ
limits, presumably), and I can't figure out exactly why it's getting
killed.

Background info:
- kernel 2.6.18-53.1.19.el5.028stab053.14
- CentOS5 host, 2 CentOS5 guests
- The host has 2GB memory, 0.5GB swap
- One guest is only running BIND (plus the usual, sshd, syslogd,
  sendmail, crond services).  Am having no issues with this guest.
  privvmpages is set to provide it up to 256MB memory, and it's using
  about half that.
- The other guest is running postgresql, java, apache, and freeradius.

The problem is that freeradius keeps dying.  Whenever it dies, failcnt
on privvmpages goes up and indeed the maxheld privvmpages value is
above the limit value, so I guess the issue is that OpenVZ thinks that
something is taking too much memory and is killing radiusd (no other
failcnt numbers go up -- only privvmpages).

There doesn't appear to be anything logged in the dmesg output on the
host or the guest to indicate that anything was killed due to a limit
being exceeded (should there be?).

A few strange things:

- Although the maxheld privvmpages value is above the limit, I've
  never seen the held privvmpages value get anywhere near the limit,
  even checking the value only seconds before radiusd gets killed, the
  held privvmpages value is under half the limit, eg. just before
  radiusd is killed:

  2008-07-11 06:39:24:
       uid  resource           held    maxheld    barrier      limit  failcnt
            privvmpages      224497     581366     506368     557056      486

  Then 10 seconds later (radiusd was killed and possibly restarted
  sometime in this interval):

  2008-07-11 06:39:34:
       uid  resource           held    maxheld    barrier      limit  failcnt
            privvmpages      182445     581366     506368     557056      487

  (is there any way to reset the maxheld values without restarting the
  guest?)

- Similarly, the output of free doesn't indicate anything wrong:

  2008-07-11 06:39:24:
               total       used       free     shared    buffers     cached
  Mem:       2071924     898092    1173832          0          0          0
  -/+ buffers/cache:     898092    1173832

  2008-07-11 06:39:34:
               total       used       free     shared    buffers     cached
  Mem:       2071924     729884    1342040          0          0          0
  -/+ buffers/cache:     729884    1342040

- I've found that I can reproduce the issue on demand by sending many
  RADIUS requests to radiusd at once, but watching what radiusd does
  with ltrace -f doesn't show anything out of the ordinary.  I summed
  up all the malloc() requests and saw only 22MB requested.

- Finally, using strace -f to see what radiusd was doing -- there were
  only about 22MB worth of calls to brk() (matching malloc(), as you'd
  expect).  And summing the mmap() length parameters (not counting
  munmap() calls) I only came up with 300MB, well within the free
  memory.

Any ideas on debugging this?

Thanks in advance,
-- 
Geoffrey D. Bennett, RHCE, RHCX               mailto:g@netcraft.com.au
Senior Systems Engineer                          sip:g@netcraft.com.au
NetCraft Australia Pty Ltd        http://www.netcraft.com.au/geoffrey/