[Bug 209471] Listen queue overflow due to too many sockets stuck in CLOSED state

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=209471

--- Comment #6 from Robert Blayzor <rblayzor at inoc.net> ---
I do not have an environment that would allow me really to test it. This
problem does seem a lot more apparently after upgrading to 10.3 however. It's
either that, or the work around for BugID 204426 unmasked this one from
happening.

Bug 204426 we would see our processes normally die, so they never really ran
for longer period of time before we had to restart them.

We added the patch for PR 204426, and processes seem stable now, but now we
have this bug. I believe we used to see this in in 10.2 as well, but not nearly
as often now that we seem to have 204426 fixed.

The application doesn't seem to matter. Our environment has not changed. The
only real special setup we have is also described in 204426. ESXi hypervisor,
VMX NIC driver, and NFS mounted root FS. Other than the sysctl knobs previously
mentioned, nothing else special other than a non-GENERIC kernel; but all we did
was remove modules and drivers we do not use so kernel build time is faster.

One thing I did not check is to see if this is a V6 socket only issue or if V4
is also affected. I can still SSH into the server when this happens, only the
process with the full queue and lots of sockets stuck in CLOSED  seems to hung.
9 times out of 10 only a server reboot clears the issue. Rarely we can manage
to kill -9 the process and restart  it.

-- 
You are receiving this mail because:
You are on the CC list for the bug.