Re: processes state "sokva"

[David Young <dyoung%pobox.com@localhost>]

On Fri, Nov 14, 2008 at 09:19:52PM -0600, Al - Image Hosting Services wrote:
> Hi,
> 
> I have been running several servers with lighttpd on them and it hasn't 
> been a problem until lately. The process enters into a state of "sokva". I 
> think this means that it is running out of memory for socket buffers. I 
> tried building a kernel with: "option SOSEND_NO_LOAN", and although now it 
> didn't give me a state of "sokva", some processes stopped responding. I 
> built another kernel with:
> 
> options         MAXFILES=16384
> options         NMBCLUSTERS=20480
> options         DFLDSIZ=402653184
> options         SHMMAXPGS=8192
> options         SYSVMSG 
> options         SYSVSEM 
> options         SYSVSHM 
> options         MSGMNB=16384
> options         MSGSSZ=64
> options         MSGTQL=512
> 
> But I commented out option SOSEND_NO_LOAN.
> 
> I switched to apache, and still I was getting some processes that have a 
> state of "sokva". I then "sysctl -w kern.somaxkva=33554432", it has not 
> done it since then, but I haven't been pushing it as hard as I was.

Al,

I believe that I have seen Apache stuck in a classic livelock with
the kernel: it has sockets ready to read, the kernel socket buffers
are exhausted, and Apache tries and tries to write to a socket instead
of reading.  (Reading would free some socket buffers.)  At one time, the
Apache code (apr, actually) looked like it could get into this livelock
situation.  lighttpd may have a similar livelock mode.  You may want to
look into this after you have chased down the other leads.

Dave

-- 
David Young             OJC Technologies
dyoung%ojctech.com@localhost      Urbana, IL * (217) 278-3933 ext 24