Re: Session Affinity after Graceful Apache restart? [was Re: Adding working dynamically with mod_jk status?]

["William A. Rowe, Jr." <wr...@rowe-clan.net>]

At 02:08 PM 6/7/2005, Mladen Turk wrote:

>It works of course, but IIRC you are planning to use it for
>reconfiguring mod_jk. I think you might get into the problems
>with shared memory (particularly on unix) because some child
>might have a different idea about shared memory addresses.

Oh - so this is a bug/design flaw in mod_jk?  Does the same
affect the prefork/worker MPMs?

One aspect is inheritance of handles (e.g. handle to SHM
segment).  I've always ment to make that more generic, and it
would be nice to solve for Apache 2.2.  It's used now to hand
of the score and listener handles from the parent to child.

The trivial solution is to use an anon segment which Win32 will
not inherit, which means that SHM is unique to a single generation,
and won't be polluted by changes in another worker.

This does sound like a good use case, though, for handing off
memory segments, mutexes and so forth to child processes.  With
a pointer or two, I'll be happy to look at it.

Bill  


---------------------------------------------------------------------
To unsubscribe, e-mail: tomcat-dev-unsubscribe@jakarta.apache.org
For additional commands, e-mail: tomcat-dev-help@jakarta.apache.org

Re: Session Affinity after Graceful Apache restart? [was Re: Adding working dynamically with mod_jk status?]

[Mladen Turk <mt...@apache.org>]

William A. Rowe, Jr. wrote:
> At 02:08 PM 6/7/2005, Mladen Turk wrote:
> 
> 
>>It works of course, but IIRC you are planning to use it for
>>reconfiguring mod_jk. I think you might get into the problems
>>with shared memory (particularly on unix) because some child
>>might have a different idea about shared memory addresses.
> 
> 
> Oh - so this is a bug/design flaw in mod_jk?  Does the same
> affect the prefork/worker MPMs?
> 

Actually the WIN32 has no such problems because it uses the
plain memory instead of shared.
Problem with WIN32 is because the closing child and new child
will not share the worker data during restart, so you will loose
the statistics during restart.

With unixes the problem is because the workers are allocated
according to the worker.list and balance_members.
If you change those lists in the config on restart the child
processes could have different shared memory segment addresses.

Like said if you only add to the end of list then everything
will probably be OK.

To resolve that we would actually need a database like shared memory
with worker name as a key.

Regards,
Mladen.

---------------------------------------------------------------------
To unsubscribe, e-mail: tomcat-dev-unsubscribe@jakarta.apache.org
For additional commands, e-mail: tomcat-dev-help@jakarta.apache.org