httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Roy T. Fielding" <field...@kiwi.ICS.UCI.EDU>
Subject Re: graceful restarts, take 5
Date Tue, 22 Apr 1997 19:53:04 GMT
>I will only give this stuff a +0 for 1.2.  But there's so many +1s now I
>guess it don't matter. 

The problem is that it forced me to actually test the graceful restart
code in the current cvs, and I found that it completely screws up a
server (at least on Solaris) -- you have to kill -9 each child afterwords.
I hadn't realized it was *that* bad.  Given the choice is between me going
through and deleting the existing code, or just applying a patch which
appears to work great, I'd rather take a chance on the optimistic change
rather than a pessimistic one.

>For take 6 I am going to include Roy's second suggested change to
>new_block().  But I'm going to skip the ap_foo() cleanups suggested by Jim
>and submit them post 1.2. 

You mean moving them to conf.h, right?  I do like that change inside
http_main.c (and I've already done the work of testing it :).  What
we really need is a multiple platform test like the ones Marc does.

One thing I'd like to establish is a policy of not changing the API
magic number on 1.2.x releases; if it requires 3rd-party module writers
to recompile and redistribute their code, then it is more than a minor

>I'll also bang on it on a busy solaris server and busy irix server.  But
>note that I've never used SIGHUP on these so I probably won't be testing
>that... and oh yeah, a caveat that I didn't mention: it probably breaks
>inetd mode.  (hint hint)

I tested SIGHUP on a non-busy Solaris server.  It fixed the memory leak
that was causing the parent to bulge, which is significant for sites
that still do a nightly SIGHUP (since all the children inherit the
parent's size as well).  Does anyone have a busy-but-not-critical
server that they can run a looping kill test on?  That would be for
testing the code for race conditions.


View raw message