httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chuck Murcko <ch...@telebase.com>
Subject Re: [Fwd: Re: socket error on APACHE v1.0]
Date Mon, 12 Feb 1996 05:32:14 GMT
Paul Richards liltingly intones:
> 
> In reply to Chuck Murcko who said
> > 
> > Systems where the socket library is really wriiten on top of streams
> > transport layer (SVR4). Eventually you hit a 1-deep connection accept
> > queue. That's what puzzles me about the problems with UnixWare 2.x.
> > accept() fails after HUP. Sounds like 'sockfds' never closing, or
> > ending up in some dumb state like TIME_WAIT.
> 
> 
> I don't follow, what is the locking supposed to achieve? I don't see what
> affect it would have on the queue being 1-deep.

Sorry, it doesn't. I was daydreaming of a different queue in streams.

I don't have my SVR4 stuff here, but on at least Solaris, IRIX, and Unixware,
you get accept() failures by the boatload if the accept() isn't locked.
Somewhere in streams (I'll dig it out again) is a piece of code that
can only deal with one accept() at a time. My books are at work, so I'll
write you tomorrow (er, later today).

> This should be ifdef'd for broken OS's since it's totally unecessary on
> working systems and slows things down.
> 
The accept_mutex_* functions are only defined for the broken OSs now.
They're empty for all the ones that work. The conf.h flag is
FCNTL_SERIALIZED_ACCEPT.

chuck
Chuck Murcko	Telebase Systems, Inc.	Wayne PA	chuck@telebase.com
And now, on a lighter note:
Boys will be boys, and so will a lot of middle-aged men.
		-- Kin Hubbard

Mime
View raw message