httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ryan Bloom <...@covalent.net>
Subject Re: MPM design abuse (was: cvs commit: httpd-2.0/server/mpm/prefork prefork.c)
Date Tue, 13 Nov 2001 17:39:30 GMT
On Tuesday 13 November 2001 02:00 am, Greg Stein wrote:
> On Tue, Nov 13, 2001 at 07:15:36AM -0000, rbb@apache.org wrote:
> >
> >   --- core.c	2001/11/13 02:09:07	1.91
> >   +++ core.c	2001/11/13 07:15:36	1.92
> >   @@ -3329,6 +3329,19 @@
> >        return net->c;
> >    }
> >
> >   +static int core_add_listeners(apr_pollfd_t *pollset,
> >   +                              apr_socket_t **listensocks, int
> > num_listensocks) +{
> >   +    int i;
> >   +    ap_listen_rec *lr;
> >   +
> >   +    for (lr = ap_listeners, i = 0; i < num_listensocks; lr = lr->next,
> > i++) { +        listensocks[i] = lr->sd;
> >   +        apr_poll_socket_add(pollset, listensocks[i], APR_POLLIN);
> >   +    }
> >   +    return OK;
> >   +}
>
> I think this is *very* wrong, and I'm even tempted to -1 the darn thing. I
> think we need a bit more discussion before continuing here.
>
> The problem is that now an MPM must implement ap_listeners for use by
> core.c. The MPM must also use a pollset.

I am going to be removing the ap_listeners from the MPM, so that the core will be
the only thing that touches it.  Yes, MPMs will need to use a pollset.  This allows
any module to add information to the pollset, which means that we can have a 
single server that responds to multiple transport layers.  All of the transport
layers must exist in the same pollset, or we will starve some of them.

> That is NOT right. The MPM is the thing that defines how sockets are to be
> set up, listened to, and accepted. It then defines the mapping from those
> sockets to backend workers.
>
> Think about the above code: the MPM sets up ap_listeners, and then it calls
> an external function to process them, to put them into a pollset that it
> created. It's like a damned little GOSUB to get some of its own work done.
> The core is not adding any value here. It grabs some data from the MPM (the
> ap_listeners variable) and puts it right back into an MPM structure (the
> passed pollset variable).

True, today this is a minor change.  But when ap_listeners is back in the core,
where it belongs, the MPM becomes a much simpler beast.  The MPM has a 
single purpose, to map threads to 

> And the interface is just whacked anyhow. The ap_listeners is a global, but
> the size of it is a passed parameter. If you want to do this right, then
> ap_listeners should not be a global (we've got a serious over-reliance on
> globals!), but should be a passed parameter. Yet if that is done, then
> again: the hook is not adding any value. The MPM passes everything it
> needs; the hook just runs a loop for the MPM. Kinda silly...

Yep, give me time to keep cleaning this up.

> And with this structure, how is an MPM supposed to set up different sets of
> listeners for different threads? Oh... sorry, but the MPM can't define that
> any more. I think with a bunch of monkey business, the MPM might be able to
> define different sets for child processes. But the MPM is going to be
> working *around* the issue when it should be *in charge*.

The same way they do today, with a small twist.  Today, they loop through
the ap_listeners list, tomorrow, they loop through a pollset.

.......  Changed my mind........

I am going to back out this change.  Instead of doing this, I am going to allow
modules to add sockets to the ap_listeners array.  The MPM will own the
ap_listeners list.


> I like the code changes for reducing the socket usage from Apache proper.
> But removing it from the MPM isn't right.
>
> Please explain :-(

Don't think of it as removing the socket usage from the MPM.  The point of this
is to augment what the MPM is doing.  Here's my point, the core already does
all of the direct network logic.  The MPM maps the network to a thread, it 
doesn't do any real network logic, nor should it.  Currently, our MPMs are
too close to the network.  Also, don't think of MPMs as separate from the core.
They are a pluggable part of the core.

Ryan

______________________________________________________________
Ryan Bloom				rbb@apache.org
Covalent Technologies			rbb@covalent.net
--------------------------------------------------------------

Mime
View raw message