At 04:30 PM 4/23/2003, David S. Miller wrote:
> From: Max Krasnyansky <maxk@xxxxxxxxxxxx>
> Date: Wed, 23 Apr 2003 15:51:11 -0700
> >This is just the first part, DaveM already merged the second part,
> >that deals with struct sock
> That's exactly what surprised me. He rejected complete patch and
> accepted something incomplete and broken.
>No, it was not broken, because he told me completely where he
>was going with his changes.
Of course it was and still is. New socket is allocated without incrementing
modules ref count in sys_accept().
>He was building infrastructure piece by piece, and that's always an acceptable
>way to do things as long as it is explained where one is going with the
Oh, I see. And I just sent a patch without any explanation. Ok.
(you might want to reread our original discussion again).
>Your stuff was unacceptable from the start because you didn't put
>the ->owner into the protocol ops.
But you didn't tell me that. You just said that it's "an ugly hack" without
giving any other feedback.
->owner field in protocol ops did come up during discussion (I think Rusty
that up) and I explained why it shouldn't be there. But again there was no feed
from you. You just ignored that thread at some point.
btw I still don't see ->owner in protocol ops. I read archives of netdev. You
didn't even talk about that.
Anyway it's not important who said what now. You chose to ignore stuff that I
What about this though
>>struct sock *sk_alloc(int family, int priority, int zero_it, kmem_cache_t
>>- struct sock *sk;
>>+ struct sock *sk = NULL;
>>+ if (!net_family_get(family))
>>+ goto out;
>Ok. This is wrong. Which should be clear from reading the thread that I
>Owner of the net_proto_family is not necessarily the owner of the 'struct
>Example: af_inet module registers net_proto_family but udp module owns the
>(not that IPv4 code is modular but just an example). Another example would be
>We have Bluetooth core that registers Bluetooth proto_family and several
>that handle specific protocols (l2cap, rfcomm, etc).
>Also net_proto_family has pretty much nothing to do with the struct sock. The
>we would want to hold reference to the module is if the module has replaced
>callbacks (i.e. sk->data_ready(), etc).
>So my point is we need sk->owner field.
>I'd also prefer to see sock->owner which gives more flexibility (i.e.
>be unregistered while struct socket still exist, etc).
>net_family_get/put() makes no sense net_proto_family has only one function
>which is referenced only from net/socket.c. struct socket should inherit owner
>struct net_proto_family by default but protocol should be able to assign
>ownership to a
>different module if it needs to.