Re: [PATCH/RFC] New module refcounting for net_proto_family

[Max Krasnyansky]

At 04:30 PM 2/20/2003, Rusty Russell wrote:
>In message <[e-mail]> you write:
>> >There has been talk of this, but OTOH, the admin has explicitly gone
>> >out of their way to remove this module.  They really don't want anyone
>> >new using it.  Presumably at this very moment they are killing off all
>> >the processes they can find with such a socket.
>> The thing is that once those processes are killed sockets will be 
>> destroyed and release the module anyway. i.e. There is no reason to
>> sort of artificially force accept() to fail. Everything will be cleaned 
>> up once the process is gone.
>
>Yes, but in practical terms it's probably going to fork a child with
>that socket.
But it will also be killed.

>> >I think it can be argued both ways, honestly.
>> Yep. And I'd argue in for of module_get() :)
>
>My only real insistence in this is that such an interface be called
>__try_module_get(), because the "__" warn people that it's a "you'd
>better know *exactly* what you are doing", even though the "try" is a
>bit of a misnomer.
Yeah, I think 'try' is definitely be a misnomer in this case.
How about something like this ?

static inline void __module_get(struct module *mod)
{
#ifdef CONFIG_MODULE_DETECT_API_VIOLATION
        if (!module_refcount(mod))
                __unsafe(mod);
#endif
        local_inc(&mod->ref[get_cpu()].count);
        put_cpu();
}

We will be able to compile the kernel with CONFIG_MODULE_DETECT_API_VIOLATION
and easily find all modules that call __module_get() without holding a reference.

Comments ?

Max