Re: [PATCH] inet_diag: make config INET_DIAG bool

On Mon, 2012-09-24 at 16:48 +0800, Gao feng wrote:

于 2012年09月23日 21:40, Jan Engelhardt 写道:

quoted

On Sunday 2012-09-23 06:36, Stephen Hemminger wrote:

quoted

quoted

when inet_diag being compiled as module, inet_diag_handler_dump
set netlink_dump_control.dump to inet_diag_dump,so if module
inet_diag is unloaded,netlink will still try to call this function
in netlink_dump. this will cause kernel panic.

Another way to handle this to just get rid of inet_diag_exit
so that module can be loaded but not unloaded.

Why don't we unset netlink_dump_control.dump on exit?

Though I like this idea,but it may cause dead lock.
netlink_dimp [mutex_lock(netlink_sock->cb_mutex) here]
   |->inet_diag_dump
	|->__inet_diag_dump
	    |->inet_diag_lock_handler [may try to load tcp_diag here,
					so we need module_mutex lock]

And on module exit,we already get module_mutex lock,
if we unset netlink_sock->cb,we need to get mutex lock of netlink_sock->cb_mutex.

I think this will cause dead lock.

I don't know if I should change this patch as Stephen said,because I don't know
witch one is better.

Any comments?

In fact I didnt fully understand the problem you try to address.

If you want to prevent module being unloaded, you need to add proper
module_get()/module_put()

So I would add a "struct module *module;" in struct sock_diag_handler
and use it appropriately.

But then, I might have totally misunderstood the problem.

Care to explain how you trigger the bug ?