Re: xfrm_state locking regression...

["Timo Teräs" <timo.teras@iki.fi>]

Herbert Xu wrote:
> On Tue, Sep 23, 2008 at 04:01:29PM +0300, Timo Teräs wrote:
>> So, what to do?
>> 1. Go back to: list_del_rcu, xfrm_state_hold(all.next) on delete and
>>    xfrm_state_put(all.next) on destruct.
>> 2. Add per-entry hlist of walkers currently referencing it.
>> 3. Use the global walker list.
>>
>> 1 can keep memory allocated until userland wakes up. 2 & 3 can make
>> the delete of that entry slow if there's many walkers suspended.
> 
> I'd cross 3 off the list because 2 is just so much better :)
> 
> I'd slightly lean towards 2 but now that you mention it yes even
> that is vulnerable to loads of dumpers sitting on the same entry.
> So SELINUX folks wouldn't like that :)

Umm... right. It's a tricky problem. Cannot think of perfect
solution atm. But I guess 3 is in general case the best. But in
worst case scenarios 1 performs better.

I have no strong opinion either way. So what ever you want, I'm
happy to provide a patch for.

>> Btw. the current stuff in net-next is broken. There's no locking
>> for xfrm_state_walkers list handling.
> 
> What about xfrm_cfg_mutex?

It's used only in xfrm_state_gc_task. xfrm_state_walk_{init,done}
touch xfrm_state_walks list without locking properly. At least in
the version I'm looking (= net-next-2.6 via git web interface).

- Timo