Re: xfrm_state locking regression...

["Timo Teräs" <timo.teras@iki.fi>]

Herbert Xu wrote:
> On Tue, Sep 23, 2008 at 03:08:08PM +0300, Timo Teräs wrote:
>> +		list_for_each_entry(walk, &xfrm_state_walks, list) {
>> +			if (walk->state == x)
>> +				walk->state = next;
>> +		}
> 
> Just make a new list in xfrm_state that has all the dumpers sitting
> on it.  In fact all you need is a hlist, or even just a pointer.

There can be a lot of xfrm_state structs. As in thousands. So it
will take more memory then. hlist would be enough, so it'd be a
4/8 bytes addition. Single pointer would not be enough as we can
have multiple walks on the same entry.

> Walking through an unbounded list on the fast path is not good :)

I was thinking about the same. That's why I wasn't so sure if this is
better. In practice there aren't many walks active. Maybe we limit
the maximum simultaneous walks?

But in real life, there is only a one or two walkers if any active.
So I would not consider a global update too expensive. But if you think
it might become a problem, I'd add an hlist to xfrm_state of walkers
referencing that entry.

- Timo