Re: [RFC V2 PATCH] rtnetlink: Add method to calculate dump info data size

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 09 mai 2011 à 20:17 -0700, Stephen Hemminger a écrit :
> On Tue, 10 May 2011 04:43:33 +0200
> Eric Dumazet <eric.dumazet@gmail.com> wrote:
> 
> > Le lundi 09 mai 2011 à 15:26 -0700, Greg Rose a écrit :
> > > The message size allocated for rtnl info dumps was limited to a single
> > > page.  This is not enough for additional interface info available with
> > > devices that support SR-IOV.  Calculate the amount of data required so
> > > the dump can allocate enough data to satisfy the request.
> > > 
> > > V2 of this patch adds a new argument to the rtnl_register service that
> > > allows for a new method to calculate the amount of data required to
> > > complete the info dump request.  So far the method is only implemented
> > > for the RTM_GETLINK slot.
> > > 
> > > Signed-off-by: Greg Rose <gregory.v.rose@intel.com>
> > 
> > >  
> > > +static u16 rtnl_calcit(struct sk_buff *skb)
> > > +{
> > > +	struct net *net = sock_net(skb->sk);
> > > +	int h;
> > > +	int idx = 0, s_idx;
> > > +	struct net_device *dev;
> > > +	struct hlist_head *head;
> > > +	struct hlist_node *node;
> > > +	u16 alloc_size = 0;
> > > +
> > > +	for (h = 0; h < NETDEV_HASHENTRIES; h++, s_idx = 0) {
> > > +		idx = 0;
> > > +		head = &net->dev_index_head[h];
> > > +		hlist_for_each_entry(dev, node, head, index_hlist) {
> > > +			if (idx < s_idx) {
> > > +				idx++;
> > > +				continue;
> > > +			}
> > > +			alloc_size = (u16)if_nlmsg_size(dev);
> > > +			break;
> > > +		}
> > > +	}
> > > +
> > > +	return alloc_size;
> > > +}
> > > +
> > 
> > 
> > Sorry this wont scale. Some machines have thousand of devices.
> > 
> > Just make an upper approximation, you dont need an exact one ;)
> 
> The route dump does scale, can't you use a similar logic?
> The result doesn't come back as one huge allocation.
> I regularly test 600K routes on small machines.
> 

Not sure I understand you Stephen.

In Greg patch, rtnl_calcit() would be called for every 4K/8K block "ip"
gets from kernel.

If you add a function to route dump that would scan the 600K routes to
get the max route size, surely you notice O(N^2) complexity instead of
O(N)

We only need to maintain a global variable to hold min_dump_alloc

[ and only increase this variable when necessary, not bother to decrease
it when removing a driver ]