Messages in this thread | | | Date | Thu, 19 Mar 2009 22:52:52 +0000 | From | Mel Gorman <> | Subject | Re: [PATCH 20/35] Use a pre-calculated value for num_online_nodes() |
| |
On Thu, Mar 19, 2009 at 06:42:39PM -0400, Christoph Lameter wrote: > On Thu, 19 Mar 2009, Mel Gorman wrote: > > > I posted an amalgamation. Sorry for the cross-over mails but I wanted to > > get tests going before I fell asleep. They take a few hours to complete. > > > > > > static inline void node_set_state(int node, enum node_states state) > > > > { > > > > __node_set(node, &node_states[state]); > > > > + if (state == N_ONLINE) > > > > + nr_online_nodes = num_node_state(N_ONLINE); > > > > } > > > > > > That assumes uses of node_set_state N_ONLINE. Are there such users or are > > > all using node_set_online()? > > > > > > > node_set_online() calls node_set_state(node, N_ONLINE) so it should have > > worked out. > > But this adds a surprising side effect to all uses of node_set_state. > Node_set_state is generating more code now. >
Fair point.
> > > if you want to check if the system could ever bring up a second node > > > (which would make the current optimization not viable) whereas > > > nr_online_nodes is the check for how many nodes are currently online. > > > > > > > I redid your patch to drop the nr_possible_nodes() because I couldn't convince > > myself it was correct in all cases and it isn't as important as avoiding > > num_online_nodes() in fast paths. > > I was more thinking about getting the infrastructure right so that we can > avoid future hacks like the one in slab. >
Which is fair enough and you're right in that it's worth fixing. One horribly large patchset and associcate thread at a time though so I'll be putting it on the wrong finger rather than adding this to the pile right now :).
-- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab
| |