>As an interim & admittedly "hackey" solution, we have introduced the >notion of "cnodes". For cpu & memory nodes, the cnode number is >the same as the nid (generic kernel node number). IO nodes are >guaranteed to have numbers assigned after the last nid. IO nodes >may have number >= 256. Right now there are only three uses of for_each_node(). The two in mm/slab.c only need to worry about nodes that have memory, so from your description above, they won't be at all concerned about the I/O only nodes that they will fail to scan. Any ideas on the relative proportions of nodes with/without memory? Should we consider renaming "for_each_node()" as "for_each_mnode()" (meaning only scan nodes with memory). That would stop mm/slab.c from allocating useless arraycaches for nodes with no memory. But this is turning into a discussion that would have to happen on LKML. Perhaps you could add an /*ACPI3.0-FIXME*/ comment (or some such) to the loop in Dean's patch as a reminder to fix this later? That might also serve as a clue to any janitor that tries to clean up this code back to using "for_each_node()" (and a reminder to me to not take such a patch). -Tony - : send the line "unsubscribe linux-ia64" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
|