From: Luck, T. <ton...@in...> - 2001-12-10 19:46:10
|
Manfred Spraul wrote: > But I think pointer to the per-cpu structures should be part of > kmem_cache_s: Every kmalloc and kfree must look up the per-cpu > structure, and embedding it directly would avoid one level of indirection. Yes, you are right (and that removes my need to invent a "logical cpu number on node" that I was going to need ... thanks!) > The array for foreign free calls is just an idea: If there are lots of > foreign free calls, then a per-cpu array and using only one global > spinlock for all nodes instead of a per-node spinlock might be faster. > But we need statistics to decide that. It should be relaively easy to switch algorithms once we have the basic infrastructure up and running ... then we can get some real data and see which way to go. -Tony Luck |