Messages in this thread | | | Date | Tue, 28 Jul 2009 19:04:07 +0900 (JST) | Subject | Re: [BUGFIX] set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware | From | "KAMEZAWA Hiroyuki" <> |
| |
Andi Kleen wrote: > KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> writes: > >> tested on x86-64/fake NUMA and ia64/NUMA. >> (That ia64 is a host which orignal bug report used.) >> >> Maybe this is bigger patch than expected, but NODEMASK_ALLOC() will be a >> way >> to go, anyway. (even if CPUMASK_ALLOC is not used anyware yet..) >> Kosaki tested this on ia64 NUMA. thanks. > > Note that low/high memory support in NUMA policy is only partial > anyways, e.g. the VMA based policy only supports a single zone. That > was by design choice and because NUMA has a lot of issues on 32bit due > to the limited address space and is not widely used. > > So small fixes are ok but I wouldn't go to large effort to fix NUMA > policy on 32bit. > ya, maybe you mention to something related to policy_zone. It's checked only in bind code now. (from before 2.6.30..) The bug was accessing NODE_DATA(nid), which is NULL. (All possible node doesn't have pgdat)
Thanks, -Kame
| |