On Mon, 2007-05-28 at 19:48 +0900, Daisuke Nishimura wrote:
> Hi,
>
> Machines with multi NUMA nodes may panic on bootup.
> Attached patch(for C/S15145), in which I modified
> the initialization order of buddy allocator, fixes
> this problem.
> I tested booting dom0/domVTi and kernel-make on guests.
> Any comments and feedbacks would be appreciated.
>
> I will describe the issue and the cause of it later, but
> I have a few questions:
>
> 1. I moved acpi_table_init(), acpi_numa_init(), and
> smp_build_cpu_map() to early_setup_arch() from late_setup_arch().
> It works and, as far as I read source codes, it seems there is
> no bad effect.
> What do you think ?
Thanks for finding this. I debugged this same issue on our NUMA
systems last week, but haven't had a chance to look for a solution.
> 2. The xenheap area (from xen_pstart to xenheap_phys_end) must exist
> in node0 from its design?
> (As far as I know, if xenheap is not in node0, the initialization
> process of xenheap recursively needs xenheap memory)
Yes, this can happen, and I believe it is happening on my system. HP
systems can be configured with varying amounts of interleaved and node
local memory. Interleaved takes lines of memory from each node and
shows up as a memory only node. In the default configuration, all
memory is interleaved with the CPUs on cpu-only nodes. This means
xenheap doesn't end up on node0. Thanks,
Alex
--
Alex Williamson HP Open Source & Linux Org.
_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel
|
Home