On 1/27/08, Balbir Singh [off-list ref] wrote:

* Paul Mackerras [off-list ref] [2008-01-27 22:55:43]:

quoted

Balbir Singh writes:

quoted

Here's a better and more complete fix for the problem. Could you
please see if it works for you? I tested it on a real NUMA box and it
seemed to work fine there.

There are a couple of other changes in behaviour that your patch
introduces, and I'd like to understand them better before taking the
patch.  First, with your patch we don't set nodes online if they end
up having no memory in them because of the memory limit, whereas
previously we did.  Secondly, in the case where we don't have NUMA
information, we now set node 0 online after adding each LMB, whereas
previously we only set it online once.

If in fact these changes are benign, then your patch description
should mention them and explain why they are benign.

Yes, they are. I'll try and justify the changes with a good detailed
changelog. If people prefer it, I can hide fake NUMA nodes under a
config option, so that it does not come enabled by default.

Sigh, there already *is* a fake NUMA config option: CONFIG_NUMA_EMU.

"CONFIG_NUMA_EMU:
  Enable NUMA emulation. A flat machine will be split
  into virtual nodes when booted with "numa=fake=N", where N is the
  number of nodes. This is only useful for debugging."

I have to assume your patch is implementing the same feature for
powerpc (really just extending the x86_64 one), and thus should share
the config option.

Any chance you can just make some of that code common? Maybe as a
follow-on patch. I expect that some of Mel's (added to Cc) work to
allow NUMA to be set on x86 more easily will flow quite simply into
adding fake NUMA support there as well. So moving the code to a common
place (at least the parsing) makes sense.

I also feel like you want to be able to online memoryless nodes --
that's where we've been hitting a number of bugs lately in the VM. I
can't tell from Paul's comment if your patch prevents that from being
faked or not.

Thanks,
Nish