Re: NUMA topology question wrt. d4edc5b6

[Nishanth Aravamudan <nacc@linux.vnet.ibm.com>]

On 23.05.2014 [02:18:05 +0530], Srivatsa S. Bhat wrote:
> 
> [ Adding a few more CC's ]
> 
> On 05/22/2014 01:34 AM, Nishanth Aravamudan wrote:
> > Hi Srivatsa,
> > 
> > After d4edc5b6 ("powerpc: Fix the setup of CPU-to-Node mappings during
> > CPU online"), cpu_to_node() looks like:
> > 
> > static inline int cpu_to_node(int cpu)
> > {
> >         int nid;
> > 
> >         nid = numa_cpu_lookup_table[cpu];
> > 
> >         /*
> >          * During early boot, the numa-cpu lookup table might not have been
> >          * setup for all CPUs yet. In such cases, default to node 0.
> >          */
> >         return (nid < 0) ? 0 : nid;
> > }
> > 
> > However, I'm curious if this is correct in all cases. I have seen
> > several LPARs that do not have any CPUs on node 0. In fact, because node
> > 0 is statically set online in the initialization of the N_ONLINE
> > nodemask, 0 is always present to Linux, whether it is present on the
> > system. I'm not sure what the best thing to do here is, but I'm curious
> > if you have any ideas? I would like to remove the static initialization
> > of node 0, as it's confusing to users to see an empty node (particularly
> > when it's completely separate in the numbering from other nodes), but
> > we trip a panic (refer to:
> > http://www.spinics.net/lists/linux-mm/msg73321.html).
> > 
> 
> Ah, I see. I didn't have any particular reason to default it to zero.
> I just did that because the existing code before this patch did the same
> thing. (numa_cpu_lookup_table[] is a global array, so it will be initialized
> with zeros. So if we access it before populating it via numa_setup_cpu(),
> it would return 0. So I retained that behaviour with the above conditional).

Ok, that seems reasonable to me (keeping the behavior the same as it was
before).

> Will something like the below [totally untested] patch solve the boot-panic?
> I understand that as of today first_online_node will still pick 0 since
> N_ONLINE is initialized statically, but with your proposed change to that
> init code, I guess the following patch should avoid the boot panic.
> 
> [ But note that first_online_node is hard-coded to 0, if MAX_NUMNODES is = 1.
> So we'll have to fix that if powerpc can have a single node system whose node
> is numbered something other than 0. Can that happen as well? ]

I think all single-node systems are only Node 0, but I'm not 100% on
that.

> And regarding your question about what is the best way to fix this
> whole Linux MM's assumption about node0, I'm not really sure.. since I
> am not really aware of the extent to which the MM subsystem is
> intertwined with this assumption and what it would take to cure that
> :-(

Well, at this point, it might be fine to just leave it alone, as it
seems to be more trouble than it's worth -- and really the only
confusion is on those LPARs where there really isn't a Node 0. I'll take
another look later this week.

Thanks,
Nish