Re: [PATCH] mm: vmstat: Use zeroed stats for unpopulated zones

[Michal Hocko <mhocko@suse.com>]

On Wed 06-05-20 17:09:51, Vlastimil Babka wrote:
> On 5/6/20 4:02 PM, Michal Hocko wrote:
> > On Wed 06-05-20 15:33:36, Vlastimil Babka wrote:
> >> On 5/4/20 12:26 PM, Michal Hocko wrote:
> >> > On Mon 04-05-20 12:33:04, Sandipan Das wrote:
> >> >> For unpopulated zones, the pagesets point to the common
> >> >> boot_pageset which can have non-zero vm_numa_stat counts.
> >> >> Because of this memory-less nodes end up having non-zero
> >> >> NUMA statistics. This can be observed on any architecture
> >> >> that supports memory-less NUMA nodes.
> >> >> 
> >> >> E.g.
> >> >> 
> >> >>   $ numactl -H
> >> >>   available: 2 nodes (0-1)
> >> >>   node 0 cpus: 0 1 2 3
> >> >>   node 0 size: 0 MB
> >> >>   node 0 free: 0 MB
> >> >>   node 1 cpus: 4 5 6 7
> >> >>   node 1 size: 8131 MB
> >> >>   node 1 free: 6980 MB
> >> >>   node distances:
> >> >>   node   0   1
> >> >>     0:  10  40
> >> >>     1:  40  10
> >> >> 
> >> >>   $ numastat
> >> >>                              node0           node1
> >> >>   numa_hit                     108           56495
> >> >>   numa_miss                      0               0
> >> >>   numa_foreign                   0               0
> >> >>   interleave_hit                 0            4537
> >> >>   local_node                   108           31547
> >> >>   other_node                     0           24948
> >> >> 
> >> >> Hence, return zero explicitly for all the stats of an
> >> >> unpopulated zone.
> >> > 
> >> > I hope I am not just confused but I would expect that at least
> >> > numa_foreign and other_node to be non zero.
> >> Hmm, checking zone_statistics():
> >> 
> >> NUMA_FOREIGN increment uses preferred zone, which is the first in zone in
> >> zonelist, so it will be a zone from node 1 even for allocations on cpu
> >> associated to node 0 - assuming node 0's unpopulated zones are not included in
> >> node 0's zonelist.
> > 
> > But the allocation could have been requested for node 0 regardless of
> > the amount of memory the node has.
> 
> Yes, if we allocate from cpu 0-3 then it should be a miss on node 0. But the
> zonelists are optimized in a way that they don't include empty zones -
> build_zonerefs_node() checks managed_zone(). As a result, node 0 zonelist has no
> node 0 zones, which confuses the stats code. We should probably document that
> numa stats are bogus on systems with memoryless nodes. This patch makes it
> somewhat more obvious by presenting nice zeroes on the memoryless node itself,
> but node 1 now include stats from node 0.

Thanks for the clarification. So the underlying problem is that zone_statistics
operates on a preferred zone rather than node. This would be fixable but
I am not sure whether this is something worth bothering. Maybe it would
just be more convenient to document the unfortunate memory less nodes
stats situation and be done with it. Or do we have any consumers that
really do care?

> >> NUMA_OTHER uses numa_node_id(), which would mean the node 0's cpus have node 1
> >> in their numa_node_id() ? Is that correct?
> > 
> > numa_node_id should reflect the real node the CPU is associated with.
> 
> You're right, numa_node_id() is probably fine. But NUMA_OTHER is actually
> incremented at the zone where the allocation succeeds. This probably doesn't
> match Documentation/admin-guide/numastat.rst, even on a non-memoryless-node systems:
> 
> other_node      A process ran on this node and got memory from another node.

Yeah, the documentation doesn't match the implementation. Maybe we
should just fix the documentation because this has been the case for
ages.

-- 
Michal Hocko
SUSE Labs