A bit late, but technically the week isn't over yet! :D

On 23/07/21 20:09, Srikar Dronamraju wrote:

* Valentin Schneider [off-list ref] [2021-07-13 17:32:14]:

quoted

Now, let's take examples from your cover letter:

  node distances:
  node   0   1   2   3   4   5   6   7
    0:  10  20  40  40  40  40  40  40
    1:  20  10  40  40  40  40  40  40
    2:  40  40  10  20  40  40  40  40
    3:  40  40  20  10  40  40  40  40
    4:  40  40  40  40  10  20  40  40
    5:  40  40  40  40  20  10  40  40
    6:  40  40  40  40  40  40  10  20
    7:  40  40  40  40  40  40  20  10

But the system boots with just nodes 0 and 1, thus only this distance
matrix is valid:

  node   0   1
    0:  10  20
    1:  20  10

topology_span_sane() is going to use tl->mask(cpu), and as you reported the
NODE topology level should cause issues. Let's assume all offline nodes say
they're 10 distance away from everyone else, and that we have one CPU per
node. This would give us:

No,
All offline nodes would be at a distance of 10 from node 0 only.
So here node distance of all offline nodes from node 1 would be 20.

quoted

  NODE->mask(0) == 0,2-7
  NODE->mask(1) == 1-7

so

NODE->mask(0) == 0,2-7
NODE->mask(1) should be 1
and NODE->mask(2-7) == 0,2-7

Ok, so that shouldn't trigger the warning.

quoted

The intersection is 2-7, we'll trigger the WARN_ON().
Now, with the above snippet, we'll check if that intersection covers any
online CPU. For sched_init_domains(), cpu_map is cpu_active_mask, so we'd
end up with an empty intersection and we shouldn't warn - that's the theory
at least.

Now lets say we onlined CPU 3 and node 3 which was at a actual distance
of 20 from node 0.

(If we only consider online CPUs, and since scheduler masks like
sched_domains_numa_masks arent updated with offline CPUs,)
then

NODE->mask(0) == 0
NODE->mask(1) == 1
NODE->mask(3) == 0,3

Wait, doesn't the distance matrix (without any offline node) say

  distance(0, 3) == 40

? We should have at the very least:

  node   0   1   2   3
    0:  10  20  ??  40
    1:  20  20  ??  40
    2:  ??  ??  ??  ??
    3:  40  40  ??  10

Regardless, NODE->mask(x) is sched_domains_numa_masks[0][x], if

  distance(0,3) > LOCAL_DISTANCE

then

  node0 ∉ NODE->mask(3)

cpumask_and(intersect, tl->mask(cpu), tl->mask(i));
if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) && cpumask_intersects(intersect, cpu_map))

cpu_map is 0,1,3
intersect is 0

From above NODE->mask(0) is !equal to NODE->mask(1) and
cpumask_intersects(intersect, cpu_map) is also true.

I picked Node 3 since if Node 1 is online, we would have faked distance
for Node 2 to be at distance of 40.

Any node from 3 to 7, we would have faced the same problem.

quoted

Looking at sd_numa_mask(), I think there's a bug with topology_span_sane():
it doesn't run in the right place wrt where sched_domains_curr_level is
updated. Could you try the below on top of the previous snippet?

If that doesn't help, could you share the node distances / topology masks
that lead to the WARN_ON()? Thanks.

---

diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index b77ad49dc14f..cda69dfa4065 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c

@@ -1516,13 +1516,6 @@ sd_init(struct sched_domain_topology_level *tl,
     int sd_id, sd_weight, sd_flags = 0;
     struct cpumask *sd_span;

-#ifdef CONFIG_NUMA
-	/*
-	 * Ugly hack to pass state to sd_numa_mask()...
-	 */
-	sched_domains_curr_level = tl->numa_level;
-#endif
-
     sd_weight = cpumask_weight(tl->mask(cpu));

     if (tl->sd_flags)

@@ -2131,7 +2124,12 @@ build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *att

             sd = NULL;
             for_each_sd_topology(tl) {
-
+#ifdef CONFIG_NUMA
+			/*
+			 * Ugly hack to pass state to sd_numa_mask()...
+			 */
+			sched_domains_curr_level = tl->numa_level;
+#endif
                     if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
                             goto error;

I tested with the above patch too. However it still not helping.

Here is the log from my testing.

At Boot.

(Do remember to arrive at sched_max_numa_levels we faked the
numa_distance of node 1 to be at 20 from node 0. All other offline
nodes are at a distance of 10 from node 0.)

[...]

[  168.026621] name=NODE mask(0/0)=0-7 mask(88/7)=0-7,88-89 numa-level=0 curr_level=0
[  168.026626] intersect=0-7 cpu_map=0-19,24-27,32-35,40-43,48-51,56-59,64-67,72-89

Ok so to condense the info, we have:

  node   0   5   7
    0:  10  40  40
    5:  40  10  20
    7:  40  20  10

  node0: 0-7
  node5: 8-29, 32-35, 40-43, 48-51, 56-59, 64-67, 72-87
  node7: 88-95

With the above distance map, we should have

  NODE->mask(CPU0) == 0-7
  NODE->mask(CPU8) == 8-29, 32-35, 40-43, 48-51, 56-59, 64-67, 72-87
  NODE->mask(CPU88) == 88-95

(this is sched_domains_numa_masks[0][CPUx], and
sched_domains_numa_distance[0] == LOCAL_DISTANCE, thus the mask of CPUs
LOCAL_DISTANCE away from CPUx).

For some reason you end up with node0 being part of node7's NODE
mask. Neither nodes are offline, and per the above distance table that
shouldn't happen.

Now this keeps repeating.

I know I have mentioned this before. (So sorry for repeating)

It can't hurt to reformulate ;)

Generally on Power node distance is not populated for offline nodes.
However to arrive at sched_max_numa_levels, we thought of faking few
node distances. In the above case, we faked distance of node 1 as 20
(from node 0) node 5 was already at distance of 40 from node 0.

Right, again that gives us the right set of unique distances (10, 20, 40).

So when sched_domains_numa_masks_set is called to update sd_numa_mask or
sched_domains_numa_masks, all CPUs under node 0 get updated for node 2
too. (since node 2 is shown as at a local distance from node 0). Do
look at the node mask of CPU 88 in the dmesg. It should have been 88,
however its 0-7,88 where 0-7 are coming from node 0.

Even if we skip updation of sched_domains_numa_masks for offline nodes,
on online of a node (i.e when we get the correct node distance), we have
to update the sched_domains_numa_masks to ensure CPUs that were already
present within a certain distance and skipped are added back. And this
was what I tried to do in my patch.

Ok, so it looks like we really can't do without that part - even if we get
"sensible" distance values for the online nodes, we can't divine values for
the offline ones.

--
Thanks and Regards
Srikar Dronamraju