Re: [PATCH] powerpc/numa: Fix percpu allocations to be NUMA aware

Nicholas Piggin [off-list ref] writes:

On Fri,  2 Jun 2017 15:14:47 +1000
Michael Ellerman [off-list ref] wrote:

quoted

In commit 8c272261194d ("powerpc/numa: Enable USE_PERCPU_NUMA_NODE_ID"), we
switched to the generic implementation of cpu_to_node(), which uses a percpu
variable to hold the NUMA node for each CPU.

Unfortunately we neglected to notice that we use cpu_to_node() in the allocation
of our percpu areas, leading to a chicken and egg problem. In practice what
happens is when we are setting up the percpu areas, cpu_to_node() reports that
all CPUs are on node 0, so we allocate all percpu areas on node 0.

This is visible in the dmesg output, as all pcpu allocs being in group 0:

  pcpu-alloc: [0] 00 01 02 03 [0] 04 05 06 07
  pcpu-alloc: [0] 08 09 10 11 [0] 12 13 14 15
  pcpu-alloc: [0] 16 17 18 19 [0] 20 21 22 23
  pcpu-alloc: [0] 24 25 26 27 [0] 28 29 30 31
  pcpu-alloc: [0] 32 33 34 35 [0] 36 37 38 39
  pcpu-alloc: [0] 40 41 42 43 [0] 44 45 46 47

To fix it we need an early_cpu_to_node() which can run prior to percpu being
setup. We already have the numa_cpu_lookup_table we can use, so just plumb it
in. With the patch dmesg output shows two groups, 0 and 1:

  pcpu-alloc: [0] 00 01 02 03 [0] 04 05 06 07
  pcpu-alloc: [0] 08 09 10 11 [0] 12 13 14 15
  pcpu-alloc: [0] 16 17 18 19 [0] 20 21 22 23
  pcpu-alloc: [1] 24 25 26 27 [1] 28 29 30 31
  pcpu-alloc: [1] 32 33 34 35 [1] 36 37 38 39
  pcpu-alloc: [1] 40 41 42 43 [1] 44 45 46 47

We can also check the data_offset in the paca of various CPUs, with the fix we
see:

  CPU 0:  data_offset = 0x0ffe8b0000
  CPU 24: data_offset = 0x1ffe5b0000

And we can see from dmesg that CPU 24 has an allocation on node 1:

  node   0: [mem 0x0000000000000000-0x0000000fffffffff]
  node   1: [mem 0x0000001000000000-0x0000001fffffffff]

Nice bug :)

Yeah what a shocker.

I wonder what happens if you put a WARN if cpu_to_node is used
before it is set up?

Yeah we definitely need some way to debug these things.

quoted

@@ -672,10 +672,19 @@ static void __init pcpu_fc_free(void *ptr, size_t size)
 
 static int pcpu_cpu_distance(unsigned int from, unsigned int to)
 {
-	if (cpu_to_node(from) == cpu_to_node(to))
-		return LOCAL_DISTANCE;
-	else
-		return REMOTE_DISTANCE;
+#ifndef CONFIG_NUMA
+	return LOCAL_DISTANCE;
+#else
+	int from_nid, to_nid;
+
+	from_nid = early_cpu_to_node(from);
+	to_nid   = early_cpu_to_node(to);
+
+	if (from_nid == -1 || to_nid == -1)
+		return LOCAL_DISTANCE;	/* Or assume remote? */
+
+	return node_distance(from_nid, to_nid);

If you made node_distance() return LOCAL_NODE for !NUMA, this
should fall out and not require the ifdef?

Maybe yeah. This is designed to be minimal for backporting though.

cheers