Re: [RFC PATCH] powerpc/numa: add ability to disable and debug topology updates

Hi Michael,

On 15.09.2014 [15:05:36 +1000], Michael Ellerman wrote:

On Tue, 2014-09-09 at 13:09 -0700, Nishanth Aravamudan wrote:

quoted

We have hit a few customer issues with the topology update code (VPHN
and PRRN). It would be nice to be able to debug the notifications coming
from the hypervisor in both cases to the LPAR, as well as to disable
reacting to the notifications, to narrow down the source of the
problems. Add a basic level of such functionality, similar to the numa=
command-line parameter.

Signed-off-by: Nishanth Aravamudan <redacted>
---
This is pretty rough, but has been useful in the field already. I'm not
sure if more information would be useful than this basic amount.

diff --git a/Documentation/kernel-parameters.txt b/Documentation/kernel-parameters.txt
index 5ae8608ca9f5..6e3b9e3a2ab4 100644
--- a/Documentation/kernel-parameters.txt
+++ b/Documentation/kernel-parameters.txt

@@ -3370,6 +3370,13 @@ bytes respectively. Such letter suffixes can also be entirely omitted.
 			e.g. base its process migration decisions on it.
 			Default is on.
 
+	topology_updates= [KNL, PPC, NUMA]
+			Format: {off | debug}
+			Specify if the kernel should ignore (off) or
+			emit more information (debug) when the
+			hypervisor sends NUMA topology updates to an
+			LPAR.

Boot-time parameters are kind of a pain, not least because they
require a reboot to activate.

Thanks for the review, this is a good point. In the case we're
hitting, we're actually crashing at boot because of possibly dodgy
firmware data.

Does it really need to be a boot param, or could it be a debugfs or
sysctl flag? ie. do we need to disable it immediately at boot or would
it be OK if it was /etc/rc.local or similar that turned it off ?

We need it off at boot, potentially. An LPAR does not indicate that it
will or will not respond to the events in any synchronous fashion, so
the hypervisor is free to send them to us whenever.

It looks like arch_update_cpu_topology() is called quite early, from
init_sched_domains(), but in that case I don't see how
cpu_associativity_changes_mask can be non-zero, ie. we'll never do any
work via that path.

Correct, that path is probably fine. But we can get these hypervisor
events fairly early in boot.

As far as the debug goes, we could just use pr_debug() with
CONFIG_DYNAMIC_DEBUG, it's not quite as easy to enable as a kernel
parameter but for the odd bit of debugging it should be fine.

That's a good point, I wonder with that mechanism if we should perhaps
extend/remove other static debugging methods in favor of that (e.g.,
numa=debug)?

I guess the downside there is you have to do some work before you know
if you're printing anything out. More below ...

quoted

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index d7737a542fd7..72c5ad313cbe 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -1160,6 +1160,28 @@ static int __init early_numa(char *p)
 }
 early_param("numa", early_numa);
 
+static int topology_updates_enabled = 1;
+static int topology_updates_debug = 0;
+
+static int __init early_topology_updates(char *p)
+{
+	if (!p)
+		return 0;
+
+	if (strstr(p, "off")) {
+		printk(KERN_INFO "Disabling topology updates\n");
+		topology_updates_enabled = 0;
+	}

Can you add a:
	#define pr_fmt(fmt) "numa: " fmt

At the top of the file and then use pr_xxx() for these?

Yes.

quoted

+	if (strstr(p, "debug")) {
+		printk(KERN_INFO "Enabling topology updates debug\n");
+		topology_updates_debug = 1;
+	}
+
+	return 0;
+}
+early_param("topology_updates", early_topology_updates);
+
 #ifdef CONFIG_MEMORY_HOTPLUG
 /*
  * Find the node associated with a hot added memory section for

@@ -1546,6 +1568,9 @@ int arch_update_cpu_topology(void)
 	struct device *dev;
 	int weight, new_nid, i = 0;
 
+	if (!topology_updates_enabled)
+		return 0;
+

	weight = cpumask_weight(&cpu_associativity_changes_mask);
	if (!weight)
		return 0;

@@ -1610,6 +1635,25 @@ int arch_update_cpu_topology(void)
 	 *
 	 * And for the similar reason, we will skip all the following updating.
 	 */

The comment should stay attached to the check below of updated_cpus, ie. this
block should preceed the comment.

Yep.

quoted

+	if (topology_updates_debug) {
+		char *buf = kmalloc_array(NR_CPUS*5, sizeof(char), GFP_KERNEL);
+		cpumask_scnprintf(buf, NR_CPUS*5, &updated_cpus);
+		printk(KERN_DEBUG "Topology update for the following CPUs:\n");
+		printk(KERN_DEBUG " %s\n", buf);
+		printk(KERN_DEBUG "cpumask_weight(&updated_cpus)) = %u\n",
+						cpumask_weight(&updated_cpus));

Do we really need to print the cpumask? The same information is
available below right, just in a more verbose form?

Yes, sorry, this is a "living" patch, and I haven't fully cleaned it up.
Will remove the cpumask bodge in v2.

Assuming we can skip that we can just use pr_debug() for these I think
and drop the debug flag. Or is this a super hot-path that I'm not
aware of?

Not super hot, but when it breaks, your LPAR is unusable.

quoted

+		if (cpumask_weight(&updated_cpus)) {
+			for (ud = &updates[0]; ud; ud = ud->next) {
+				printk(KERN_DEBUG "cpu %d moving from node %d "
+						  "to %d\n", ud->cpu,
+						  ud->old_nid, ud->new_nid);
+			}
+		}
+		kfree(buf);
+	}
+
 	if (!cpumask_weight(&updated_cpus))
 		goto out;
 

@@ -1807,8 +1851,10 @@ static const struct file_operations topology_ops = {
 
 static int topology_update_init(void)
 {
-	start_topology_update();
-	proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops);
+	if (topology_updates_enabled) {
+		start_topology_update();
+		proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops);
+	}

It was already broken, but can you fix the (lack of) error handling
here please?

Yep, will make it the first patch in the series.

-Nish