Also fetch node info for non-PCI devices

[ryncsn]

Followring error is observed on some large servers:

[ 40.345505] irq 4: Affinity broken due to vector space exhaustion.

Irq 4 is on node 0 but irqbalance put it in node 1. Root cause is that irq 4 is not an from an PCI device so irqbalance failed to get the right numa node number.

This PR fixes this problem.

[ryncsn]

Hmm, I just noticed there is a issue #169 after I making this PR. That issue is almost the same with the one I'm experiencing.

On my machine after applying this patch the issue is gone. Hopefully I'm getting it right, will this PR will fix the issue?

[nhorman]

yeah, I think you did. We honestly shouldn't have to do this, as the kernel shouldn't issue that error just because the irq is affined to a non-local numa node, but regardless, its better to assign the irq to a local node anyway, so I'll take this.

[legoater]

Hi @ryncsn @nhorman ,

While working on a fix related to IPIs of hotplugged CPUs on PPC, I found out that if Linux maps an IRQ associated with an empty node (hence not exposed under /sys/devices/system/node/), this change will lead to a core dump in place_irq_in_node() because irq_numa_node(info) is NULL.

It would be good to protect the code with a test on place_irq_in_node() or with something like below :

diff --git a/classify.c b/classify.c
index c08144f432bb..9b4640cd4754 100644
--- a/classify.c
+++ b/classify.c
@@ -376,6 +376,11 @@ get_numa_node:
        else
                new->numa_node = get_numa_node(numa_node);
 
+       if (!new->numa_node) {
+               log(TO_CONSOLE, LOG_WARNING, "IRQ %d has an unknown node\n", irq);
+               new->numa_node = get_numa_node(NUMA_NO_NODE);
+       }
+
        cpus_setall(new->cpumask);
        if (devpath != NULL) {
                sprintf(path, "%s/local_cpus", devpath);

What would you prefer ?

Thanks,

C.

[nhorman]

I think that would be fine, open a PR please