Skip to contentSkip to conversation

Report NUMA node in lspci -v

PATCH In review Matthew Wilcox 2 participants 2 messages Apr 22, 2015 – Sep 14, 2015

Subsystems (2)
PCI Library code
lore
Share & tools
lore

Matthew Wilcox

author
#0 lore raw

PATCH 1/1 Report NUMA node in lspci -v

In multi-socket systems, it's useful to see which node a particular PCI device belongs to. Linux provides this information through sysfs, but some users don't like poking through sysfs themselves to find it, and it's pretty straightforward to report it in lspci.

I should note that when there is no NUMA node for a particular device, Linux reports -1. I've chosen to continue that convention in pciutils, and simply omit the information if the device does not belong to a NUMA node (eg on single-socket systems, or devices which are not preferentially attached to a particular node, like Nehalem-based systems).

This is going to break the library ABI again (and I don't feel confident adding those bits to this patch); might it be worth adding some more padding to struct pci_dev?

lib/access.c +1 -0
index d292085..a547d2c 100644
@@ -28,6 +28,7 @@ pci_alloc_dev(struct pci_access *a)
   d->access = a;
   d->methods = a->methods;
   d->hdrtype = -1;
+  d->numa_node = -1;
   if (d->methods->init_dev)
     d->methods->init_dev(d);
   return d;
lib/pci.h +1 -0
index bd6f6a1..0ccfbe3 100644
@@ -127,6 +127,7 @@ struct pci_dev {
   u16 vendor_id, device_id;		/* Identity of the device */
   u16 device_class;			/* PCI device class */
   int irq;				/* IRQ number */
+  int numa_node;			/* NUMA node */
   pciaddr_t base_addr[6];		/* Base addresses including flags in lower bits */
   pciaddr_t size[6];			/* Region sizes */
   pciaddr_t rom_base_addr;		/* Expansion ROM base address */
lib/sysfs.c +1 -0
index 9f348bb..800fb4e 100644
@@ -192,6 +192,7 @@ static void sysfs_scan(struct pci_access *a)
 	{
 	  sysfs_get_resources(d);
 	  d->irq = sysfs_get_value(d, "irq");
+	  d->numa_node = sysfs_get_value(d, "numa_node");
 	  /*
 	   *  We could read these faster from the config registers, but we want to give
 	   *  the kernel a chance to fix up ID's and especially classes of broken devices.
lspci.c +2 -0
index 6f16142..a78eb26 100644
@@ -757,6 +757,8 @@ show_verbose(struct device *d)
 	printf(", latency %d", latency);
       if (irq)
 	printf(", IRQ " PCIIRQ_FMT, irq);
+      if (p->numa_node != -1)
+	printf(", NUMA node %d", p->numa_node);
       putchar('\n');
     }
#1 lore raw

Hi Matthew!

Quoted 4 lines

In multi-socket systems, it's useful to see which node a particular PCI device belongs to. Linux provides this information through sysfs, but some users don't like poking through sysfs themselves to find it, and it's pretty straightforward to report it in lspci.

First of all, sorry for the delay.

I have applied your patch and improved it slightly: NUMA nodes are reported also in "-vv" and "-nv" modes and everything is done in an ABI-compatible way. Look at the public Git repo, if you wish.

Thanks for the patch!

Have a nice fortnight

Signature · 3 lines
Martin `MJ' Mares                          <[email protected]>   http://mj.ucw.cz/
Faculty of Math and Physics, Charles University, Prague, Czech Rep., Earth
System going down at 5 pm to install scheduler bug.