Hello,

I have a question regarding the /proc/interrupts file. Let's say for example I had a server with 2xQuad-Core processors (so we have CPU0 to CPU7 ), and with 5 network interfaces.

SMP affinity to all 5 interfaces was set to "ff", so all interfaces have done interrupts on all the processors.

The network interfaces interrupts counters should look like this:

24: 32650776 32670506 50315017 32677739 32672119 32677935 32656299 32667496 IO-APIC-fasteoi eth2
27: 35233448 35285546 35317201 52657622 35315835 35339998 35221092 35246597 IO-APIC-fasteoi eth3
55: 37573056 35870363 35874607 35858010 37728061 35880275 35882340 35882127 PCI-MSI-edge eth1
56: 35861392 35863516 35865123 35865342 35864604 40515822 35861250 35863585 PCI-MSI-edge eth0
58: 30173318 30109146 30002990 30016870 30084937 30034759 46517278 30139103 PCI-MSI-edge eth4

Now let's say I've set the affinity to each of the network interfaces, so that it balances one per processor, so we have 5 processors working only for the network interfaces each processor with it's own eth.

Now let's assume that the network interfaces generate very few interrupts, and that they show up every 5-6 seconds, so watching cat /proc/interrupts doesn`t exactly underline the modifications unless you look with a ruler.

Is there a way to reset the counters on /proc/interrupts so that they all start at 0, without rebooting the system ?

Cheers,

AFAIK not.

And a quick look at the code would appear to confirm that. It merely passes back the counter - no logic to accept input from userspace.

I was thinking that each change in interrupt values does an echo 'cat /some/path/to/some/file/that/contains/the/last/interrupt/value' + 1 > /same/path , and that reset would be a simple echo 0 > /some/path/to/some/file/that/contains/the/last/interrupt/value , to reset the counters, but it`s clearly much more complicated than this.

I don't know about /proc/interrupts but if you for instance look at the kernel source for the /proc/net/netstat counters you'll see the kernel uses a SNMP-like approach to displaying counters, passively emitting values, and most tunables you'd find with 'sysctl -a'.

Nothing is ever that easy. The (kernelspace) code that create the /proc entry determines it attributes. Whether it is writeable from userspace for example - and what to do with anything that is.
For modules exposing (new) data it's easy - for data that already exists, all that generally gets done is exposing the data. Too many codepoints that already use that structure would need (at least) inspection and modification.

In this case the OP could just save the data, then compare it to the previous data each time. Bash awk, perl, python ... pick your favourite.

You could write a driver to do it. The key is that you have to do it from kernel space.

I don't know what would happen if you reset those counters. Should the kernel be using those values for something, someplace, then you could have unexpected consequences.

Easiest just to do it in userspace, keeping the last count and comparing the count you get this time to see what happened in between times.