irqstats

#!/usr/bin/perl -w

use strict;
use warnings;

=head1 NAME

irqstats - Munin plugin to monitor individual interrupts on a Linux machine

=head1 APPLICABLE SYSTEMS

Any Linux system

=head1 CONFIGURATION

Specified devices may be skipped.

   [irqstats]
   env.skipdevs blkif,vif

None needed

=head1 USAGE

Link this plugin to /etc/munin/plugins/ and restart the munin-node.

=head1 INTERPRETATION

The plugin simply shows the number of interrupts on each interrupt
line.  The lines are identified by the attached equipment rather than
the IRQ numbers.

High interrupt rates may coincide with high loads on disk, network or
other I/O, this is normal.

=head1 MAGIC MARKERS

  #%# family=auto
  #%# capabilities=autoconf

=head1 BUGS

None known

=head1 AUTHOR

Unknown

=head1 LICENSE

GPLv2

=cut

use Munin::Plugin;
use Scalar::Util qw(looks_like_number);

my $skipdevs = $ENV{'skipdevs'} || "";

if (defined $ARGV[0] && $ARGV[0] eq 'autoconf') {
    if(-r '/proc/interrupts') {
	print "yes\n";
	exit(0);
    } else {
	print "no (file /proc/interrupts not readable)\n";
	exit(1);
    }
}

my $sun;
$sun = 1 if (`uname -m` =~ /sparc/);

my $in;
my @cpus;
my $cpus;

if ($sun) {
    @cpus = grep (/^cpu\d+/, `cat /proc/stat`);
    $cpus = scalar @cpus;
} else {
    open $in, '<', '/proc/interrupts'
      or die "Can't open /proc/interrupts: $!\n";
    @cpus = split(' ', <$in>);
    $cpus = scalar @cpus;
}

my $cpu;

if ($0 =~ /(?:.*\/)?irqstats_(.+)/) {
    $cpu = $1;
    if ($cpu > $#cpus) {
	die "Requested CPU $cpu out of bounds (0..$#cpus)\n";
    }
}
my @irqs;


sub sum (@) {
    my $sum = 0;
    $sum += (looks_like_number($_) ? $_ : 0) || 0 for @_;	# Avoid complaints about empty strings
    return $sum;
}

LINE: while (my $line = <$in>) {
    my ($irq, $label, $type);
    my @data;
    if ($sun) {
	@data = split (' ', $line, $cpus + 2);
    } else {
	@data = split(' ', $line, $cpus + 3);
    }
    chomp @data;
    $irq = shift @data;
    next unless defined($irq) and length $irq;
    chop $irq;
    if ($irq =~ /^\d+$/) {
	# For numbered interrupts the format is set
	$label = pop @data;
	$type = pop @data;
	# work around "MPIC \d" on powerpc
	if ($type =~ "MPIC" && $label =~ s/^([12])\s+//) {
	    $type .= " $1";;
	}
	# handle hwirq
	if ($label =~ /^(\d+)\s+(.+)/) {
	    $label = $1 eq $irq ? $2 : "$2 [$1]";
	}
	# handle level/edge
	$label =~ s/^(Level|Edge)\s+//;
    } else {
	# For named interrupts it's more confusing, slice and dice
	# to divorce data from commentary
	$label = join(" ",@data[$cpus..$#data]);
	@data = @data[0..$cpus-1];
    }

    # skip devices specified in config
    foreach my $skipdev (split(',', $skipdevs)) {
        next LINE if ($label =~ $skipdev);
    }

    # Skip non-per-cpu values for per-cpu stats
    # This will ignore ERR: and MIS: lines on multi cpu systems
    # (checked on 2.6.22, 2.6.24)
    # next if defined($cpu) and $cpus > @data;

    push @irqs, {
		 irq => $irq,
		 label => $label,
		 count => defined($cpu) ? $data[$cpu] : sum(@data)
		};
}
close $in;

if (defined $ARGV[0] && $ARGV[0] eq 'config') {
    print 'graph_title Individual interrupts',
      defined($cpu) ? " on CPU $cpu\n" : "\n";
    print <<EOM;
graph_args --base 1000 --logarithmic
graph_vlabel interrupts / \${graph_period}
graph_category system
graph_info Shows the number of different IRQs received by the kernel.  High disk or network traffic can cause a high number of interrupts (with good hardware and drivers this will be less so). Sudden high interrupt activity with no associated higher system activity is not normal.

EOM
    print join(' ', 'graph_order', map {"i" . $_->{irq}} @irqs), "\n";
    for my $irq (@irqs) {
	my $f = ($irq->{label} || $irq->{irq});
	$f = $irq->{irq} if (length ($f) > 47);
	print "i", $irq->{irq}, '.label ', $f, "\n";
	if ( $irq->{label} ) {
	  print "i", $irq->{irq}, '.info Interrupt ', $irq->{irq}, ', for device(s): ', $irq->{label}, "\n"
	} elsif ( $irq->{irq} =~ /NMI/i ) {
	  print "i", $irq->{irq}, ".info Nonmaskable interrupt. Either 0 or quite high. If it's normally 0 then just one NMI will often mark some hardware failure.\n";
	} elsif ( $irq->{irq} =~ /LOC/i ) {
	  print "i", $irq->{irq}, ".info Local (per CPU core) APIC timer interrupt.  Until 2.6.21 normally 250 or 1000 per second. On modern 'tickless' kernels it more or less reflects how busy the machine is.\n";
	}

	print "i", $irq->{irq}, ".type DERIVE\n";
	print "i", $irq->{irq}, ".min 0\n";
	print_thresholds("i$irq->{irq}");
    }
} else {
    print "i", $_->{irq}, '.value ', $_->{count}, "\n" for @irqs;
}

__END__