Nagios監控SSD健康的腳本

生產中使用了SSD磁盤，使用smartctl -a /dev/sdb可以查看到該磁盤的各項值，我們可以根據這些Value和Worst等於或小於thresh時候就要注意了，下面上腳本:

 

#!/usr/bin/perl

=head
check ssd infomation

usage: add "nagios  ALL=(root)      NOPASSWD: /usr/sbin/smartctl" to /etc/sudoers file;
=cut

use strict;
use warnings;
use Data::Dumper;
use Getopt::Long;

my ($result,$device,$h,$debug);
my ($start_time,$use_time) = (time,0.00);;
my %ssd_attribute = (   5   =>  "Reallocated_Sector_Ct",
                        184 =>  "End_to_End_Error_Detection_Count",
                        225 =>  "Raw_Read_Error_Rate",
                        232 =>  "Available_Reserver_Space",
                        233 =>  "Media_Wearout_Indicator",
                        9   =>  "Power_On_Hours",
                    );

$result = GetOptions ("device=s" => \$device,
                      "debug"    => \$debug);

$device ||= 'sdb';
$debug ||=0;

$h->{$device}->{output} = "";
$h->{$device}->{perfdata} = "";
$h->{$device}->{status} = 0;
$h->{$device}->{total_info} = `sudo /usr/sbin/smartctl -a /dev/$device 2>&1`;
if ( $h->{$device}->{total_info} =~ m{===\s+START\s+OF\s+INFORMATION\s+SECTION\s+===(.*)===\s+START\s+OF\s+READ\s+SMART\s+DATA\s+SECTION\s+===\s+SMART\s+overall-health\s+self-assessment\s+test\s+result:\s+(\w+)[\d\D]+Vendor\s+Specific\s+SMART\s+Attributes\s+with\s+Thresholds([\d\D]+)SMART\s+Error\s+Log\s+Version:\s+(\d+)}is) {
    $h->{$device}->{info_section} = $1;
    $h->{$device}->{smart_test_result} = $2;
    $h->{$device}->{healt_result} = $3;
    print "************************************** get $device healt info sta **************************************\n" if $debug;
    foreach my $line (split /\n/,$h->{$device}->{healt_result}) {
        #if ($line =~ m{(\d+)\s+([^\s]+)\s+([\dx]+)\s+(\d+)\s+(\d+)\s+(\d+)\s+(\w+)\s+(\w+)\s+([^\s]+)\s+(\d+)}i) {
        if ($line =~ m{(\d+)\s+([^\s]+)\s+([^\s]+)\s+([^\s]+)\s+([^\s]+)\s+([^\s]+)\s+([^\s]+)\s+([^\s]+)\s+([^\s]+)\s+([^\s]+)}i) {
            print "ID:$1\tATTRIBUTE_NAME:$2\tFLAG:$3\tVALUE:$4\tWORST:$5\tTHRESH:$6\tTYPE:$7\tUPDATED:$8\tWHEN_FAILED:$9\tRAW_VALUE:$10\n" if $debug;
            $h->{$device}->{healt}->{$1}->{id} = $1;
            $h->{$device}->{healt}->{$1}->{attribute_name} = $2;
            $h->{$device}->{healt}->{$1}->{flag} = $3;
            $h->{$device}->{healt}->{$1}->{value} = $4;
            $h->{$device}->{healt}->{$1}->{worst} = $5;
            $h->{$device}->{healt}->{$1}->{thresh} = $6;
            $h->{$device}->{healt}->{$1}->{type} = $7;
            $h->{$device}->{healt}->{$1}->{updated} = $8;
            $h->{$device}->{healt}->{$1}->{when_failed} = $9;
            $h->{$device}->{healt}->{$1}->{raw_value} = $10;
        }
    }
    print "************************************** get $device healt info end **************************************\n" if $debug;
    $h->{$device}->{smart_error_log_version} = $4;
    $h->{$device}->{match} = 1;
} else {
    $h->{$device}->{match} = 0;
}

print "runging..... `sudo /usr/sbin/smartctl -a /dev/$device 2>&1`\n" if $debug;
print "\n\n-------------------------------- Dumper \$h sta --------------------------------\n" if $debug;
print Dumper $h if $debug;
print "-------------------------------- Dumper \$h end --------------------------------\n\n" if $debug;

$use_time = sprintf("%0.2f",time - $start_time);
if ( (exists $h->{$device}->{match} && $h->{$device}->{match} == 0) ) {
    print "CRITICAL - smartctl get $device total info fail|status=1 time=$use_time\n";
    exit (2);
} elsif ( ! exists $h->{$device}->{healt} ) {
    print "WARNING - smartctl get $device healt info fail|status=1 time=$use_time\n";
    exit (1);
} else {
    while ( (my ($id,$id_hash)) = (each %{$h->{$device}->{healt}} ) ) {
        if ( !exists $ssd_attribute{$id}) {
            print "not exists \$ssd_attribute{\$id},now next\n" if $debug;
            next;
        }
        print "----------------------------- loop \$h->{\$device}->{healt} hash -----------------------------\n" if $debug;
        print $h->{$device}->{healt}->{$id}->{worst} . "\t" if $debug;
        print $h->{$device}->{healt}->{$id}->{value} . "\t" if $debug;
        print $h->{$device}->{healt}->{$id}->{thresh} . "\n" if $debug;
        if ($h->{$device}->{healt}->{$id}->{value} <= $h->{$device}->{healt}->{$id}->{thresh}){
            $h->{$device}->{output} .= "CRITICAL - $device " if ($h->{$device}->{output} eq "");
            $h->{$device}->{output} .= "id:$id attribute_name:" . $ssd_attribute{$id} . " value:" . $h->{$device}->{healt}->{$id}->{value} . " ";
            $h->{$device}->{status} = 2;
            print 'value <= thersh' . "\n" if $debug;
        } elsif ($h->{$device}->{healt}->{$id}->{worst} <= $h->{$device}->{healt}->{$id}->{thresh}) {
            $h->{$device}->{output} .= "WARNING - $device " if ($h->{$device}->{output} eq "");
            $h->{$device}->{output} .= "id:$id attribute_name:" . $ssd_attribute{$id} . " value:" . $h->{$device}->{healt}->{$id}->{value} . " ";
            $h->{$device}->{status} = 1;
            print 'value <= thersh' . "\n" if $debug;
        }
    }
}


if ($h->{$device}->{status} == 2) {
    print $h->{$device}->{output} . "|status=" . $h->{$device}->{status} . " time=$use_time\n";
    exit ($h->{$device}->{status});
} elsif ($h->{$device}->{status} == 1) {
    print $h->{$device}->{output} . "|status=" . $h->{$device}->{status} . " time=$use_time\n";
    exit ($h->{$device}->{status});
} else {
    print "OK - $device SSD attribute healt" .  "|status=" . $h->{$device}->{status} . " time=$use_time\n";
    exit ($h->{$device}->{status});
}