#!/usr/bin/perl
#	iorate.pl 1.8 2004-01-20 18:29:07-05 root Exp

# show i/o rates from the /proc/partitions and /proc/stat files

sub read_init;
sub get_partition_names;
sub Usage;
sub dbug;

use Getopt::Std;
getopts("zpPs:m:Hd") || Usage;

# sanity check
Usage if $opt_P && $opt_p;

$SleepTime = ( $opt_s ? $opt_s : 60 );
$MaxSamples = ( $opt_m ? $opt_m : 0 );
$Debug = 0 + $opt_d;
if ($opt_H) {
	$hdr = sprintf("%-5s %10s %10s %10s %10s %10s", "",
		"tot-io/s", "rio/s", "rMB/s", "wio/s", "wMB/s");
}

# get the initial values
$PartFull = $opt_P;
dbug("main", "before call get_part_names");
get_partition_names;
dbug("main", "after names, start init");
read_init(\%baseval);
$Nkeys = keys %baseval;
dbug("main", "after 1st init, hda:rio is $baseval{hda}{rio}, keys: $Nkeys");

# save the base, enter a loop to calculate differences
for $v (keys %baseval) { $current{$v} = $baseval{$v} }

# main output loop
for (;;) {
	my ($new, $did_io);
	my ($NeedHdr) = 1;
	sleep $SleepTime;
	read_init(\%new);

	for $v (sort keys %new) {
		local ($OLD, $NEW);
		*OLD = $current{$v};
		*NEW = $new{$v};
		$did_io = ($NEW{opps}>$OLD{opps});
		dbug("loop", "info: $v, name: $partname{$v}");
		if ((!$opt_z || $did_io) &&
			( $NEW{minor} % 16 ?
				($opt_p || $opt_P) && defined $partname{$v} :
				!$opt_P
			)
		) {
			if ($NeedHdr) {
				# about to print a value, no header yet
				$NeedHdr = 0;
				print ( $opt_H ? "$hdr\n" : "\n" );
			}
			printf("%-5s %10.1f %10.1f %10.2f %10.1f %10.2f %s\n", $v,
				($NEW{opps} - $OLD{opps})/$SleepTime,
				($NEW{rio} - $OLD{rio})/$SleepTime,
				($NEW{rsect} - $OLD{rsect})/($SleepTime*2*1024),
				($NEW{wio} - $OLD{wio})/$SleepTime,
				($NEW{wsect} - $OLD{wsect})/($SleepTime*2*1024),
				$partname{$v}
			);
		}
	}

	# copy the values and delete the fields
	for $v (keys %new) {
		$current{$v} = $new{$v};
		delete $new{$v};
	}

	# end logic here
	last if $MaxSamples && --$MaxSamples == 0;
}	

exit 0;

sub read_init
{
	local (*partinfo) = @_;
	my ($major, $minor, $size, $name,
			$rio, $rmerg, $rsect, $ruse,
			$wio, $wmerg, $wsect, $wuse,
			$running, $use, $avQ
	);
	my ($device, $lun, $wptr, $isdisk, $tmp);
	my ($RHproc, $v26) = (0.0);

	# try to open diskstats first, then partitions
	if (open(PT, "/proc/diskstats")) {
		$v26 = 1;
	} elsif (open(PT, "/proc/partitions")) {
		# check for Redhat or similar with io stats
		$_ = <PT>;
		@headers = split;
		die "No io stats in /proc/partitions" if $#headers < 4;
		$RHproc = 1;
	} else {
		die "No proc f/s? ($!)";
	}
	while (<PT>) {
		if ($v26) {
			($major, $minor, $name, @tmp) = split;
			dbug("v26 split", sprintf("%s %d %d %d", $name, $major, $minor, $#tmp));
			if ($#tmp == 3) {
				# partition
				($rio, $rsect, $wio, $wsect) = @tmp;
				dbug("v26_part", "$name, rio: $rio, wio: $wio");
			} else {
				($rio, $rmerg, $rsect, $ruse,
					$wio, $wmerg, $wsect, $wuse,
					$running, $use, $avQ
				) = @tmp;
			}
		} elsif ($RHproc) {
			($major, $minor, $size, $name,
				$rio, $rmerg, $rsect, $ruse,
				$wio, $wmerg, $wsect, $wuse,
				$running, $use, $avQ
			) = split;
		} else {
			die "stat type not defined";
		}

		dbug("loop-ar", "name=$name, rio=$rio");
		$device = int($minor / 16);
		$lun = $minor % 16;
		$isdisk = $name =~ m/[sh]d[a-z]+$/;
		# check for incomplete partition info
#		next if $ispart{$name} && !($PartFull || ($opt_p && ($rio || $wio)));
		# save a pointer to the base data
		if (defined $partname{$name}) {
			$partinfo{$name} = {};
			*wptr = $partinfo{$name};
			$wptr{major} = $major;
			$wptr{minor} = $minor;
			$wptr{rio} = $rio;
			$wptr{rsect} = $rsect;
			$wptr{wio} = $wio;
			$wptr{wsect} = $wsect;
			$wptr{opps} = $rio + $wio;
		}
	}
	close PT;

	dbug("exit_init", sprintf("keys: %d", keys %partinfo));
}

# get partition names - mount point or "swap"

sub get_partition_names
{
	my ($pname, $size1, $size2, $size3, $used, $mpoint, @tmp);


	# do all the known devices, just to validate
	open(PR, "/proc/partitions") || die "no partition table (no proc fs?)";
	while (<PR>) {
		@tmp = split;
		if ($tmp[0] > 0 && $tmp[2] > 1) {
			$pname = $tmp[3];
			$partname{$pname} = ( $pname =~ m/[hs]d[a-z]+$/ ? "" : "unmounted" );
			$ispart{$pname} = ($tmp[1] > 0);
		}
	}
	close PT;

	# mounts first
	open(DF, "df|") || warn "Can't run df, nameless partitions";
	while (<DF>) {
		($pname, $size1, $size2, $size3, $used, $mpoint) = split;
		if ($pname =~ m#/dev/(\S+)#) {
			$partname{$1} = $mpoint;
		}
	}
	close DF;

	# now identify the swap partitions
	if (open(SW, "/proc/swaps")) {
		while (<SW>) {
			$partname{$1} = "swap" if m#/dev/(\S+)#;
		}
		close SW;
	}
}

# provide usage info on user request

sub Usage
{
	my ($rev) = '1.8';

	print STDERR "diorate - $rev\n"
	. "\n"
	. "Usage:\n"
	. "  diorate [ options ]\n"
	. "\n"
	. "Displays the total, read, and write activity for each drive\n"
	. "on a repeating basis. This may be modified by options to show\n"
	. "activity per partition instead of or in addition to the drive\n"
	. "display.\n"
	. "\n"
	. "Options:\n"
	. " -H   display a header with each data set\n"
	. " -P   display per-partition activity instead of per-drive\n"
	. " -p   display per-partition activity in addition to pre-drive\n"
	. " -z   do not display drives or partitions with no activity\n"
	. " -sNN repeat every NN sec (def: 60)\n"
	. " -mNN display max NN datasets\n"
	. "\n"
	. "Note: -p and -P are mutually exclusive!\n"
	. "\n";

	exit 1;
}

# print debugging info

sub dbug
{
	my ($caller, $text) = @_;

	printf STDERR "%-10s %s\n", $caller, $text if $Debug;
}

# 11 fields
#
#  1 - reads completed
#  2 - reads merged
#  3 - sectors read
#  4 - ms spent reading
#  5 - writes completed
#  6 - writes merged
#  7 - sectors written
#  8 - ms spent writing
#  9 - io in progress
# 10 - ms spent doing io (clear is 9 zero)
# 11 - ms spent doing all io (?)
#
# 2.6 partition info
#
#  1 - reads issues
#  2 - sectors read
#  3 - writes issues
#  4 - sectors written