]>
git.proxmox.com Git - pve-storage.git/blob - PVE/Diskmanage.pm
1 package PVE
::Diskmanage
;
11 use PVE
::Tools
qw(extract_param run_command file_get_contents file_read_firstline dir_glob_regex dir_glob_foreach trim);
13 my $SMARTCTL = "/usr/sbin/smartctl";
14 my $ZPOOL = "/sbin/zpool";
15 my $SGDISK = "/sbin/sgdisk";
16 my $PVS = "/sbin/pvs";
17 my $LVS = "/sbin/lvs";
18 my $UDEVADM = "/bin/udevadm";
19 my $LSBLK = "/bin/lsblk";
21 sub verify_blockdev_path
{
24 die "missing path" if !$rel_path;
25 my $path = abs_path
($rel_path);
26 die "failed to get absolute path to $rel_path\n" if !$path;
28 die "got unusual device path '$path'\n" if $path !~ m
|^/dev/(.*)$|;
30 $path = "/dev/$1"; # untaint
32 assert_blockdev
($path);
38 my ($dev, $noerr) = @_;
40 if ($dev !~ m
|^/dev/| || !(-b
$dev)) {
41 return undef if $noerr;
42 die "not a valid block device\n";
49 my ($disk, $uuid) = @_;
51 assert_blockdev
($disk);
53 # we should already have checked if it is in use in the api call
54 # but we check again for safety
55 die "disk $disk is already in use\n" if disk_is_used
($disk);
57 my $id = $uuid || 'R';
58 run_command
([$SGDISK, $disk, '-U', $id]);
68 my $disklist = get_disks
($dev, 1);
70 die "'$disk' is not a valid local disk\n" if !defined($disklist->{$dev});
71 return 1 if $disklist->{$dev}->{used
};
77 my ($disk, $healthonly) = @_;
79 assert_blockdev
($disk);
85 if ($disk =~ m!^/dev/(nvme\d+n\d+)$!) {
86 my $info = get_sysdir_info
("/sys/block/$1");
87 $disk = "/dev/".($info->{device
}
88 or die "failed to get nvme controller device for $disk\n");
91 my $cmd = [$SMARTCTL, '-H'];
92 push @$cmd, '-A', '-f', 'brief' if !$healthonly;
96 $returncode = run_command
($cmd, noerr
=> 1, outfunc
=> sub{
99 # ATA SMART attributes, e.g.:
100 # ID# ATTRIBUTE_NAME FLAGS VALUE WORST THRESH FAIL RAW_VALUE
101 # 1 Raw_Read_Error_Rate POSR-K 100 100 000 - 0
103 # SAS and NVME disks, e.g.:
104 # Data Units Written: 5,584,952 [2.85 TB]
105 # Accumulated start-stop cycles: 34
107 if (defined($type) && $type eq 'ata' && $line =~ m/^([ \d]{2}\d)\s+(\S+)\s+(\S{6})\s+(\d+)\s+(\d+)\s+(\S+)\s+(\S+)\s+(.*)$/) {
111 $entry->{name
} = $2 if defined $2;
112 $entry->{flags
} = $3 if defined $3;
113 # the +0 makes a number out of the strings
114 $entry->{value
} = $4+0 if defined $4;
115 $entry->{worst
} = $5+0 if defined $5;
116 # some disks report the default threshold as --- instead of 000
117 if (defined($6) && $6 eq '---') {
118 $entry->{threshold
} = 0;
120 $entry->{threshold
} = $6+0 if defined $6;
122 $entry->{fail
} = $7 if defined $7;
123 $entry->{raw
} = $8 if defined $8;
124 $entry->{id
} = $1 if defined $1;
125 push @{$smartdata->{attributes
}}, $entry;
126 } elsif ($line =~ m/(?:Health Status|self\-assessment test result): (.*)$/ ) {
127 $smartdata->{health
} = $1;
128 } elsif ($line =~ m/Vendor Specific SMART Attributes with Thresholds:/) {
130 delete $smartdata->{text
};
131 } elsif ($line =~ m/=== START OF (READ )?SMART DATA SECTION ===/) {
133 } elsif (defined($type) && $type eq 'text') {
134 $smartdata->{text
} = '' if !defined $smartdata->{text
};
135 $smartdata->{text
} .= "$line\n";
136 # extract wearout from nvme text, allow for decimal values
137 if ($line =~ m/Percentage Used:.*(\d+(?:\.\d+)?)\%/i) {
138 $smartdata->{wearout
} = 100 - $1;
140 } elsif ($line =~ m/SMART Disabled/) {
141 $smartdata->{health
} = "SMART Disabled";
147 # bit 0 and 1 mark an severe smartctl error
148 # all others are for disk status, so ignore them
150 if ((defined($returncode) && ($returncode & 0b00000011)) || $err) {
151 die "Error getting S.M.A.R.T. data: Exit code: $returncode\n";
154 $smartdata->{type
} = $type;
159 sub get_parttype_info
() {
160 my $cmd = [$LSBLK, '--json', '-o', 'path,parttype'];
164 run_command
($cmd, outfunc
=> sub {
166 $output .= "$line\n";
170 return $res if $output eq '';
172 my $parsed = eval { decode_json
($output) };
174 my $list = $parsed->{blockdevices
} // [];
176 foreach my $dev (@$list) {
177 next if !($dev->{parttype
});
178 my $type = $dev->{parttype
};
179 $res->{$type} = [] if !defined($res->{$type});
180 push @{$res->{$type}}, $dev->{path
};
186 my $get_devices_by_partuuid = sub {
187 my ($parttype_map, $uuids, $res) = @_;
189 $res = {} if !defined($res);
191 foreach my $uuid (sort keys %$uuids) {
192 map { $res->{$_} = $uuids->{$uuid} } @{$parttype_map->{$uuid}};
198 sub get_zfs_devices
{
199 my ($parttype_map) = @_;
202 return {} if ! -x
$ZPOOL;
204 # use zpool and parttype uuid,
205 # because log and cache do not have
208 run_command
([$ZPOOL, 'list', '-HPLv'], outfunc
=> sub {
211 if ($line =~ m
|^\t([^\t]+)\t|) {
218 # because maybe zfs tools are not installed
222 "6a898cc3-1dd2-11b2-99a6-080020736631" => 1, # apple
223 "516e7cba-6ecf-11d6-8ff8-00022d09712b" => 1, # bsd
227 $res = $get_devices_by_partuuid->($parttype_map, $uuids, $res);
232 sub get_lvm_devices
{
233 my ($parttype_map) = @_;
236 run_command
([$PVS, '--noheadings', '--readonly', '-o', 'pv_name'], outfunc
=> sub{
239 if ($line =~ m
|^/dev/|) {
245 # if something goes wrong, we do not want
246 # to give up, but indicate an error has occured
250 "e6d6d379-f507-44c2-a23c-238f2a3df928" => 1,
253 $res = $get_devices_by_partuuid->($parttype_map, $uuids, $res);
258 sub get_ceph_journals
{
259 my ($parttype_map) = @_;
263 '45b0969e-9b03-4f30-b4c6-b4b80ceff106' => 1, # journal
264 '30cd0809-c2b2-499c-8879-2d6b78529876' => 2, # db
265 '5ce17fce-4087-4169-b7ff-056cc58473f9' => 3, # wal
266 'cafecafe-9b03-4f30-b4c6-b4b80ceff106' => 4, # block
269 $res = $get_devices_by_partuuid->($parttype_map, $uuids, $res);
274 # reads the lv_tags and matches them with the devices
275 sub get_ceph_volume_infos
{
278 my $cmd = [ $LVS, '-S', 'lv_name=~^osd-', '-o', 'devices,lv_name,lv_tags',
279 '--noheadings', '--readonly', '--separator', ';' ];
281 run_command
($cmd, outfunc
=> sub {
283 $line =~ s/(?:^\s+)|(?:\s+$)//g; # trim whitespaces
285 my $fields = [ split(';', $line) ];
287 # lvs syntax is /dev/sdX(Y) where Y is the start (which we do not need)
288 my ($dev) = $fields->[0] =~ m
|^(/dev/[a-z
]+)|;
289 if ($fields->[1] =~ m
|^osd-
([^-]+)-|) {
291 # $result autovivification is wanted, to not creating empty hashes
292 if (($type eq 'block' || $type eq 'data') && $fields->[2] =~ m/ceph.osd_id=([^,]+)/) {
293 $result->{$dev}->{osdid
} = $1;
294 $result->{$dev}->{bluestore
} = ($type eq 'block');
296 # undef++ becomes '1' (see `perldoc perlop`: Auto-increment)
297 $result->{$dev}->{$type}++;
311 run_command
([$UDEVADM, 'info', '-p', $dev, '--query', 'all'], outfunc
=> sub {
317 return undef if !$info;
319 return undef if $info !~ m/^E: DEVTYPE=disk$/m;
320 return undef if $info =~ m/^E: ID_CDROM/m;
322 # we use this, because some disks are not simply in /dev
323 # e.g. /dev/cciss/c0d0
324 if ($info =~ m/^E: DEVNAME=(\S+)$/m) {
325 $data->{devpath
} = $1;
327 return if !defined($data->{devpath
});
329 $data->{serial
} = 'unknown';
330 if ($info =~ m/^E: ID_SERIAL_SHORT=(\S+)$/m) {
331 $data->{serial
} = $1;
335 if ($info =~ m/^E: ID_PART_TABLE_TYPE=gpt$/m) {
341 if ($info =~ m/^E: ID_ATA_ROTATION_RATE_RPM=(\d+)$/m) {
345 if ($info =~ m/^E: ID_BUS=usb$/m) {
349 if ($info =~ m/^E: ID_MODEL=(.+)$/m) {
353 $data->{wwn
} = 'unknown';
354 if ($info =~ m/^E: ID_WWN=(.*)$/m) {
361 sub get_sysdir_info
{
364 return undef if ! -d
"$sysdir/device";
368 my $size = file_read_firstline
("$sysdir/size");
369 return undef if !$size;
371 # linux always considers sectors to be 512 bytes,
372 # independently of real block size
373 $data->{size
} = $size * 512;
375 # dir/queue/rotational should be 1 for hdd, 0 for ssd
376 $data->{rotational
} = file_read_firstline
("$sysdir/queue/rotational") // -1;
378 $data->{vendor
} = file_read_firstline
("$sysdir/device/vendor") || 'unknown';
379 $data->{model
} = file_read_firstline
("$sysdir/device/model") || 'unknown';
381 if (defined(my $device = readlink("$sysdir/device"))) {
382 # strip directory and untaint:
383 ($data->{device
}) = $device =~ m!([^/]+)$!;
389 sub get_wear_leveling_info
{
390 my ($smartdata, $model) = @_;
391 my $attributes = $smartdata->{attributes
};
393 if (defined($smartdata->{wearout
})) {
394 return $smartdata->{wearout
};
408 # find target attr id
412 foreach my $vendor (keys %$vendormap) {
413 if ($model =~ m/$vendor/i) {
414 $attrid = $vendormap->{$vendor};
415 # found the attribute
421 $attrid = $vendormap->{default};
424 foreach my $attr (@$attributes) {
425 next if $attr->{id
} != $attrid;
426 $wearout = $attr->{value
};
436 my $dh = IO
::Dir-
>new ($dir);
439 while (defined(my $tmp = $dh->read)) {
440 next if $tmp eq '.' || $tmp eq '..';
451 if (-l
$sysdir && readlink($sysdir) =~ m
|host
[^/]*/session
[^/]*|) {
459 my ($disks, $nosmart) = @_;
464 my $mounts = PVE
::ProcFSTools
::parse_proc_mounts
();
466 foreach my $mount (@$mounts) {
467 next if $mount->[0] !~ m
|^/dev/|;
468 $mounted->{abs_path
($mount->[0])} = $mount->[1];
471 my $dev_is_mounted = sub {
473 return $mounted->{$dev};
476 my $parttype_map = get_parttype_info
();
478 my $journalhash = get_ceph_journals
($parttype_map);
479 my $ceph_volume_infos = get_ceph_volume_infos
();
481 my $zfshash = get_zfs_devices
($parttype_map);
483 my $lvmhash = get_lvm_devices
($parttype_map);
485 my $disk_regex = ".*";
486 if (defined($disks)) {
489 } elsif (ref($disks) ne 'ARRAY') {
490 die "disks is not a string or array reference\n";
492 # we get cciss/c0d0 but need cciss!c0d0
493 map { s
|cciss
/|cciss
!| } @$disks;
495 $disk_regex = "(?:" . join('|', @$disks) . ")";
498 dir_glob_foreach
('/sys/block', $disk_regex, sub {
500 # whitelisting following devices
501 # hdX: ide block device
502 # sdX: sd block device
503 # vdX: virtual block device
504 # xvdX: xen virtual block device
505 # nvmeXnY: nvme devices
506 # cciss!cXnY: cciss devices
507 return if $dev !~ m/^(h|s|x?v)d[a-z]+$/ &&
508 $dev !~ m/^nvme\d+n\d+$/ &&
509 $dev !~ m/^cciss\!c\d+d\d+$/;
511 my $data = get_udev_info
("/sys/block/$dev");
512 return if !defined($data);
513 my $devpath = $data->{devpath
};
515 my $sysdir = "/sys/block/$dev";
517 # we do not want iscsi devices
518 return if is_iscsi
($sysdir);
520 my $sysdata = get_sysdir_info
($sysdir);
521 return if !defined($sysdata);
523 my $type = 'unknown';
525 if ($sysdata->{rotational
} == 0) {
528 } elsif ($sysdata->{rotational
} == 1) {
529 if ($data->{rpm
} != -1) {
531 } elsif ($data->{usb
}) {
537 my $health = 'UNKNOWN';
542 my $smartdata = get_smart_data
($devpath, ($type ne 'ssd'));
543 $health = $smartdata->{health
} if $smartdata->{health
};
545 if ($type eq 'ssd') {
546 # if we have an ssd we try to get the wearout indicator
547 my $wearval = get_wear_leveling_info
($smartdata, $data->{model
} || $sysdata->{model
});
548 $wearout = $wearval if defined($wearval);
555 $used = 'LVM' if $lvmhash->{$devpath};
557 $used = 'mounted' if &$dev_is_mounted($devpath);
559 $used = 'ZFS' if $zfshash->{$devpath};
561 # we replaced cciss/ with cciss! above
562 # but in the result we need cciss/ again
563 # because the caller might want to check the
564 # result again with the original parameter
565 if ($dev =~ m
|^cciss
!|) {
566 $dev =~ s
|^cciss
!|cciss
/|;
569 $disklist->{$dev} = {
570 vendor
=> $sysdata->{vendor
},
571 model
=> $data->{model
} || $sysdata->{model
},
572 size
=> $sysdata->{size
},
573 serial
=> $data->{serial
},
586 my $journal_count = 0;
590 my $found_partitions;
592 my $found_mountpoints;
595 my $partpath = $devpath;
597 # remove part after last / to
598 # get the base path for the partitions
599 # e.g. from /dev/cciss/c0d0 get /dev/cciss
600 $partpath =~ s/\/[^\/]+$//;
602 dir_glob_foreach
("$sysdir", "$dev.+", sub {
605 $found_partitions = 1;
607 if (my $mp = &$dev_is_mounted("$partpath/$part")) {
608 $found_mountpoints = 1;
609 if ($mp =~ m
|^/var/lib
/ceph/osd/ceph-
(\d
+)$|) {
614 if ($lvmhash->{"$partpath/$part"}) {
618 if ($zfshash->{"$partpath/$part"}) {
622 if (my $journal_part = $journalhash->{"$partpath/$part"}) {
623 $journal_count++ if $journal_part == 1;
624 $db_count++ if $journal_part == 2;
625 $wal_count++ if $journal_part == 3;
626 $bluestore = 1 if $journal_part == 4;
629 if (!dir_is_empty
("$sysdir/$part/holders") && !$found_lvm) {
634 if (my $ceph_volume = $ceph_volume_infos->{$devpath}) {
635 $journal_count += $ceph_volume->{journal
} // 0;
636 $db_count += $ceph_volume->{db
} // 0;
637 $wal_count += $ceph_volume->{wal
} // 0;
638 if (defined($ceph_volume->{osdid
})) {
639 $osdid = $ceph_volume->{osdid
};
640 $bluestore = 1 if $ceph_volume->{bluestore
};
644 $used = 'mounted' if $found_mountpoints && !$used;
645 $used = 'LVM' if $found_lvm && !$used;
646 $used = 'ZFS' if $found_zfs && !$used;
647 $used = 'Device Mapper' if $found_dm && !$used;
648 $used = 'partitions' if $found_partitions && !$used;
650 # multipath, software raid, etc.
651 # this check comes in last, to show more specific info
653 $used = 'Device Mapper' if !$used && !dir_is_empty
("$sysdir/holders");
655 $disklist->{$dev}->{used
} = $used if $used;
656 $disklist->{$dev}->{osdid
} = $osdid;
657 $disklist->{$dev}->{journals
} = $journal_count if $journal_count;
658 $disklist->{$dev}->{bluestore
} = $bluestore if $osdid != -1;
659 $disklist->{$dev}->{db
} = $db_count if $db_count;
660 $disklist->{$dev}->{wal
} = $wal_count if $wal_count;
668 my ($part_path) = @_;
670 my ($mode, $rdev) = (stat($part_path))[2,6];
672 next if !$mode || !S_ISBLK
($mode) || !$rdev;
673 my $major = PVE
::Tools
::dev_t_major
($rdev);
674 my $minor = PVE
::Tools
::dev_t_minor
($rdev);
675 my $partnum_path = "/sys/dev/block/$major:$minor/";
679 $partnum = file_read_firstline
("${partnum_path}partition");
681 die "Partition does not exists\n" if !defined($partnum);
683 #untaint and ensure it is a int
684 if ($partnum =~ m/(\d+)/) {
686 die "Partition number $partnum is invalid\n" if $partnum > 128;
688 die "Failed to get partition number\n";
695 my ($part_path) = @_;
697 my $dev = $1 if $part_path =~ m
|^/dev/(.*)$|;
698 my $link = readlink "/sys/class/block/$dev";
699 my $block_dev = $1 if $link =~ m
|([^/]*)/$dev$|;
701 die "Can't parse parent device\n" if !defined($block_dev);
702 die "No valid block device\n" if index($dev, $block_dev) == -1;
704 $block_dev = "/dev/$block_dev";
705 die "Block device does not exsists\n" if !(-b
$block_dev);
710 sub locked_disk_action
{
712 my $res = PVE
::Tools
::lock_file
('/run/lock/pve-diskmanage.lck', undef, $sub);
717 sub assert_disk_unused
{
720 die "device '$dev' is already in use\n" if disk_is_used
($dev);
725 sub append_partition
{
726 my ($dev, $size) = @_;
729 $devname =~ s
|^/dev/||;
732 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*?
(\d
+)/, sub {
733 my ($part, $partid) = @_;
735 if ($partid >= $newpartid) {
736 $newpartid = $partid + 1;
740 $size = PVE
::Tools
::convert_size
($size, 'b' => 'mb');
742 run_command
([ $SGDISK, '-n', "$newpartid:0:+${size}M", $dev ],
743 errmsg
=> "error creating partition '$newpartid' on '$dev'");
747 # loop again to detect the real partiton device which does not always follow
748 # a strict $devname$partition scheme like /dev/nvme0n1 -> /dev/nvme0n1p1
749 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*$newpartid/, sub {
752 $partition = "/dev/$part";