]>
git.proxmox.com Git - pve-storage.git/blob - PVE/Diskmanage.pm
7c07e52c7cc46d9d0d62f5357670fc093bb51822
1 package PVE
::Diskmanage
;
13 use PVE
::Tools
qw(extract_param run_command file_get_contents file_read_firstline dir_glob_regex dir_glob_foreach trim);
15 my $SMARTCTL = "/usr/sbin/smartctl";
16 my $ZPOOL = "/sbin/zpool";
17 my $SGDISK = "/sbin/sgdisk";
18 my $PVS = "/sbin/pvs";
19 my $LVS = "/sbin/lvs";
20 my $LSBLK = "/bin/lsblk";
22 sub verify_blockdev_path
{
25 die "missing path" if !$rel_path;
26 my $path = abs_path
($rel_path);
27 die "failed to get absolute path to $rel_path\n" if !$path;
29 die "got unusual device path '$path'\n" if $path !~ m
|^/dev/(.*)$|;
31 $path = "/dev/$1"; # untaint
33 assert_blockdev
($path);
39 my ($dev, $noerr) = @_;
41 if ($dev !~ m
|^/dev/| || !(-b
$dev)) {
42 return undef if $noerr;
43 die "not a valid block device\n";
50 my ($disk, $uuid) = @_;
52 assert_blockdev
($disk);
54 # we should already have checked if it is in use in the api call
55 # but we check again for safety
56 die "disk $disk is already in use\n" if disk_is_used
($disk);
58 my $id = $uuid || 'R';
59 run_command
([$SGDISK, $disk, '-U', $id]);
69 my $disklist = get_disks
($dev, 1);
71 die "'$disk' is not a valid local disk\n" if !defined($disklist->{$dev});
72 return 1 if $disklist->{$dev}->{used
};
78 my ($disk, $healthonly) = @_;
80 assert_blockdev
($disk);
86 if ($disk =~ m!^/dev/(nvme\d+n\d+)$!) {
87 my $info = get_sysdir_info
("/sys/block/$1");
88 $disk = "/dev/".($info->{device
}
89 or die "failed to get nvme controller device for $disk\n");
92 my $cmd = [$SMARTCTL, '-H'];
93 push @$cmd, '-A', '-f', 'brief' if !$healthonly;
97 $returncode = run_command
($cmd, noerr
=> 1, outfunc
=> sub{
100 # ATA SMART attributes, e.g.:
101 # ID# ATTRIBUTE_NAME FLAGS VALUE WORST THRESH FAIL RAW_VALUE
102 # 1 Raw_Read_Error_Rate POSR-K 100 100 000 - 0
104 # SAS and NVME disks, e.g.:
105 # Data Units Written: 5,584,952 [2.85 TB]
106 # Accumulated start-stop cycles: 34
108 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/sas text, allow for decimal values
137 if ($line =~ m/Percentage Used(?: endurance indicator)?:\s*(\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_lsblk_info
() {
160 my $cmd = [$LSBLK, '--json', '-o', 'path,parttype,fstype'];
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
} // [];
178 parttype
=> $_->{parttype
},
179 fstype
=> $_->{fstype
}
186 my $get_devices_by_partuuid = sub {
187 my ($lsblk_info, $uuids, $res) = @_;
189 $res = {} if !defined($res);
191 foreach my $dev (sort keys %{$lsblk_info}) {
192 my $uuid = $lsblk_info->{$dev}->{parttype
};
193 next if !defined($uuid) || !defined($uuids->{$uuid});
194 $res->{$dev} = $uuids->{$uuid};
200 sub get_zfs_devices
{
201 my ($lsblk_info) = @_;
204 return {} if ! -x
$ZPOOL;
206 # use zpool and parttype uuid,
207 # because log and cache do not have
210 run_command
([$ZPOOL, 'list', '-HPLv'], outfunc
=> sub {
213 if ($line =~ m
|^\t([^\t]+)\t|) {
220 # because maybe zfs tools are not installed
224 "6a898cc3-1dd2-11b2-99a6-080020736631" => 1, # apple
225 "516e7cba-6ecf-11d6-8ff8-00022d09712b" => 1, # bsd
229 $res = $get_devices_by_partuuid->($lsblk_info, $uuids, $res);
234 sub get_lvm_devices
{
235 my ($lsblk_info) = @_;
238 run_command
([$PVS, '--noheadings', '--readonly', '-o', 'pv_name'], outfunc
=> sub{
241 if ($line =~ m
|^/dev/|) {
247 # if something goes wrong, we do not want
248 # to give up, but indicate an error has occured
252 "e6d6d379-f507-44c2-a23c-238f2a3df928" => 1,
255 $res = $get_devices_by_partuuid->($lsblk_info, $uuids, $res);
260 sub get_ceph_journals
{
261 my ($lsblk_info) = @_;
265 '45b0969e-9b03-4f30-b4c6-b4b80ceff106' => 1, # journal
266 '30cd0809-c2b2-499c-8879-2d6b78529876' => 2, # db
267 '5ce17fce-4087-4169-b7ff-056cc58473f9' => 3, # wal
268 'cafecafe-9b03-4f30-b4c6-b4b80ceff106' => 4, # block
271 $res = $get_devices_by_partuuid->($lsblk_info, $uuids, $res);
276 # reads the lv_tags and matches them with the devices
277 sub get_ceph_volume_infos
{
280 my $cmd = [ $LVS, '-S', 'lv_name=~^osd-', '-o', 'devices,lv_name,lv_tags',
281 '--noheadings', '--readonly', '--separator', ';' ];
283 run_command
($cmd, outfunc
=> sub {
285 $line =~ s/(?:^\s+)|(?:\s+$)//g; # trim whitespaces
287 my $fields = [ split(';', $line) ];
289 # lvs syntax is /dev/sdX(Y) where Y is the start (which we do not need)
290 my ($dev) = $fields->[0] =~ m
|^(/dev/[a-z
]+[^(]*)|;
291 if ($fields->[1] =~ m
|^osd-
([^-]+)-|) {
293 # $result autovivification is wanted, to not creating empty hashes
294 if (($type eq 'block' || $type eq 'data') && $fields->[2] =~ m/ceph.osd_id=([^,]+)/) {
295 $result->{$dev}->{osdid
} = $1;
296 $result->{$dev}->{bluestore
} = ($type eq 'block');
297 if ($fields->[2] =~ m/ceph\.encrypted=1/) {
298 $result->{$dev}->{encrypted
} = 1;
301 # undef++ becomes '1' (see `perldoc perlop`: Auto-increment)
302 $result->{$dev}->{$type}++;
316 run_command
(['udevadm', 'info', '-p', $dev, '--query', 'all'], outfunc
=> sub {
322 return undef if !$info;
324 return undef if $info !~ m/^E: DEVTYPE=disk$/m;
325 return undef if $info =~ m/^E: ID_CDROM/m;
327 # we use this, because some disks are not simply in /dev
328 # e.g. /dev/cciss/c0d0
329 if ($info =~ m/^E: DEVNAME=(\S+)$/m) {
330 $data->{devpath
} = $1;
332 return if !defined($data->{devpath
});
334 $data->{serial
} = 'unknown';
335 if ($info =~ m/^E: ID_SERIAL_SHORT=(\S+)$/m) {
336 $data->{serial
} = $1;
340 if ($info =~ m/^E: ID_PART_TABLE_TYPE=gpt$/m) {
346 if ($info =~ m/^E: ID_ATA_ROTATION_RATE_RPM=(\d+)$/m) {
350 if ($info =~ m/^E: ID_BUS=usb$/m) {
354 if ($info =~ m/^E: ID_MODEL=(.+)$/m) {
358 $data->{wwn
} = 'unknown';
359 if ($info =~ m/^E: ID_WWN=(.*)$/m) {
363 if ($info =~ m/^E: DEVLINKS=(.+)$/m) {
364 my @devlinks = grep(m
#^/dev/disk/by-id/(ata|scsi|nvme(?!-eui))#, split (/ /, $1));
365 $data->{by_id_link
} = $devlinks[0] if defined($devlinks[0]);
371 sub get_sysdir_size
{
374 my $size = file_read_firstline
("$sysdir/size");
377 # linux always considers sectors to be 512 bytes,
378 # independently of real block size
382 sub get_sysdir_info
{
385 return undef if ! -d
"$sysdir/device";
389 $data->{size
} = get_sysdir_size
($sysdir) or return;
391 # dir/queue/rotational should be 1 for hdd, 0 for ssd
392 $data->{rotational
} = file_read_firstline
("$sysdir/queue/rotational") // -1;
394 $data->{vendor
} = file_read_firstline
("$sysdir/device/vendor") || 'unknown';
395 $data->{model
} = file_read_firstline
("$sysdir/device/model") || 'unknown';
397 if (defined(my $device = readlink("$sysdir/device"))) {
398 # strip directory and untaint:
399 ($data->{device
}) = $device =~ m!([^/]+)$!;
405 sub get_wear_leveling_info
{
406 my ($smartdata) = @_;
407 my $attributes = $smartdata->{attributes
};
409 if (defined($smartdata->{wearout
})) {
410 return $smartdata->{wearout
};
415 # Common register names that represent percentage values of potential
416 # failure indicators used in drivedb.h of smartmontool's. Order matters,
417 # as some drives may have multiple definitions
418 my @wearoutregisters = (
419 "Media_Wearout_Indicator",
421 "Wear_Leveling_Count",
422 "Perc_Write\/Erase_Ct_BC",
423 "Perc_Rated_Life_Remain",
424 "Remaining_Lifetime_Perc",
425 "Percent_Lifetime_Remain",
427 "PCT_Life_Remaining",
428 "Lifetime_Remaining",
429 "Percent_Life_Remaining",
430 "Percent_Lifetime_Used",
431 "Perc_Rated_Life_Used"
434 # Search for S.M.A.R.T. attributes for known register
435 foreach my $register (@wearoutregisters) {
436 last if defined $wearout;
437 foreach my $attr (@$attributes) {
438 next if $attr->{name
} !~ m/$register/;
439 $wearout = $attr->{value
};
450 my $dh = IO
::Dir-
>new ($dir);
453 while (defined(my $tmp = $dh->read)) {
454 next if $tmp eq '.' || $tmp eq '..';
465 if (-l
$sysdir && readlink($sysdir) =~ m
|host
[^/]*/session
[^/]*|) {
474 return $type eq 'ssd' || $type eq 'nvme';
478 my ($disks, $nosmart) = @_;
483 my $mounts = PVE
::ProcFSTools
::parse_proc_mounts
();
485 foreach my $mount (@$mounts) {
486 next if $mount->[0] !~ m
|^/dev/|;
487 $mounted->{abs_path
($mount->[0])} = $mount->[1];
490 my $lsblk_info = get_lsblk_info
();
492 my $journalhash = get_ceph_journals
($lsblk_info);
493 my $ceph_volume_infos = get_ceph_volume_infos
();
495 my $zfshash = get_zfs_devices
($lsblk_info);
497 my $lvmhash = get_lvm_devices
($lsblk_info);
499 my $disk_regex = ".*";
500 if (defined($disks)) {
503 } elsif (ref($disks) ne 'ARRAY') {
504 die "disks is not a string or array reference\n";
506 # we get cciss/c0d0 but need cciss!c0d0
507 $_ =~ s
|cciss
/|cciss
!| for @$disks;
509 $disk_regex = "(?:" . join('|', @$disks) . ")";
512 dir_glob_foreach
('/sys/block', $disk_regex, sub {
514 # whitelisting following devices
515 # hdX: ide block device
516 # sdX: sd block device
517 # vdX: virtual block device
518 # xvdX: xen virtual block device
519 # nvmeXnY: nvme devices
520 # cciss!cXnY: cciss devices
521 return if $dev !~ m/^(h|s|x?v)d[a-z]+$/ &&
522 $dev !~ m/^nvme\d+n\d+$/ &&
523 $dev !~ m/^cciss\!c\d+d\d+$/;
525 my $data = get_udev_info
("/sys/block/$dev");
526 return if !defined($data);
527 my $devpath = $data->{devpath
};
529 my $sysdir = "/sys/block/$dev";
531 # we do not want iscsi devices
532 return if is_iscsi
($sysdir);
534 my $sysdata = get_sysdir_info
($sysdir);
535 return if !defined($sysdata);
537 my $type = 'unknown';
539 if ($sysdata->{rotational
} == 0) {
541 $type = 'nvme' if $dev =~ m/^nvme\d+n\d+$/;
543 } elsif ($sysdata->{rotational
} == 1) {
544 if ($data->{rpm
} != -1) {
546 } elsif ($data->{usb
}) {
552 my $health = 'UNKNOWN';
557 my $smartdata = get_smart_data
($devpath, !is_ssdlike
($type));
558 $health = $smartdata->{health
} if $smartdata->{health
};
560 if (is_ssdlike
($type)) {
561 # if we have an ssd we try to get the wearout indicator
562 my $wearval = get_wear_leveling_info
($smartdata);
563 $wearout = $wearval if defined($wearval);
568 # we replaced cciss/ with cciss! above
569 # but in the result we need cciss/ again
570 # because the caller might want to check the
571 # result again with the original parameter
572 if ($dev =~ m
|^cciss
!|) {
573 $dev =~ s
|^cciss
!|cciss
/|;
576 $disklist->{$dev} = {
577 vendor
=> $sysdata->{vendor
},
578 model
=> $data->{model
} || $sysdata->{model
},
579 size
=> $sysdata->{size
},
580 serial
=> $data->{serial
},
590 my $by_id_link = $data->{by_id_link
};
591 $disklist->{$dev}->{by_id_link
} = $by_id_link if defined($by_id_link);
595 my $osdencrypted = 0;
597 my $journal_count = 0;
601 my $partpath = $devpath;
603 # remove part after last / to
604 # get the base path for the partitions
605 # e.g. from /dev/cciss/c0d0 get /dev/cciss
606 $partpath =~ s/\/[^\/]+$//;
608 my $determine_usage = sub {
609 my ($devpath, $sysdir, $is_partition) = @_;
611 return 'LVM' if $lvmhash->{$devpath};
612 return 'ZFS' if $zfshash->{$devpath};
614 my $info = $lsblk_info->{$devpath} // {};
616 my $parttype = $info->{parttype
};
617 if (defined($parttype)) {
619 if $parttype eq '21686148-6449-6e6f-744e-656564454649';
621 if $parttype eq 'c12a7328-f81f-11d2-ba4b-00a0c93ec93b';
622 return 'ZFS reserved'
623 if $parttype eq '6a945a3b-1dd2-11b2-99a6-080020736631';
626 my $fstype = $info->{fstype
};
627 if (defined($fstype)) {
628 return "${fstype} (mounted)" if $mounted->{$devpath};
631 return 'mounted' if $mounted->{$devpath};
633 return if !$is_partition;
635 # for devices, this check is done explicitly later
636 return 'Device Mapper' if !dir_is_empty
("$sysdir/holders");
641 my $collect_ceph_info = sub {
644 my $ceph_volume = $ceph_volume_infos->{$devpath} or return;
645 $journal_count += $ceph_volume->{journal
} // 0;
646 $db_count += $ceph_volume->{db
} // 0;
647 $wal_count += $ceph_volume->{wal
} // 0;
648 if (defined($ceph_volume->{osdid
})) {
649 $osdid = $ceph_volume->{osdid
};
650 $bluestore = 1 if $ceph_volume->{bluestore
};
651 $osdencrypted = 1 if $ceph_volume->{encrypted
};
654 my $result = { %{$ceph_volume} };
655 $result->{journals
} = delete $result->{journal
}
656 if $result->{journal
};
662 dir_glob_foreach
("$sysdir", "$dev.+", sub {
665 $partitions->{$part} = $collect_ceph_info->("$partpath/$part");
666 my $lvm_based_osd = defined($partitions->{$part});
668 $partitions->{$part}->{devpath
} = "$partpath/$part";
669 $partitions->{$part}->{gpt
} = $data->{gpt
};
670 $partitions->{$part}->{size
} =
671 get_sysdir_size
("$sysdir/$part") // 0;
672 $partitions->{$part}->{used
} =
673 $determine_usage->("$partpath/$part", "$sysdir/$part", 1);
674 $partitions->{$part}->{osdid
} //= -1;
676 # Avoid counting twice (e.g. partition on which the LVM for the
677 # DB OSD resides is present in the $journalhash)
678 return if $lvm_based_osd;
680 # Legacy handling for non-LVM based OSDs
682 if (my $mp = $mounted->{"$partpath/$part"}) {
683 if ($mp =~ m
|^/var/lib
/ceph/osd/ceph-
(\d
+)$|) {
685 $partitions->{$part}->{osdid
} = $osdid;
689 if (my $journal_part = $journalhash->{"$partpath/$part"}) {
690 $journal_count++ if $journal_part == 1;
691 $db_count++ if $journal_part == 2;
692 $wal_count++ if $journal_part == 3;
693 $bluestore = 1 if $journal_part == 4;
695 $partitions->{$part}->{journals
} = 1 if $journal_part == 1;
696 $partitions->{$part}->{db
} = 1 if $journal_part == 2;
697 $partitions->{$part}->{wal
} = 1 if $journal_part == 3;
698 $partitions->{$part}->{bluestore
} = 1 if $journal_part == 4;
702 my $used = $determine_usage->($devpath, $sysdir, 0);
703 foreach my $part (sort keys %{$partitions}) {
704 next if $partitions->{$part}->{used
} eq 'partition';
705 $used //= $partitions->{$part}->{used
};
707 $used //= 'partitions' if scalar(keys %{$partitions});
708 # multipath, software raid, etc.
709 # this check comes in last, to show more specific info
711 $used //= 'Device Mapper' if !dir_is_empty
("$sysdir/holders");
713 $disklist->{$dev}->{used
} = $used if $used;
715 $collect_ceph_info->($devpath);
717 $disklist->{$dev}->{osdid
} = $osdid;
718 $disklist->{$dev}->{journals
} = $journal_count if $journal_count;
719 $disklist->{$dev}->{bluestore
} = $bluestore if $osdid != -1;
720 $disklist->{$dev}->{osdencrypted
} = $osdencrypted if $osdid != -1;
721 $disklist->{$dev}->{db
} = $db_count if $db_count;
722 $disklist->{$dev}->{wal
} = $wal_count if $wal_count;
730 my ($part_path) = @_;
732 my $st = stat($part_path);
734 next if !$st->mode || !S_ISBLK
($st->mode) || !$st->rdev;
735 my $major = PVE
::Tools
::dev_t_major
($st->rdev);
736 my $minor = PVE
::Tools
::dev_t_minor
($st->rdev);
737 my $partnum_path = "/sys/dev/block/$major:$minor/";
741 $partnum = file_read_firstline
("${partnum_path}partition");
743 die "Partition does not exist\n" if !defined($partnum);
745 #untaint and ensure it is a int
746 if ($partnum =~ m/(\d+)/) {
748 die "Partition number $partnum is invalid\n" if $partnum > 128;
750 die "Failed to get partition number\n";
757 my ($part_path) = @_;
759 my ($dev, $block_dev);
760 if ($part_path =~ m
|^/dev/(.*)$|) {
762 my $link = readlink "/sys/class/block/$dev";
763 $block_dev = $1 if $link =~ m
|([^/]*)/$dev$|;
766 die "Can't parse parent device\n" if !defined($block_dev);
767 die "No valid block device\n" if index($dev, $block_dev) == -1;
769 $block_dev = "/dev/$block_dev";
770 die "Block device does not exsists\n" if !(-b
$block_dev);
775 sub locked_disk_action
{
777 my $res = PVE
::Tools
::lock_file
('/run/lock/pve-diskmanage.lck', undef, $sub);
782 sub assert_disk_unused
{
785 die "device '$dev' is already in use\n" if disk_is_used
($dev);
790 sub append_partition
{
791 my ($dev, $size) = @_;
794 $devname =~ s
|^/dev/||;
797 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*?
(\d
+)/, sub {
798 my ($part, $partid) = @_;
800 if ($partid >= $newpartid) {
801 $newpartid = $partid + 1;
805 $size = PVE
::Tools
::convert_size
($size, 'b' => 'mb');
807 run_command
([ $SGDISK, '-n', "$newpartid:0:+${size}M", $dev ],
808 errmsg
=> "error creating partition '$newpartid' on '$dev'");
812 # loop again to detect the real partiton device which does not always follow
813 # a strict $devname$partition scheme like /dev/nvme0n1 -> /dev/nvme0n1p1
814 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*$newpartid/, sub {
817 $partition = "/dev/$part";