]>
git.proxmox.com Git - pve-storage.git/blob - PVE/Diskmanage.pm
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'];
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
} // [];
177 $_->{path
} => { parttype
=> $_->{parttype
} }
183 my $get_devices_by_partuuid = sub {
184 my ($lsblk_info, $uuids, $res) = @_;
186 $res = {} if !defined($res);
188 foreach my $dev (sort keys %{$lsblk_info}) {
189 my $uuid = $lsblk_info->{$dev}->{parttype
};
190 next if !defined($uuid) || !defined($uuids->{$uuid});
191 $res->{$dev} = $uuids->{$uuid};
197 sub get_zfs_devices
{
198 my ($lsblk_info) = @_;
201 return {} if ! -x
$ZPOOL;
203 # use zpool and parttype uuid,
204 # because log and cache do not have
207 run_command
([$ZPOOL, 'list', '-HPLv'], outfunc
=> sub {
210 if ($line =~ m
|^\t([^\t]+)\t|) {
217 # because maybe zfs tools are not installed
221 "6a898cc3-1dd2-11b2-99a6-080020736631" => 1, # apple
222 "516e7cba-6ecf-11d6-8ff8-00022d09712b" => 1, # bsd
226 $res = $get_devices_by_partuuid->($lsblk_info, $uuids, $res);
231 sub get_lvm_devices
{
232 my ($lsblk_info) = @_;
235 run_command
([$PVS, '--noheadings', '--readonly', '-o', 'pv_name'], outfunc
=> sub{
238 if ($line =~ m
|^/dev/|) {
244 # if something goes wrong, we do not want
245 # to give up, but indicate an error has occured
249 "e6d6d379-f507-44c2-a23c-238f2a3df928" => 1,
252 $res = $get_devices_by_partuuid->($lsblk_info, $uuids, $res);
257 sub get_ceph_journals
{
258 my ($lsblk_info) = @_;
262 '45b0969e-9b03-4f30-b4c6-b4b80ceff106' => 1, # journal
263 '30cd0809-c2b2-499c-8879-2d6b78529876' => 2, # db
264 '5ce17fce-4087-4169-b7ff-056cc58473f9' => 3, # wal
265 'cafecafe-9b03-4f30-b4c6-b4b80ceff106' => 4, # block
268 $res = $get_devices_by_partuuid->($lsblk_info, $uuids, $res);
273 # reads the lv_tags and matches them with the devices
274 sub get_ceph_volume_infos
{
277 my $cmd = [ $LVS, '-S', 'lv_name=~^osd-', '-o', 'devices,lv_name,lv_tags',
278 '--noheadings', '--readonly', '--separator', ';' ];
280 run_command
($cmd, outfunc
=> sub {
282 $line =~ s/(?:^\s+)|(?:\s+$)//g; # trim whitespaces
284 my $fields = [ split(';', $line) ];
286 # lvs syntax is /dev/sdX(Y) where Y is the start (which we do not need)
287 my ($dev) = $fields->[0] =~ m
|^(/dev/[a-z
]+)|;
288 if ($fields->[1] =~ m
|^osd-
([^-]+)-|) {
290 # $result autovivification is wanted, to not creating empty hashes
291 if (($type eq 'block' || $type eq 'data') && $fields->[2] =~ m/ceph.osd_id=([^,]+)/) {
292 $result->{$dev}->{osdid
} = $1;
293 $result->{$dev}->{bluestore
} = ($type eq 'block');
294 if ($fields->[2] =~ m/ceph\.encrypted=1/) {
295 $result->{$dev}->{encrypted
} = 1;
298 # undef++ becomes '1' (see `perldoc perlop`: Auto-increment)
299 $result->{$dev}->{$type}++;
313 run_command
(['udevadm', 'info', '-p', $dev, '--query', 'all'], outfunc
=> sub {
319 return undef if !$info;
321 return undef if $info !~ m/^E: DEVTYPE=disk$/m;
322 return undef if $info =~ m/^E: ID_CDROM/m;
324 # we use this, because some disks are not simply in /dev
325 # e.g. /dev/cciss/c0d0
326 if ($info =~ m/^E: DEVNAME=(\S+)$/m) {
327 $data->{devpath
} = $1;
329 return if !defined($data->{devpath
});
331 $data->{serial
} = 'unknown';
332 if ($info =~ m/^E: ID_SERIAL_SHORT=(\S+)$/m) {
333 $data->{serial
} = $1;
337 if ($info =~ m/^E: ID_PART_TABLE_TYPE=gpt$/m) {
343 if ($info =~ m/^E: ID_ATA_ROTATION_RATE_RPM=(\d+)$/m) {
347 if ($info =~ m/^E: ID_BUS=usb$/m) {
351 if ($info =~ m/^E: ID_MODEL=(.+)$/m) {
355 $data->{wwn
} = 'unknown';
356 if ($info =~ m/^E: ID_WWN=(.*)$/m) {
360 if ($info =~ m/^E: DEVLINKS=(.+)$/m) {
361 my @devlinks = grep(m
#^/dev/disk/by-id/(ata|scsi|nvme(?!-eui))#, split (/ /, $1));
362 $data->{by_id_link
} = $devlinks[0] if defined($devlinks[0]);
368 sub get_sysdir_info
{
371 return undef if ! -d
"$sysdir/device";
375 my $size = file_read_firstline
("$sysdir/size");
376 return undef if !$size;
378 # linux always considers sectors to be 512 bytes,
379 # independently of real block size
380 $data->{size
} = $size * 512;
382 # dir/queue/rotational should be 1 for hdd, 0 for ssd
383 $data->{rotational
} = file_read_firstline
("$sysdir/queue/rotational") // -1;
385 $data->{vendor
} = file_read_firstline
("$sysdir/device/vendor") || 'unknown';
386 $data->{model
} = file_read_firstline
("$sysdir/device/model") || 'unknown';
388 if (defined(my $device = readlink("$sysdir/device"))) {
389 # strip directory and untaint:
390 ($data->{device
}) = $device =~ m!([^/]+)$!;
396 sub get_wear_leveling_info
{
397 my ($smartdata) = @_;
398 my $attributes = $smartdata->{attributes
};
400 if (defined($smartdata->{wearout
})) {
401 return $smartdata->{wearout
};
406 # Common register names that represent percentage values of potential
407 # failure indicators used in drivedb.h of smartmontool's. Order matters,
408 # as some drives may have multiple definitions
409 my @wearoutregisters = (
410 "Media_Wearout_Indicator",
412 "Wear_Leveling_Count",
413 "Perc_Write\/Erase_Ct_BC",
414 "Perc_Rated_Life_Remain",
415 "Remaining_Lifetime_Perc",
416 "Percent_Lifetime_Remain",
418 "PCT_Life_Remaining",
419 "Lifetime_Remaining",
420 "Percent_Life_Remaining",
421 "Percent_Lifetime_Used",
422 "Perc_Rated_Life_Used"
425 # Search for S.M.A.R.T. attributes for known register
426 foreach my $register (@wearoutregisters) {
427 last if defined $wearout;
428 foreach my $attr (@$attributes) {
429 next if $attr->{name
} !~ m/$register/;
430 $wearout = $attr->{value
};
441 my $dh = IO
::Dir-
>new ($dir);
444 while (defined(my $tmp = $dh->read)) {
445 next if $tmp eq '.' || $tmp eq '..';
456 if (-l
$sysdir && readlink($sysdir) =~ m
|host
[^/]*/session
[^/]*|) {
465 return $type eq 'ssd' || $type eq 'nvme';
469 my ($disks, $nosmart) = @_;
474 my $mounts = PVE
::ProcFSTools
::parse_proc_mounts
();
476 foreach my $mount (@$mounts) {
477 next if $mount->[0] !~ m
|^/dev/|;
478 $mounted->{abs_path
($mount->[0])} = $mount->[1];
481 my $lsblk_info = get_lsblk_info
();
483 my $journalhash = get_ceph_journals
($lsblk_info);
484 my $ceph_volume_infos = get_ceph_volume_infos
();
486 my $zfshash = get_zfs_devices
($lsblk_info);
488 my $lvmhash = get_lvm_devices
($lsblk_info);
490 my $disk_regex = ".*";
491 if (defined($disks)) {
494 } elsif (ref($disks) ne 'ARRAY') {
495 die "disks is not a string or array reference\n";
497 # we get cciss/c0d0 but need cciss!c0d0
498 $_ =~ s
|cciss
/|cciss
!| for @$disks;
500 $disk_regex = "(?:" . join('|', @$disks) . ")";
503 dir_glob_foreach
('/sys/block', $disk_regex, sub {
505 # whitelisting following devices
506 # hdX: ide block device
507 # sdX: sd block device
508 # vdX: virtual block device
509 # xvdX: xen virtual block device
510 # nvmeXnY: nvme devices
511 # cciss!cXnY: cciss devices
512 return if $dev !~ m/^(h|s|x?v)d[a-z]+$/ &&
513 $dev !~ m/^nvme\d+n\d+$/ &&
514 $dev !~ m/^cciss\!c\d+d\d+$/;
516 my $data = get_udev_info
("/sys/block/$dev");
517 return if !defined($data);
518 my $devpath = $data->{devpath
};
520 my $sysdir = "/sys/block/$dev";
522 # we do not want iscsi devices
523 return if is_iscsi
($sysdir);
525 my $sysdata = get_sysdir_info
($sysdir);
526 return if !defined($sysdata);
528 my $type = 'unknown';
530 if ($sysdata->{rotational
} == 0) {
532 $type = 'nvme' if $dev =~ m/^nvme\d+n\d+$/;
534 } elsif ($sysdata->{rotational
} == 1) {
535 if ($data->{rpm
} != -1) {
537 } elsif ($data->{usb
}) {
543 my $health = 'UNKNOWN';
548 my $smartdata = get_smart_data
($devpath, !is_ssdlike
($type));
549 $health = $smartdata->{health
} if $smartdata->{health
};
551 if (is_ssdlike
($type)) {
552 # if we have an ssd we try to get the wearout indicator
553 my $wearval = get_wear_leveling_info
($smartdata);
554 $wearout = $wearval if defined($wearval);
561 $used = 'LVM' if $lvmhash->{$devpath};
563 $used = 'mounted' if $mounted->{$devpath};
565 $used = 'ZFS' if $zfshash->{$devpath};
567 # we replaced cciss/ with cciss! above
568 # but in the result we need cciss/ again
569 # because the caller might want to check the
570 # result again with the original parameter
571 if ($dev =~ m
|^cciss
!|) {
572 $dev =~ s
|^cciss
!|cciss
/|;
575 $disklist->{$dev} = {
576 vendor
=> $sysdata->{vendor
},
577 model
=> $data->{model
} || $sysdata->{model
},
578 size
=> $sysdata->{size
},
579 serial
=> $data->{serial
},
589 my $by_id_link = $data->{by_id_link
};
590 $disklist->{$dev}->{by_id_link
} = $by_id_link if defined($by_id_link);
594 my $osdencrypted = 0;
596 my $journal_count = 0;
600 my $found_partitions;
602 my $found_mountpoints;
605 my $partpath = $devpath;
607 # remove part after last / to
608 # get the base path for the partitions
609 # e.g. from /dev/cciss/c0d0 get /dev/cciss
610 $partpath =~ s/\/[^\/]+$//;
612 dir_glob_foreach
("$sysdir", "$dev.+", sub {
615 $found_partitions = 1;
617 if (my $mp = $mounted->{"$partpath/$part"}) {
618 $found_mountpoints = 1;
619 if ($mp =~ m
|^/var/lib
/ceph/osd/ceph-
(\d
+)$|) {
624 if ($lvmhash->{"$partpath/$part"}) {
628 if ($zfshash->{"$partpath/$part"}) {
632 if (my $journal_part = $journalhash->{"$partpath/$part"}) {
633 $journal_count++ if $journal_part == 1;
634 $db_count++ if $journal_part == 2;
635 $wal_count++ if $journal_part == 3;
636 $bluestore = 1 if $journal_part == 4;
639 if (!dir_is_empty
("$sysdir/$part/holders") && !$found_lvm) {
644 if (my $ceph_volume = $ceph_volume_infos->{$devpath}) {
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
};
655 $used = 'mounted' if $found_mountpoints && !$used;
656 $used = 'LVM' if $found_lvm && !$used;
657 $used = 'ZFS' if $found_zfs && !$used;
658 $used = 'Device Mapper' if $found_dm && !$used;
659 $used = 'partitions' if $found_partitions && !$used;
661 # multipath, software raid, etc.
662 # this check comes in last, to show more specific info
664 $used = 'Device Mapper' if !$used && !dir_is_empty
("$sysdir/holders");
666 $disklist->{$dev}->{used
} = $used if $used;
667 $disklist->{$dev}->{osdid
} = $osdid;
668 $disklist->{$dev}->{journals
} = $journal_count if $journal_count;
669 $disklist->{$dev}->{bluestore
} = $bluestore if $osdid != -1;
670 $disklist->{$dev}->{osdencrypted
} = $osdencrypted if $osdid != -1;
671 $disklist->{$dev}->{db
} = $db_count if $db_count;
672 $disklist->{$dev}->{wal
} = $wal_count if $wal_count;
680 my ($part_path) = @_;
682 my $st = stat($part_path);
684 next if !$st->mode || !S_ISBLK
($st->mode) || !$st->rdev;
685 my $major = PVE
::Tools
::dev_t_major
($st->rdev);
686 my $minor = PVE
::Tools
::dev_t_minor
($st->rdev);
687 my $partnum_path = "/sys/dev/block/$major:$minor/";
691 $partnum = file_read_firstline
("${partnum_path}partition");
693 die "Partition does not exist\n" if !defined($partnum);
695 #untaint and ensure it is a int
696 if ($partnum =~ m/(\d+)/) {
698 die "Partition number $partnum is invalid\n" if $partnum > 128;
700 die "Failed to get partition number\n";
707 my ($part_path) = @_;
709 my ($dev, $block_dev);
710 if ($part_path =~ m
|^/dev/(.*)$|) {
712 my $link = readlink "/sys/class/block/$dev";
713 $block_dev = $1 if $link =~ m
|([^/]*)/$dev$|;
716 die "Can't parse parent device\n" if !defined($block_dev);
717 die "No valid block device\n" if index($dev, $block_dev) == -1;
719 $block_dev = "/dev/$block_dev";
720 die "Block device does not exsists\n" if !(-b
$block_dev);
725 sub locked_disk_action
{
727 my $res = PVE
::Tools
::lock_file
('/run/lock/pve-diskmanage.lck', undef, $sub);
732 sub assert_disk_unused
{
735 die "device '$dev' is already in use\n" if disk_is_used
($dev);
740 sub append_partition
{
741 my ($dev, $size) = @_;
744 $devname =~ s
|^/dev/||;
747 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*?
(\d
+)/, sub {
748 my ($part, $partid) = @_;
750 if ($partid >= $newpartid) {
751 $newpartid = $partid + 1;
755 $size = PVE
::Tools
::convert_size
($size, 'b' => 'mb');
757 run_command
([ $SGDISK, '-n', "$newpartid:0:+${size}M", $dev ],
758 errmsg
=> "error creating partition '$newpartid' on '$dev'");
762 # loop again to detect the real partiton device which does not always follow
763 # a strict $devname$partition scheme like /dev/nvme0n1 -> /dev/nvme0n1p1
764 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*$newpartid/, sub {
767 $partition = "/dev/$part";