]>
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_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
} = ($type eq 'block');
295 if ($fields->[2] =~ m/ceph\.encrypted=1/) {
296 $result->{$dev}->{encrypted
} = 1;
299 # undef++ becomes '1' (see `perldoc perlop`: Auto-increment)
300 $result->{$dev}->{$type}++;
314 run_command
(['udevadm', 'info', '-p', $dev, '--query', 'all'], outfunc
=> sub {
320 return undef if !$info;
322 return undef if $info !~ m/^E: DEVTYPE=disk$/m;
323 return undef if $info =~ m/^E: ID_CDROM/m;
325 # we use this, because some disks are not simply in /dev
326 # e.g. /dev/cciss/c0d0
327 if ($info =~ m/^E: DEVNAME=(\S+)$/m) {
328 $data->{devpath
} = $1;
330 return if !defined($data->{devpath
});
332 $data->{serial
} = 'unknown';
333 if ($info =~ m/^E: ID_SERIAL_SHORT=(\S+)$/m) {
334 $data->{serial
} = $1;
338 if ($info =~ m/^E: ID_PART_TABLE_TYPE=gpt$/m) {
344 if ($info =~ m/^E: ID_ATA_ROTATION_RATE_RPM=(\d+)$/m) {
348 if ($info =~ m/^E: ID_BUS=usb$/m) {
352 if ($info =~ m/^E: ID_MODEL=(.+)$/m) {
356 $data->{wwn
} = 'unknown';
357 if ($info =~ m/^E: ID_WWN=(.*)$/m) {
361 if ($info =~ m/^E: DEVLINKS=(.+)$/m) {
362 my @devlinks = grep(m
#^/dev/disk/by-id/(ata|scsi|nvme(?!-eui))#, split (/ /, $1));
363 $data->{by_id_link
} = $devlinks[0] if defined($devlinks[0]);
369 sub get_sysdir_info
{
372 return undef if ! -d
"$sysdir/device";
376 my $size = file_read_firstline
("$sysdir/size");
377 return undef if !$size;
379 # linux always considers sectors to be 512 bytes,
380 # independently of real block size
381 $data->{size
} = $size * 512;
383 # dir/queue/rotational should be 1 for hdd, 0 for ssd
384 $data->{rotational
} = file_read_firstline
("$sysdir/queue/rotational") // -1;
386 $data->{vendor
} = file_read_firstline
("$sysdir/device/vendor") || 'unknown';
387 $data->{model
} = file_read_firstline
("$sysdir/device/model") || 'unknown';
389 if (defined(my $device = readlink("$sysdir/device"))) {
390 # strip directory and untaint:
391 ($data->{device
}) = $device =~ m!([^/]+)$!;
397 sub get_wear_leveling_info
{
398 my ($smartdata) = @_;
399 my $attributes = $smartdata->{attributes
};
401 if (defined($smartdata->{wearout
})) {
402 return $smartdata->{wearout
};
407 # Common register names that represent percentage values of potential
408 # failure indicators used in drivedb.h of smartmontool's. Order matters,
409 # as some drives may have multiple definitions
410 my @wearoutregisters = (
411 "Media_Wearout_Indicator",
413 "Wear_Leveling_Count",
414 "Perc_Write\/Erase_Ct_BC",
415 "Perc_Rated_Life_Remain",
416 "Remaining_Lifetime_Perc",
417 "Percent_Lifetime_Remain",
419 "PCT_Life_Remaining",
420 "Lifetime_Remaining",
421 "Percent_Life_Remaining",
422 "Percent_Lifetime_Used",
423 "Perc_Rated_Life_Used"
426 # Search for S.M.A.R.T. attributes for known register
427 foreach my $register (@wearoutregisters) {
428 last if defined $wearout;
429 foreach my $attr (@$attributes) {
430 next if $attr->{name
} !~ m/$register/;
431 $wearout = $attr->{value
};
442 my $dh = IO
::Dir-
>new ($dir);
445 while (defined(my $tmp = $dh->read)) {
446 next if $tmp eq '.' || $tmp eq '..';
457 if (-l
$sysdir && readlink($sysdir) =~ m
|host
[^/]*/session
[^/]*|) {
466 return $type eq 'ssd' || $type eq 'nvme';
470 my ($disks, $nosmart) = @_;
475 my $mounts = PVE
::ProcFSTools
::parse_proc_mounts
();
477 foreach my $mount (@$mounts) {
478 next if $mount->[0] !~ m
|^/dev/|;
479 $mounted->{abs_path
($mount->[0])} = $mount->[1];
482 my $parttype_map = get_parttype_info
();
484 my $journalhash = get_ceph_journals
($parttype_map);
485 my $ceph_volume_infos = get_ceph_volume_infos
();
487 my $zfshash = get_zfs_devices
($parttype_map);
489 my $lvmhash = get_lvm_devices
($parttype_map);
491 my $disk_regex = ".*";
492 if (defined($disks)) {
495 } elsif (ref($disks) ne 'ARRAY') {
496 die "disks is not a string or array reference\n";
498 # we get cciss/c0d0 but need cciss!c0d0
499 $_ =~ s
|cciss
/|cciss
!| for @$disks;
501 $disk_regex = "(?:" . join('|', @$disks) . ")";
504 dir_glob_foreach
('/sys/block', $disk_regex, sub {
506 # whitelisting following devices
507 # hdX: ide block device
508 # sdX: sd block device
509 # vdX: virtual block device
510 # xvdX: xen virtual block device
511 # nvmeXnY: nvme devices
512 # cciss!cXnY: cciss devices
513 return if $dev !~ m/^(h|s|x?v)d[a-z]+$/ &&
514 $dev !~ m/^nvme\d+n\d+$/ &&
515 $dev !~ m/^cciss\!c\d+d\d+$/;
517 my $data = get_udev_info
("/sys/block/$dev");
518 return if !defined($data);
519 my $devpath = $data->{devpath
};
521 my $sysdir = "/sys/block/$dev";
523 # we do not want iscsi devices
524 return if is_iscsi
($sysdir);
526 my $sysdata = get_sysdir_info
($sysdir);
527 return if !defined($sysdata);
529 my $type = 'unknown';
531 if ($sysdata->{rotational
} == 0) {
533 $type = 'nvme' if $dev =~ m/^nvme\d+n\d+$/;
535 } elsif ($sysdata->{rotational
} == 1) {
536 if ($data->{rpm
} != -1) {
538 } elsif ($data->{usb
}) {
544 my $health = 'UNKNOWN';
549 my $smartdata = get_smart_data
($devpath, !is_ssdlike
($type));
550 $health = $smartdata->{health
} if $smartdata->{health
};
552 if (is_ssdlike
($type)) {
553 # if we have an ssd we try to get the wearout indicator
554 my $wearval = get_wear_leveling_info
($smartdata);
555 $wearout = $wearval if defined($wearval);
562 $used = 'LVM' if $lvmhash->{$devpath};
564 $used = 'mounted' if $mounted->{$devpath};
566 $used = 'ZFS' if $zfshash->{$devpath};
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 $found_partitions;
603 my $found_mountpoints;
606 my $partpath = $devpath;
608 # remove part after last / to
609 # get the base path for the partitions
610 # e.g. from /dev/cciss/c0d0 get /dev/cciss
611 $partpath =~ s/\/[^\/]+$//;
613 dir_glob_foreach
("$sysdir", "$dev.+", sub {
616 $found_partitions = 1;
618 if (my $mp = $mounted->{"$partpath/$part"}) {
619 $found_mountpoints = 1;
620 if ($mp =~ m
|^/var/lib
/ceph/osd/ceph-
(\d
+)$|) {
625 if ($lvmhash->{"$partpath/$part"}) {
629 if ($zfshash->{"$partpath/$part"}) {
633 if (my $journal_part = $journalhash->{"$partpath/$part"}) {
634 $journal_count++ if $journal_part == 1;
635 $db_count++ if $journal_part == 2;
636 $wal_count++ if $journal_part == 3;
637 $bluestore = 1 if $journal_part == 4;
640 if (!dir_is_empty
("$sysdir/$part/holders") && !$found_lvm) {
645 if (my $ceph_volume = $ceph_volume_infos->{$devpath}) {
646 $journal_count += $ceph_volume->{journal
} // 0;
647 $db_count += $ceph_volume->{db
} // 0;
648 $wal_count += $ceph_volume->{wal
} // 0;
649 if (defined($ceph_volume->{osdid
})) {
650 $osdid = $ceph_volume->{osdid
};
651 $bluestore = 1 if $ceph_volume->{bluestore
};
652 $osdencrypted = 1 if $ceph_volume->{encrypted
};
656 $used = 'mounted' if $found_mountpoints && !$used;
657 $used = 'LVM' if $found_lvm && !$used;
658 $used = 'ZFS' if $found_zfs && !$used;
659 $used = 'Device Mapper' if $found_dm && !$used;
660 $used = 'partitions' if $found_partitions && !$used;
662 # multipath, software raid, etc.
663 # this check comes in last, to show more specific info
665 $used = 'Device Mapper' if !$used && !dir_is_empty
("$sysdir/holders");
667 $disklist->{$dev}->{used
} = $used if $used;
668 $disklist->{$dev}->{osdid
} = $osdid;
669 $disklist->{$dev}->{journals
} = $journal_count if $journal_count;
670 $disklist->{$dev}->{bluestore
} = $bluestore if $osdid != -1;
671 $disklist->{$dev}->{osdencrypted
} = $osdencrypted if $osdid != -1;
672 $disklist->{$dev}->{db
} = $db_count if $db_count;
673 $disklist->{$dev}->{wal
} = $wal_count if $wal_count;
681 my ($part_path) = @_;
683 my $st = stat($part_path);
685 next if !$st->mode || !S_ISBLK
($st->mode) || !$st->rdev;
686 my $major = PVE
::Tools
::dev_t_major
($st->rdev);
687 my $minor = PVE
::Tools
::dev_t_minor
($st->rdev);
688 my $partnum_path = "/sys/dev/block/$major:$minor/";
692 $partnum = file_read_firstline
("${partnum_path}partition");
694 die "Partition does not exist\n" if !defined($partnum);
696 #untaint and ensure it is a int
697 if ($partnum =~ m/(\d+)/) {
699 die "Partition number $partnum is invalid\n" if $partnum > 128;
701 die "Failed to get partition number\n";
708 my ($part_path) = @_;
710 my ($dev, $block_dev);
711 if ($part_path =~ m
|^/dev/(.*)$|) {
713 my $link = readlink "/sys/class/block/$dev";
714 $block_dev = $1 if $link =~ m
|([^/]*)/$dev$|;
717 die "Can't parse parent device\n" if !defined($block_dev);
718 die "No valid block device\n" if index($dev, $block_dev) == -1;
720 $block_dev = "/dev/$block_dev";
721 die "Block device does not exsists\n" if !(-b
$block_dev);
726 sub locked_disk_action
{
728 my $res = PVE
::Tools
::lock_file
('/run/lock/pve-diskmanage.lck', undef, $sub);
733 sub assert_disk_unused
{
736 die "device '$dev' is already in use\n" if disk_is_used
($dev);
741 sub append_partition
{
742 my ($dev, $size) = @_;
745 $devname =~ s
|^/dev/||;
748 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*?
(\d
+)/, sub {
749 my ($part, $partid) = @_;
751 if ($partid >= $newpartid) {
752 $newpartid = $partid + 1;
756 $size = PVE
::Tools
::convert_size
($size, 'b' => 'mb');
758 run_command
([ $SGDISK, '-n', "$newpartid:0:+${size}M", $dev ],
759 errmsg
=> "error creating partition '$newpartid' on '$dev'");
763 # loop again to detect the real partiton device which does not always follow
764 # a strict $devname$partition scheme like /dev/nvme0n1 -> /dev/nvme0n1p1
765 dir_glob_foreach
("/sys/block/$devname", qr
/\Q
$devname\E.*$newpartid/, sub {
768 $partition = "/dev/$part";