PVE/QemuServer.pm

   1 package PVE::QemuServer;
   2
   3 use strict;
   4 use POSIX;
   5 use IO::Handle;
   6 use IO::Select;
   7 use IO::File;
   8 use IO::Dir;
   9 use IO::Socket::UNIX;
  10 use File::Basename;
  11 use File::Path;
  12 use File::stat;
  13 use Getopt::Long;
  14 use Digest::SHA;
  15 use Fcntl ':flock';
  16 use Cwd 'abs_path';
  17 use IPC::Open3;
  18 use JSON;
  19 use Fcntl;
  20 use PVE::SafeSyslog;
  21 use Storable qw(dclone);
  22 use PVE::Exception qw(raise raise_param_exc);
  23 use PVE::Storage;
  24 use PVE::Tools qw(run_command lock_file file_read_firstline);
  25 use PVE::JSONSchema qw(get_standard_option);
  26 use PVE::Cluster qw(cfs_register_file cfs_read_file cfs_write_file cfs_lock_file);
  27 use PVE::INotify;
  28 use PVE::ProcFSTools;
  29 use PVE::QMPClient;
  30 use PVE::RPCEnvironment;
  31 use Time::HiRes qw(gettimeofday);
  32
  33 my $cpuinfo = PVE::ProcFSTools::read_cpuinfo();
  34
  35 # Note about locking: we use flock on the config file protect
  36 # against concurent actions.
  37 # Aditionaly, we have a 'lock' setting in the config file. This
  38 # can be set to 'migrate', 'backup', 'snapshot' or 'rollback'. Most actions are not
  39 # allowed when such lock is set. But you can ignore this kind of
  40 # lock with the --skiplock flag.
  41
  42 cfs_register_file('/qemu-server/',
  43                   \&parse_vm_config,
  44                   \&write_vm_config);
  45
  46 PVE::JSONSchema::register_standard_option('skiplock', {
  47     description => "Ignore locks - only root is allowed to use this option.",
  48     type => 'boolean',
  49     optional => 1,
  50 });
  51
  52 PVE::JSONSchema::register_standard_option('pve-qm-stateuri', {
  53     description => "Some command save/restore state from this location.",
  54     type => 'string',
  55     maxLength => 128,
  56     optional => 1,
  57 });
  58
  59 PVE::JSONSchema::register_standard_option('pve-snapshot-name', {
  60     description => "The name of the snapshot.",
  61     type => 'string', format => 'pve-configid',
  62     maxLength => 40,
  63 });
  64
  65 #no warnings 'redefine';
  66
  67 unless(defined(&_VZSYSCALLS_H_)) {
  68     eval 'sub _VZSYSCALLS_H_ () {1;}' unless defined(&_VZSYSCALLS_H_);
  69     require 'sys/syscall.ph';
  70     if(defined(&__x86_64__)) {
  71         eval 'sub __NR_fairsched_vcpus () {499;}' unless defined(&__NR_fairsched_vcpus);
  72         eval 'sub __NR_fairsched_mknod () {504;}' unless defined(&__NR_fairsched_mknod);
  73         eval 'sub __NR_fairsched_rmnod () {505;}' unless defined(&__NR_fairsched_rmnod);
  74         eval 'sub __NR_fairsched_chwt () {506;}' unless defined(&__NR_fairsched_chwt);
  75         eval 'sub __NR_fairsched_mvpr () {507;}' unless defined(&__NR_fairsched_mvpr);
  76         eval 'sub __NR_fairsched_rate () {508;}' unless defined(&__NR_fairsched_rate);
  77         eval 'sub __NR_setluid () {501;}' unless defined(&__NR_setluid);
  78         eval 'sub __NR_setublimit () {502;}' unless defined(&__NR_setublimit);
  79     }
  80     elsif(defined( &__i386__) ) {
  81         eval 'sub __NR_fairsched_mknod () {500;}' unless defined(&__NR_fairsched_mknod);
  82         eval 'sub __NR_fairsched_rmnod () {501;}' unless defined(&__NR_fairsched_rmnod);
  83         eval 'sub __NR_fairsched_chwt () {502;}' unless defined(&__NR_fairsched_chwt);
  84         eval 'sub __NR_fairsched_mvpr () {503;}' unless defined(&__NR_fairsched_mvpr);
  85         eval 'sub __NR_fairsched_rate () {504;}' unless defined(&__NR_fairsched_rate);
  86         eval 'sub __NR_fairsched_vcpus () {505;}' unless defined(&__NR_fairsched_vcpus);
  87         eval 'sub __NR_setluid () {511;}' unless defined(&__NR_setluid);
  88         eval 'sub __NR_setublimit () {512;}' unless defined(&__NR_setublimit);
  89     } else {
  90         die("no fairsched syscall for this arch");
  91     }
  92     require 'asm/ioctl.ph';
  93     eval 'sub KVM_GET_API_VERSION () { &_IO(0xAE, 0x);}' unless defined(&KVM_GET_API_VERSION);
  94 }
  95
  96 sub fairsched_mknod {
  97     my ($parent, $weight, $desired) = @_;
  98
  99     return syscall(&__NR_fairsched_mknod, int($parent), int($weight), int($desired));
 100 }
 101
 102 sub fairsched_rmnod {
 103     my ($id) = @_;
 104
 105     return syscall(&__NR_fairsched_rmnod, int($id));
 106 }
 107
 108 sub fairsched_mvpr {
 109     my ($pid, $newid) = @_;
 110
 111     return syscall(&__NR_fairsched_mvpr, int($pid), int($newid));
 112 }
 113
 114 sub fairsched_vcpus {
 115     my ($id, $vcpus) = @_;
 116
 117     return syscall(&__NR_fairsched_vcpus, int($id), int($vcpus));
 118 }
 119
 120 sub fairsched_rate {
 121     my ($id, $op, $rate) = @_;
 122
 123     return syscall(&__NR_fairsched_rate, int($id), int($op), int($rate));
 124 }
 125
 126 use constant FAIRSCHED_SET_RATE  => 0;
 127 use constant FAIRSCHED_DROP_RATE => 1;
 128 use constant FAIRSCHED_GET_RATE  => 2;
 129
 130 sub fairsched_cpulimit {
 131     my ($id, $limit) = @_;
 132
 133     my $cpulim1024 = int($limit * 1024 / 100);
 134     my $op = $cpulim1024 ? FAIRSCHED_SET_RATE : FAIRSCHED_DROP_RATE;
 135
 136     return fairsched_rate($id, $op, $cpulim1024);
 137 }
 138
 139 my $nodename = PVE::INotify::nodename();
 140
 141 mkdir "/etc/pve/nodes/$nodename";
 142 my $confdir = "/etc/pve/nodes/$nodename/qemu-server";
 143 mkdir $confdir;
 144
 145 my $var_run_tmpdir = "/var/run/qemu-server";
 146 mkdir $var_run_tmpdir;
 147
 148 my $lock_dir = "/var/lock/qemu-server";
 149 mkdir $lock_dir;
 150
 151 my $pcisysfs = "/sys/bus/pci";
 152
 153 my $confdesc = {
 154     onboot => {
 155         optional => 1,
 156         type => 'boolean',
 157         description => "Specifies whether a VM will be started during system bootup.",
 158         default => 0,
 159     },
 160     autostart => {
 161         optional => 1,
 162         type => 'boolean',
 163         description => "Automatic restart after crash (currently ignored).",
 164         default => 0,
 165     },
 166     hotplug => {
 167         optional => 1,
 168         type => 'boolean',
 169         description => "Activate hotplug for disk and network device",
 170         default => 0,
 171     },
 172     reboot => {
 173         optional => 1,
 174         type => 'boolean',
 175         description => "Allow reboot. If set to '0' the VM exit on reboot.",
 176         default => 1,
 177     },
 178     lock => {
 179         optional => 1,
 180         type => 'string',
 181         description => "Lock/unlock the VM.",
 182         enum => [qw(migrate backup snapshot rollback)],
 183     },
 184     cpulimit => {
 185         optional => 1,
 186         type => 'integer',
 187         description => "Limit of CPU usage in per cent. Note if the computer has 2 CPUs, it has total of 200% CPU time. Value '0' indicates no CPU limit.\n\nNOTE: This option is currently ignored.",
 188         minimum => 0,
 189         default => 0,
 190     },
 191     cpuunits => {
 192         optional => 1,
 193         type => 'integer',
 194         description => "CPU weight for a VM. Argument is used in the kernel fair scheduler. The larger the number is, the more CPU time this VM gets. Number is relative to weights of all the other running VMs.\n\nNOTE: You can disable fair-scheduler configuration by setting this to 0.",
 195         minimum => 0,
 196         maximum => 500000,
 197         default => 1000,
 198     },
 199     memory => {
 200         optional => 1,
 201         type => 'integer',
 202         description => "Amount of RAM for the VM in MB. This is the maximum available memory when you use the balloon device.",
 203         minimum => 16,
 204         default => 512,
 205     },
 206     balloon => {
 207         optional => 1,
 208         type => 'integer',
 209         description => "Amount of target RAM for the VM in MB. Using zero disables the ballon driver.",
 210         minimum => 0,
 211     },
 212     shares => {
 213         optional => 1,
 214         type => 'integer',
 215         description => "Amount of memory shares for auto-ballooning. The larger the number is, the more memory this VM gets. Number is relative to weights of all other running VMs. Using zero disables auto-ballooning",
 216         minimum => 0,
 217         maximum => 50000,
 218         default => 1000,
 219     },
 220     keyboard => {
 221         optional => 1,
 222         type => 'string',
 223         description => "Keybord layout for vnc server. Default is read from the datacenter configuration file.",
 224         enum => PVE::Tools::kvmkeymaplist(),
 225         default => 'en-us',
 226     },
 227     name => {
 228         optional => 1,
 229         type => 'string', format => 'dns-name',
 230         description => "Set a name for the VM. Only used on the configuration web interface.",
 231     },
 232     scsihw => {
 233         optional => 1,
 234         type => 'string',
 235         description => "scsi controller model",
 236         enum => [qw(lsi virtio-scsi-pci megasas)],
 237         default => 'lsi',
 238     },
 239     description => {
 240         optional => 1,
 241         type => 'string',
 242         description => "Description for the VM. Only used on the configuration web interface. This is saved as comment inside the configuration file.",
 243     },
 244     ostype => {
 245         optional => 1,
 246         type => 'string',
 247         enum => [qw(other wxp w2k w2k3 w2k8 wvista win7 win8 l24 l26)],
 248         description => <<EODESC,
 249 Used to enable special optimization/features for specific
 250 operating systems:
 251
 252 other  => unspecified OS
 253 wxp    => Microsoft Windows XP
 254 w2k    => Microsoft Windows 2000
 255 w2k3   => Microsoft Windows 2003
 256 w2k8   => Microsoft Windows 2008
 257 wvista => Microsoft Windows Vista
 258 win7   => Microsoft Windows 7
 259 win8   => Microsoft Windows 8/2012
 260 l24    => Linux 2.4 Kernel
 261 l26    => Linux 2.6/3.X Kernel
 262
 263 other|l24|l26                       ... no special behaviour
 264 wxp|w2k|w2k3|w2k8|wvista|win7|win8  ... use --localtime switch
 265 EODESC
 266     },
 267     boot => {
 268         optional => 1,
 269         type => 'string',
 270         description => "Boot on floppy (a), hard disk (c), CD-ROM (d), or network (n).",
 271         pattern => '[acdn]{1,4}',
 272         default => 'cdn',
 273     },
 274     bootdisk => {
 275         optional => 1,
 276         type => 'string', format => 'pve-qm-bootdisk',
 277         description => "Enable booting from specified disk.",
 278         pattern => '(ide|sata|scsi|virtio)\d+',
 279     },
 280     smp => {
 281         optional => 1,
 282         type => 'integer',
 283         description => "The number of CPUs. Please use option -sockets instead.",
 284         minimum => 1,
 285         default => 1,
 286     },
 287     sockets => {
 288         optional => 1,
 289         type => 'integer',
 290         description => "The number of CPU sockets.",
 291         minimum => 1,
 292         default => 1,
 293     },
 294     cores => {
 295         optional => 1,
 296         type => 'integer',
 297         description => "The number of cores per socket.",
 298         minimum => 1,
 299         default => 1,
 300     },
 301     acpi => {
 302         optional => 1,
 303         type => 'boolean',
 304         description => "Enable/disable ACPI.",
 305         default => 1,
 306     },
 307     agent => {
 308         optional => 1,
 309         type => 'boolean',
 310         description => "Enable/disable Qemu GuestAgent.",
 311         default => 0,
 312     },
 313     kvm => {
 314         optional => 1,
 315         type => 'boolean',
 316         description => "Enable/disable KVM hardware virtualization.",
 317         default => 1,
 318     },
 319     tdf => {
 320         optional => 1,
 321         type => 'boolean',
 322         description => "Enable/disable time drift fix.",
 323         default => 0,
 324     },
 325     localtime => {
 326         optional => 1,
 327         type => 'boolean',
 328         description => "Set the real time clock to local time. This is enabled by default if ostype indicates a Microsoft OS.",
 329     },
 330     freeze => {
 331         optional => 1,
 332         type => 'boolean',
 333         description => "Freeze CPU at startup (use 'c' monitor command to start execution).",
 334     },
 335     vga => {
 336         optional => 1,
 337         type => 'string',
 338         description => "Select VGA type. If you want to use high resolution modes (>= 1280x1024x16) then you should use option 'std' or 'vmware'. Default is 'std' for win8/win7/w2k8, and 'cirrur' for other OS types",
 339         enum => [qw(std cirrus vmware)],
 340     },
 341     watchdog => {
 342         optional => 1,
 343         type => 'string', format => 'pve-qm-watchdog',
 344         typetext => '[[model=]i6300esb|ib700] [,[action=]reset|shutdown|poweroff|pause|debug|none]',
 345         description => "Create a virtual hardware watchdog device.  Once enabled (by a guest action), the watchdog must be periodically polled by an agent inside the guest or else the guest will be restarted (or execute the action specified)",
 346     },
 347     startdate => {
 348         optional => 1,
 349         type => 'string',
 350         typetext => "(now | YYYY-MM-DD | YYYY-MM-DDTHH:MM:SS)",
 351         description => "Set the initial date of the real time clock. Valid format for date are: 'now' or '2006-06-17T16:01:21' or '2006-06-17'.",
 352         pattern => '(now|\d{4}-\d{1,2}-\d{1,2}(T\d{1,2}:\d{1,2}:\d{1,2})?)',
 353         default => 'now',
 354     },
 355     startup => {
 356         optional => 1,
 357         type => 'string', format => 'pve-qm-startup',
 358         typetext => '[[order=]\d+] [,up=\d+] [,down=\d+] ',
 359         description => "Startup and shutdown behavior. Order is a non-negative number defining the general startup order. Shutdown in done with reverse ordering. Additionally you can set the 'up' or 'down' delay in seconds, which specifies a delay to wait before the next VM is started or stopped.",
 360     },
 361     args => {
 362         optional => 1,
 363         type => 'string',
 364         description => <<EODESCR,
 365 Note: this option is for experts only. It allows you to pass arbitrary arguments to kvm, for example:
 366
 367 args: -no-reboot -no-hpet
 368 EODESCR
 369     },
 370     tablet => {
 371         optional => 1,
 372         type => 'boolean',
 373         default => 1,
 374         description => "Enable/disable the usb tablet device. This device is usually needed to allow absolute mouse positioning. Else the mouse runs out of sync with normal vnc clients. If you're running lots of console-only guests on one host, you may consider disabling this to save some context switches.",
 375     },
 376     migrate_speed => {
 377         optional => 1,
 378         type => 'integer',
 379         description => "Set maximum speed (in MB/s) for migrations. Value 0 is no limit.",
 380         minimum => 0,
 381         default => 0,
 382     },
 383     migrate_downtime => {
 384         optional => 1,
 385         type => 'integer',
 386         description => "Set maximum tolerated downtime (in seconds) for migrations.",
 387         minimum => 0,
 388         default => 1,
 389     },
 390     cdrom => {
 391         optional => 1,
 392         type => 'string', format => 'pve-qm-drive',
 393         typetext => 'volume',
 394         description => "This is an alias for option -ide2",
 395     },
 396     cpu => {
 397         optional => 1,
 398         description => "Emulated CPU type.",
 399         type => 'string',
 400         enum => [ qw(486 athlon pentium pentium2 pentium3 coreduo core2duo kvm32 kvm64 qemu32 qemu64 phenom Conroe Penryn Nehalem Westmere SandyBridge Haswell Opteron_G1 Opteron_G2 Opteron_G3 Opteron_G4 Opteron_G5 host) ],
 401         default => 'qemu64',
 402     },
 403     parent => get_standard_option('pve-snapshot-name', {
 404         optional => 1,
 405         description => "Parent snapshot name. This is used internally, and should not be modified.",
 406     }),
 407     snaptime => {
 408         optional => 1,
 409         description => "Timestamp for snapshots.",
 410         type => 'integer',
 411         minimum => 0,
 412     },
 413     vmstate => {
 414         optional => 1,
 415         type => 'string', format => 'pve-volume-id',
 416         description => "Reference to a volume which stores the VM state. This is used internally for snapshots.",
 417     },
 418 };
 419
 420 # what about other qemu settings ?
 421 #cpu => 'string',
 422 #machine => 'string',
 423 #fda => 'file',
 424 #fdb => 'file',
 425 #mtdblock => 'file',
 426 #sd => 'file',
 427 #pflash => 'file',
 428 #snapshot => 'bool',
 429 #bootp => 'file',
 430 ##tftp => 'dir',
 431 ##smb => 'dir',
 432 #kernel => 'file',
 433 #append => 'string',
 434 #initrd => 'file',
 435 ##soundhw => 'string',
 436
 437 while (my ($k, $v) = each %$confdesc) {
 438     PVE::JSONSchema::register_standard_option("pve-qm-$k", $v);
 439 }
 440
 441 my $MAX_IDE_DISKS = 4;
 442 my $MAX_SCSI_DISKS = 14;
 443 my $MAX_VIRTIO_DISKS = 16;
 444 my $MAX_SATA_DISKS = 6;
 445 my $MAX_USB_DEVICES = 5;
 446 my $MAX_NETS = 32;
 447 my $MAX_UNUSED_DISKS = 8;
 448 my $MAX_HOSTPCI_DEVICES = 2;
 449 my $MAX_SERIAL_PORTS = 4;
 450 my $MAX_PARALLEL_PORTS = 3;
 451
 452 my $nic_model_list = ['rtl8139', 'ne2k_pci', 'e1000',  'pcnet',  'virtio',
 453                       'ne2k_isa', 'i82551', 'i82557b', 'i82559er'];
 454 my $nic_model_list_txt = join(' ', sort @$nic_model_list);
 455
 456 my $netdesc = {
 457     optional => 1,
 458     type => 'string', format => 'pve-qm-net',
 459     typetext => "MODEL=XX:XX:XX:XX:XX:XX [,bridge=<dev>][,rate=<mbps>][,tag=<vlanid>]",
 460     description => <<EODESCR,
 461 Specify network devices.
 462
 463 MODEL is one of: $nic_model_list_txt
 464
 465 XX:XX:XX:XX:XX:XX should be an unique MAC address. This is
 466 automatically generated if not specified.
 467
 468 The bridge parameter can be used to automatically add the interface to a bridge device. The Proxmox VE standard bridge is called 'vmbr0'.
 469
 470 Option 'rate' is used to limit traffic bandwidth from and to this interface. It is specified as floating point number, unit is 'Megabytes per second'.
 471
 472 If you specify no bridge, we create a kvm 'user' (NATed) network device, which provides DHCP and DNS services. The following addresses are used:
 473
 474 10.0.2.2   Gateway
 475 10.0.2.3   DNS Server
 476 10.0.2.4   SMB Server
 477
 478 The DHCP server assign addresses to the guest starting from 10.0.2.15.
 479
 480 EODESCR
 481 };
 482 PVE::JSONSchema::register_standard_option("pve-qm-net", $netdesc);
 483
 484 for (my $i = 0; $i < $MAX_NETS; $i++)  {
 485     $confdesc->{"net$i"} = $netdesc;
 486 }
 487
 488 my $drivename_hash;
 489
 490 my $idedesc = {
 491     optional => 1,
 492     type => 'string', format => 'pve-qm-drive',
 493     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads]',
 494     description => "Use volume as IDE hard disk or CD-ROM (n is 0 to " .($MAX_IDE_DISKS -1) . ").",
 495 };
 496 PVE::JSONSchema::register_standard_option("pve-qm-ide", $idedesc);
 497
 498 my $scsidesc = {
 499     optional => 1,
 500     type => 'string', format => 'pve-qm-drive',
 501     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads]',
 502     description => "Use volume as SCSI hard disk or CD-ROM (n is 0 to " . ($MAX_SCSI_DISKS - 1) . ").",
 503 };
 504 PVE::JSONSchema::register_standard_option("pve-qm-scsi", $scsidesc);
 505
 506 my $satadesc = {
 507     optional => 1,
 508     type => 'string', format => 'pve-qm-drive',
 509     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads]',
 510     description => "Use volume as SATA hard disk or CD-ROM (n is 0 to " . ($MAX_SATA_DISKS - 1). ").",
 511 };
 512 PVE::JSONSchema::register_standard_option("pve-qm-sata", $satadesc);
 513
 514 my $virtiodesc = {
 515     optional => 1,
 516     type => 'string', format => 'pve-qm-drive',
 517     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads]',
 518     description => "Use volume as VIRTIO hard disk (n is 0 to " . ($MAX_VIRTIO_DISKS - 1) . ").",
 519 };
 520 PVE::JSONSchema::register_standard_option("pve-qm-virtio", $virtiodesc);
 521
 522 my $usbdesc = {
 523     optional => 1,
 524     type => 'string', format => 'pve-qm-usb-device',
 525     typetext => 'host=HOSTUSBDEVICE',
 526     description => <<EODESCR,
 527 Configure an USB device (n is 0 to 4). This can be used to
 528 pass-through usb devices to the guest. HOSTUSBDEVICE syntax is:
 529
 530 'bus-port(.port)*' (decimal numbers) or
 531 'vendor_id:product_id' (hexadeciaml numbers)
 532
 533 You can use the 'lsusb -t' command to list existing usb devices.
 534
 535 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 536
 537 EODESCR
 538 };
 539 PVE::JSONSchema::register_standard_option("pve-qm-usb", $usbdesc);
 540
 541 my $hostpcidesc = {
 542         optional => 1,
 543         type => 'string', format => 'pve-qm-hostpci',
 544         typetext => "HOSTPCIDEVICE",
 545         description => <<EODESCR,
 546 Map host pci devices. HOSTPCIDEVICE syntax is:
 547
 548 'bus:dev.func' (hexadecimal numbers)
 549
 550 You can us the 'lspci' command to list existing pci devices.
 551
 552 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 553
 554 Experimental: user reported problems with this option.
 555 EODESCR
 556 };
 557 PVE::JSONSchema::register_standard_option("pve-qm-hostpci", $hostpcidesc);
 558
 559 my $serialdesc = {
 560         optional => 1,
 561         type => 'string',
 562         pattern => '/dev/ttyS\d+',
 563         description =>  <<EODESCR,
 564 Map host serial devices (n is 0 to 3).
 565
 566 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 567
 568 Experimental: user reported problems with this option.
 569 EODESCR
 570 };
 571
 572 my $paralleldesc= {
 573         optional => 1,
 574         type => 'string',
 575         pattern => '/dev/parport\d+',
 576         description =>  <<EODESCR,
 577 Map host parallel devices (n is 0 to 2).
 578
 579 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 580
 581 Experimental: user reported problems with this option.
 582 EODESCR
 583 };
 584
 585 for (my $i = 0; $i < $MAX_PARALLEL_PORTS; $i++)  {
 586     $confdesc->{"parallel$i"} = $paralleldesc;
 587 }
 588
 589 for (my $i = 0; $i < $MAX_SERIAL_PORTS; $i++)  {
 590     $confdesc->{"serial$i"} = $serialdesc;
 591 }
 592
 593 for (my $i = 0; $i < $MAX_HOSTPCI_DEVICES; $i++)  {
 594     $confdesc->{"hostpci$i"} = $hostpcidesc;
 595 }
 596
 597 for (my $i = 0; $i < $MAX_IDE_DISKS; $i++)  {
 598     $drivename_hash->{"ide$i"} = 1;
 599     $confdesc->{"ide$i"} = $idedesc;
 600 }
 601
 602 for (my $i = 0; $i < $MAX_SATA_DISKS; $i++)  {
 603     $drivename_hash->{"sata$i"} = 1;
 604     $confdesc->{"sata$i"} = $satadesc;
 605 }
 606
 607 for (my $i = 0; $i < $MAX_SCSI_DISKS; $i++)  {
 608     $drivename_hash->{"scsi$i"} = 1;
 609     $confdesc->{"scsi$i"} = $scsidesc ;
 610 }
 611
 612 for (my $i = 0; $i < $MAX_VIRTIO_DISKS; $i++)  {
 613     $drivename_hash->{"virtio$i"} = 1;
 614     $confdesc->{"virtio$i"} = $virtiodesc;
 615 }
 616
 617 for (my $i = 0; $i < $MAX_USB_DEVICES; $i++)  {
 618     $confdesc->{"usb$i"} = $usbdesc;
 619 }
 620
 621 my $unuseddesc = {
 622     optional => 1,
 623     type => 'string', format => 'pve-volume-id',
 624     description => "Reference to unused volumes.",
 625 };
 626
 627 for (my $i = 0; $i < $MAX_UNUSED_DISKS; $i++)  {
 628     $confdesc->{"unused$i"} = $unuseddesc;
 629 }
 630
 631 my $kvm_api_version = 0;
 632
 633 sub kvm_version {
 634
 635     return $kvm_api_version if $kvm_api_version;
 636
 637     my $fh = IO::File->new("</dev/kvm") ||
 638         return 0;
 639
 640     if (my $v = $fh->ioctl(KVM_GET_API_VERSION(), 0)) {
 641         $kvm_api_version = $v;
 642     }
 643
 644     $fh->close();
 645
 646     return  $kvm_api_version;
 647 }
 648
 649 my $kvm_user_version;
 650
 651 sub kvm_user_version {
 652
 653     return $kvm_user_version if $kvm_user_version;
 654
 655     $kvm_user_version = 'unknown';
 656
 657     my $tmp = `kvm -help 2>/dev/null`;
 658
 659     if ($tmp =~ m/^QEMU( PC)? emulator version (\d+\.\d+(\.\d+)?)[,\s]/) {
 660         $kvm_user_version = $2;
 661     }
 662
 663     return $kvm_user_version;
 664
 665 }
 666
 667 my $kernel_has_vhost_net = -c '/dev/vhost-net';
 668
 669 sub disknames {
 670     # order is important - used to autoselect boot disk
 671     return ((map { "ide$_" } (0 .. ($MAX_IDE_DISKS - 1))),
 672             (map { "scsi$_" } (0 .. ($MAX_SCSI_DISKS - 1))),
 673             (map { "virtio$_" } (0 .. ($MAX_VIRTIO_DISKS - 1))),
 674             (map { "sata$_" } (0 .. ($MAX_SATA_DISKS - 1))));
 675 }
 676
 677 sub valid_drivename {
 678     my $dev = shift;
 679
 680     return defined($drivename_hash->{$dev});
 681 }
 682
 683 sub option_exists {
 684     my $key = shift;
 685     return defined($confdesc->{$key});
 686 }
 687
 688 sub nic_models {
 689     return $nic_model_list;
 690 }
 691
 692 sub os_list_description {
 693
 694     return {
 695         other => 'Other',
 696         wxp => 'Windows XP',
 697         w2k => 'Windows 2000',
 698         w2k3 =>, 'Windows 2003',
 699         w2k8 => 'Windows 2008',
 700         wvista => 'Windows Vista',
 701         win7 => 'Windows 7',
 702         win8 => 'Windows 8/2012',
 703         l24 => 'Linux 2.4',
 704         l26 => 'Linux 2.6',
 705     };
 706 }
 707
 708 my $cdrom_path;
 709
 710 sub get_cdrom_path {
 711
 712     return  $cdrom_path if $cdrom_path;
 713
 714     return $cdrom_path = "/dev/cdrom" if -l "/dev/cdrom";
 715     return $cdrom_path = "/dev/cdrom1" if -l "/dev/cdrom1";
 716     return $cdrom_path = "/dev/cdrom2" if -l "/dev/cdrom2";
 717 }
 718
 719 sub get_iso_path {
 720     my ($storecfg, $vmid, $cdrom) = @_;
 721
 722     if ($cdrom eq 'cdrom') {
 723         return get_cdrom_path();
 724     } elsif ($cdrom eq 'none') {
 725         return '';
 726     } elsif ($cdrom =~ m|^/|) {
 727         return $cdrom;
 728     } else {
 729         return PVE::Storage::path($storecfg, $cdrom);
 730     }
 731 }
 732
 733 # try to convert old style file names to volume IDs
 734 sub filename_to_volume_id {
 735     my ($vmid, $file, $media) = @_;
 736
 737     if (!($file eq 'none' || $file eq 'cdrom' ||
 738           $file =~ m|^/dev/.+| || $file =~ m/^([^:]+):(.+)$/)) {
 739
 740         return undef if $file =~ m|/|;
 741
 742         if ($media && $media eq 'cdrom') {
 743             $file = "local:iso/$file";
 744         } else {
 745             $file = "local:$vmid/$file";
 746         }
 747     }
 748
 749     return $file;
 750 }
 751
 752 sub verify_media_type {
 753     my ($opt, $vtype, $media) = @_;
 754
 755     return if !$media;
 756
 757     my $etype;
 758     if ($media eq 'disk') {
 759         $etype = 'images';
 760     } elsif ($media eq 'cdrom') {
 761         $etype = 'iso';
 762     } else {
 763         die "internal error";
 764     }
 765
 766     return if ($vtype eq $etype);
 767
 768     raise_param_exc({ $opt => "unexpected media type ($vtype != $etype)" });
 769 }
 770
 771 sub cleanup_drive_path {
 772     my ($opt, $storecfg, $drive) = @_;
 773
 774     # try to convert filesystem paths to volume IDs
 775
 776     if (($drive->{file} !~ m/^(cdrom|none)$/) &&
 777         ($drive->{file} !~ m|^/dev/.+|) &&
 778         ($drive->{file} !~ m/^([^:]+):(.+)$/) &&
 779         ($drive->{file} !~ m/^\d+$/)) {
 780         my ($vtype, $volid) = PVE::Storage::path_to_volume_id($storecfg, $drive->{file});
 781         raise_param_exc({ $opt => "unable to associate path '$drive->{file}' to any storage"}) if !$vtype;
 782         $drive->{media} = 'cdrom' if !$drive->{media} && $vtype eq 'iso';
 783         verify_media_type($opt, $vtype, $drive->{media});
 784         $drive->{file} = $volid;
 785     }
 786
 787     $drive->{media} = 'cdrom' if !$drive->{media} && $drive->{file} =~ m/^(cdrom|none)$/;
 788 }
 789
 790 sub create_conf_nolock {
 791     my ($vmid, $settings) = @_;
 792
 793     my $filename = config_file($vmid);
 794
 795     die "configuration file '$filename' already exists\n" if -f $filename;
 796
 797     my $defaults = load_defaults();
 798
 799     $settings->{name} = "vm$vmid" if !$settings->{name};
 800     $settings->{memory} = $defaults->{memory} if !$settings->{memory};
 801
 802     my $data = '';
 803     foreach my $opt (keys %$settings) {
 804         next if !$confdesc->{$opt};
 805
 806         my $value = $settings->{$opt};
 807         next if !$value;
 808
 809         $data .= "$opt: $value\n";
 810     }
 811
 812     PVE::Tools::file_set_contents($filename, $data);
 813 }
 814
 815 my $parse_size = sub {
 816     my ($value) = @_;
 817
 818     return undef if $value !~ m/^(\d+(\.\d+)?)([KMG])?$/;
 819     my ($size, $unit) = ($1, $3);
 820     if ($unit) {
 821         if ($unit eq 'K') {
 822             $size = $size * 1024;
 823         } elsif ($unit eq 'M') {
 824             $size = $size * 1024 * 1024;
 825         } elsif ($unit eq 'G') {
 826             $size = $size * 1024 * 1024 * 1024;
 827         }
 828     }
 829     return int($size);
 830 };
 831
 832 my $format_size = sub {
 833     my ($size) = @_;
 834
 835     $size = int($size);
 836
 837     my $kb = int($size/1024);
 838     return $size if $kb*1024 != $size;
 839
 840     my $mb = int($kb/1024);
 841     return "${kb}K" if $mb*1024 != $kb;
 842
 843     my $gb = int($mb/1024);
 844     return "${mb}M" if $gb*1024 != $mb;
 845
 846     return "${gb}G";
 847 };
 848
 849 # ideX = [volume=]volume-id[,media=d][,cyls=c,heads=h,secs=s[,trans=t]]
 850 #        [,snapshot=on|off][,cache=on|off][,format=f][,backup=yes|no]
 851 #        [,rerror=ignore|report|stop][,werror=enospc|ignore|report|stop]
 852 #        [,aio=native|threads]
 853
 854 sub parse_drive {
 855     my ($key, $data) = @_;
 856
 857     my $res = {};
 858
 859     # $key may be undefined - used to verify JSON parameters
 860     if (!defined($key)) {
 861         $res->{interface} = 'unknown'; # should not harm when used to verify parameters
 862         $res->{index} = 0;
 863     } elsif ($key =~ m/^([^\d]+)(\d+)$/) {
 864         $res->{interface} = $1;
 865         $res->{index} = $2;
 866     } else {
 867         return undef;
 868     }
 869
 870     foreach my $p (split (/,/, $data)) {
 871         next if $p =~ m/^\s*$/;
 872
 873         if ($p =~ m/^(file|volume|cyls|heads|secs|trans|media|snapshot|cache|format|rerror|werror|backup|aio|bps|mbps|bps_rd|mbps_rd|bps_wr|mbps_wr|iops|iops_rd|iops_wr|size)=(.+)$/) {
 874             my ($k, $v) = ($1, $2);
 875
 876             $k = 'file' if $k eq 'volume';
 877
 878             return undef if defined $res->{$k};
 879
 880             if ($k eq 'bps' || $k eq 'bps_rd' || $k eq 'bps_wr') {
 881                 return undef if !$v || $v !~ m/^\d+/;
 882                 $k = "m$k";
 883                 $v = sprintf("%.3f", $v / (1024*1024));
 884             }
 885             $res->{$k} = $v;
 886         } else {
 887             if (!$res->{file} && $p !~ m/=/) {
 888                 $res->{file} = $p;
 889             } else {
 890                 return undef;
 891             }
 892         }
 893     }
 894
 895     return undef if !$res->{file};
 896
 897     return undef if $res->{cache} &&
 898         $res->{cache} !~ m/^(off|none|writethrough|writeback|unsafe|directsync)$/;
 899     return undef if $res->{snapshot} && $res->{snapshot} !~ m/^(on|off)$/;
 900     return undef if $res->{cyls} && $res->{cyls} !~ m/^\d+$/;
 901     return undef if $res->{heads} && $res->{heads} !~ m/^\d+$/;
 902     return undef if $res->{secs} && $res->{secs} !~ m/^\d+$/;
 903     return undef if $res->{media} && $res->{media} !~ m/^(disk|cdrom)$/;
 904     return undef if $res->{trans} && $res->{trans} !~ m/^(none|lba|auto)$/;
 905     return undef if $res->{format} && $res->{format} !~ m/^(raw|cow|qcow|qcow2|vmdk|cloop)$/;
 906     return undef if $res->{rerror} && $res->{rerror} !~ m/^(ignore|report|stop)$/;
 907     return undef if $res->{werror} && $res->{werror} !~ m/^(enospc|ignore|report|stop)$/;
 908     return undef if $res->{backup} && $res->{backup} !~ m/^(yes|no)$/;
 909     return undef if $res->{aio} && $res->{aio} !~ m/^(native|threads)$/;
 910
 911
 912     return undef if $res->{mbps_rd} && $res->{mbps};
 913     return undef if $res->{mbps_wr} && $res->{mbps};
 914
 915     return undef if $res->{mbps} && $res->{mbps} !~ m/^\d+(\.\d+)?$/;
 916     return undef if $res->{mbps_rd} && $res->{mbps_rd} !~ m/^\d+(\.\d+)?$/;
 917     return undef if $res->{mbps_wr} && $res->{mbps_wr} !~ m/^\d+(\.\d+)?$/;
 918
 919     return undef if $res->{iops_rd} && $res->{iops};
 920     return undef if $res->{iops_wr} && $res->{iops};
 921     return undef if $res->{iops} && $res->{iops} !~ m/^\d+$/;
 922     return undef if $res->{iops_rd} && $res->{iops_rd} !~ m/^\d+$/;
 923     return undef if $res->{iops_wr} && $res->{iops_wr} !~ m/^\d+$/;
 924
 925
 926     if ($res->{size}) {
 927         return undef if !defined($res->{size} = &$parse_size($res->{size}));
 928     }
 929
 930     if ($res->{media} && ($res->{media} eq 'cdrom')) {
 931         return undef if $res->{snapshot} || $res->{trans} || $res->{format};
 932         return undef if $res->{heads} || $res->{secs} || $res->{cyls};
 933         return undef if $res->{interface} eq 'virtio';
 934     }
 935
 936     # rerror does not work with scsi drives
 937     if ($res->{rerror}) {
 938         return undef if $res->{interface} eq 'scsi';
 939     }
 940
 941     return $res;
 942 }
 943
 944 my @qemu_drive_options = qw(heads secs cyls trans media format cache snapshot rerror werror aio iops iops_rd iops_wr);
 945
 946 sub print_drive {
 947     my ($vmid, $drive) = @_;
 948
 949     my $opts = '';
 950     foreach my $o (@qemu_drive_options, 'mbps', 'mbps_rd', 'mbps_wr', 'backup') {
 951         $opts .= ",$o=$drive->{$o}" if $drive->{$o};
 952     }
 953
 954     if ($drive->{size}) {
 955         $opts .= ",size=" . &$format_size($drive->{size});
 956     }
 957
 958     return "$drive->{file}$opts";
 959 }
 960
 961 sub scsi_inquiry {
 962     my($fh, $noerr) = @_;
 963
 964     my $SG_IO = 0x2285;
 965     my $SG_GET_VERSION_NUM = 0x2282;
 966
 967     my $versionbuf = "\x00" x 8;
 968     my $ret = ioctl($fh, $SG_GET_VERSION_NUM, $versionbuf);
 969     if (!$ret) {
 970         die "scsi ioctl SG_GET_VERSION_NUM failoed - $!\n" if !$noerr;
 971         return undef;
 972     }
 973     my $version = unpack("I", $versionbuf);
 974     if ($version < 30000) {
 975         die "scsi generic interface too old\n"  if !$noerr;
 976         return undef;
 977     }
 978
 979     my $buf = "\x00" x 36;
 980     my $sensebuf = "\x00" x 8;
 981     my $cmd = pack("C x3 C x11", 0x12, 36);
 982
 983     # see /usr/include/scsi/sg.h
 984     my $sg_io_hdr_t = "i i C C s I P P P I I i P C C C C S S i I I";
 985
 986     my $packet = pack($sg_io_hdr_t, ord('S'), -3, length($cmd),
 987                       length($sensebuf), 0, length($buf), $buf,
 988                       $cmd, $sensebuf, 6000);
 989
 990     $ret = ioctl($fh, $SG_IO, $packet);
 991     if (!$ret) {
 992         die "scsi ioctl SG_IO failed - $!\n" if !$noerr;
 993         return undef;
 994     }
 995
 996     my @res = unpack($sg_io_hdr_t, $packet);
 997     if ($res[17] || $res[18]) {
 998         die "scsi ioctl SG_IO status error - $!\n" if !$noerr;
 999         return undef;
1000     }
1001
1002     my $res = {};
1003     ($res->{device}, $res->{removable}, $res->{venodor},
1004      $res->{product}, $res->{revision}) = unpack("C C x6 A8 A16 A4", $buf);
1005
1006     return $res;
1007 }
1008
1009 sub path_is_scsi {
1010     my ($path) = @_;
1011
1012     my $fh = IO::File->new("+<$path") || return undef;
1013     my $res = scsi_inquiry($fh, 1);
1014     close($fh);
1015
1016     return $res;
1017 }
1018
1019 sub print_drivedevice_full {
1020     my ($storecfg, $conf, $vmid, $drive, $bridges) = @_;
1021
1022     my $device = '';
1023     my $maxdev = 0;
1024
1025     if ($drive->{interface} eq 'virtio') {
1026         my $pciaddr = print_pci_addr("$drive->{interface}$drive->{index}", $bridges);
1027         $device = "virtio-blk-pci,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}$pciaddr";
1028     } elsif ($drive->{interface} eq 'scsi') {
1029         $maxdev = ($conf->{scsihw} && $conf->{scsihw} ne 'lsi') ? 256 : 7;
1030         my $controller = int($drive->{index} / $maxdev);
1031         my $unit = $drive->{index} % $maxdev;
1032         my $devicetype = 'hd';
1033         my $path = '';
1034         if (drive_is_cdrom($drive)) {
1035               $devicetype = 'cd';
1036           } else {
1037               if ($drive->{file} =~ m|^/|) {
1038                   $path = $drive->{file};
1039               } else {
1040                   $path = PVE::Storage::path($storecfg, $drive->{file});
1041               }
1042
1043               if($path =~ m/^iscsi\:\/\//){
1044                  $devicetype = 'generic';
1045               }
1046               else {
1047                  $devicetype = 'block' if path_is_scsi($path);
1048               }
1049          }
1050
1051         if (!$conf->{scsihw} || $conf->{scsihw} eq 'lsi'){
1052             $device = "scsi-$devicetype,bus=scsihw$controller.0,scsi-id=$unit,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}" if !$conf->{scsihw} || $conf->{scsihw} eq 'lsi';
1053         } else {
1054             $device = "scsi-$devicetype,bus=scsihw$controller.0,channel=0,scsi-id=0,lun=$drive->{index},drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}";
1055         }
1056
1057     } elsif ($drive->{interface} eq 'ide'){
1058         $maxdev = 2;
1059         my $controller = int($drive->{index} / $maxdev);
1060         my $unit = $drive->{index} % $maxdev;
1061         my $devicetype = ($drive->{media} && $drive->{media} eq 'cdrom') ? "cd" : "hd";
1062
1063         $device = "ide-$devicetype,bus=ide.$controller,unit=$unit,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}";
1064     } elsif ($drive->{interface} eq 'sata'){
1065         my $controller = int($drive->{index} / $MAX_SATA_DISKS);
1066         my $unit = $drive->{index} % $MAX_SATA_DISKS;
1067         $device = "ide-drive,bus=ahci$controller.$unit,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}";
1068     } elsif ($drive->{interface} eq 'usb') {
1069         die "implement me";
1070         #  -device ide-drive,bus=ide.1,unit=0,drive=drive-ide0-1-0,id=ide0-1-0
1071     } else {
1072         die "unsupported interface type";
1073     }
1074
1075     $device .= ",bootindex=$drive->{bootindex}" if $drive->{bootindex};
1076
1077     return $device;
1078 }
1079
1080 sub print_drive_full {
1081     my ($storecfg, $vmid, $drive) = @_;
1082
1083     my $opts = '';
1084     foreach my $o (@qemu_drive_options) {
1085         next if $o eq 'bootindex';
1086         $opts .= ",$o=$drive->{$o}" if $drive->{$o};
1087     }
1088
1089     foreach my $o (qw(bps bps_rd bps_wr)) {
1090         my $v = $drive->{"m$o"};
1091         $opts .= ",$o=" . int($v*1024*1024) if $v;
1092     }
1093
1094     # use linux-aio by default (qemu default is threads)
1095     $opts .= ",aio=native" if !$drive->{aio};
1096
1097     my $path;
1098     my $volid = $drive->{file};
1099     if (drive_is_cdrom($drive)) {
1100         $path = get_iso_path($storecfg, $vmid, $volid);
1101     } else {
1102         if ($volid =~ m|^/|) {
1103             $path = $volid;
1104         } else {
1105             $path = PVE::Storage::path($storecfg, $volid);
1106         }
1107         if (!$drive->{cache} && ($path =~ m|^/dev/| || $path =~ m|\.raw$|)) {
1108             $opts .= ",cache=none";
1109         }
1110     }
1111
1112     my $pathinfo = $path ? "file=$path," : '';
1113
1114     return "${pathinfo}if=none,id=drive-$drive->{interface}$drive->{index}$opts";
1115 }
1116
1117 sub print_netdevice_full {
1118     my ($vmid, $conf, $net, $netid, $bridges) = @_;
1119
1120     my $bootorder = $conf->{boot} || $confdesc->{boot}->{default};
1121
1122     my $device = $net->{model};
1123     if ($net->{model} eq 'virtio') {
1124          $device = 'virtio-net-pci';
1125      };
1126
1127     # qemu > 0.15 always try to boot from network - we disable that by
1128     # not loading the pxe rom file
1129     my $extra = ($bootorder !~ m/n/) ? "romfile=," : '';
1130     my $pciaddr = print_pci_addr("$netid", $bridges);
1131     my $tmpstr = "$device,${extra}mac=$net->{macaddr},netdev=$netid$pciaddr,id=$netid";
1132     $tmpstr .= ",bootindex=$net->{bootindex}" if $net->{bootindex} ;
1133     return $tmpstr;
1134 }
1135
1136 sub print_netdev_full {
1137     my ($vmid, $conf, $net, $netid) = @_;
1138
1139     my $i = '';
1140     if ($netid =~ m/^net(\d+)$/) {
1141         $i = int($1);
1142     }
1143
1144     die "got strange net id '$i'\n" if $i >= ${MAX_NETS};
1145
1146     my $ifname = "tap${vmid}i$i";
1147
1148     # kvm uses TUNSETIFF ioctl, and that limits ifname length
1149     die "interface name '$ifname' is too long (max 15 character)\n"
1150         if length($ifname) >= 16;
1151
1152     my $vhostparam = '';
1153     $vhostparam = ',vhost=on' if $kernel_has_vhost_net && $net->{model} eq 'virtio';
1154
1155     my $vmname = $conf->{name} || "vm$vmid";
1156
1157     if ($net->{bridge}) {
1158         return "type=tap,id=$netid,ifname=${ifname},script=/var/lib/qemu-server/pve-bridge$vhostparam";
1159     } else {
1160         return "type=user,id=$netid,hostname=$vmname";
1161     }
1162 }
1163
1164 sub drive_is_cdrom {
1165     my ($drive) = @_;
1166
1167     return $drive && $drive->{media} && ($drive->{media} eq 'cdrom');
1168
1169 }
1170
1171 sub parse_hostpci {
1172     my ($value) = @_;
1173
1174     return undef if !$value;
1175
1176     my $res = {};
1177
1178     if ($value =~ m/^[a-f0-9]{2}:[a-f0-9]{2}\.[a-f0-9]$/) {
1179        $res->{pciid} = $value;
1180     } else {
1181        return undef;
1182     }
1183
1184     return $res;
1185 }
1186
1187 # netX: e1000=XX:XX:XX:XX:XX:XX,bridge=vmbr0,rate=<mbps>
1188 sub parse_net {
1189     my ($data) = @_;
1190
1191     my $res = {};
1192
1193     foreach my $kvp (split(/,/, $data)) {
1194
1195         if ($kvp =~ m/^(ne2k_pci|e1000|rtl8139|pcnet|virtio|ne2k_isa|i82551|i82557b|i82559er)(=([0-9a-f]{2}(:[0-9a-f]{2}){5}))?$/i) {
1196             my $model = lc($1);
1197             my $mac = uc($3) || PVE::Tools::random_ether_addr();
1198             $res->{model} = $model;
1199             $res->{macaddr} = $mac;
1200         } elsif ($kvp =~ m/^bridge=(\S+)$/) {
1201             $res->{bridge} = $1;
1202         } elsif ($kvp =~ m/^rate=(\d+(\.\d+)?)$/) {
1203             $res->{rate} = $1;
1204         } elsif ($kvp =~ m/^tag=(\d+)$/) {
1205             $res->{tag} = $1;
1206         } else {
1207             return undef;
1208         }
1209
1210     }
1211
1212     return undef if !$res->{model};
1213
1214     return $res;
1215 }
1216
1217 sub print_net {
1218     my $net = shift;
1219
1220     my $res = "$net->{model}";
1221     $res .= "=$net->{macaddr}" if $net->{macaddr};
1222     $res .= ",bridge=$net->{bridge}" if $net->{bridge};
1223     $res .= ",rate=$net->{rate}" if $net->{rate};
1224     $res .= ",tag=$net->{tag}" if $net->{tag};
1225
1226     return $res;
1227 }
1228
1229 sub add_random_macs {
1230     my ($settings) = @_;
1231
1232     foreach my $opt (keys %$settings) {
1233         next if $opt !~ m/^net(\d+)$/;
1234         my $net = parse_net($settings->{$opt});
1235         next if !$net;
1236         $settings->{$opt} = print_net($net);
1237     }
1238 }
1239
1240 sub add_unused_volume {
1241     my ($config, $volid) = @_;
1242
1243     my $key;
1244     for (my $ind = $MAX_UNUSED_DISKS - 1; $ind >= 0; $ind--) {
1245         my $test = "unused$ind";
1246         if (my $vid = $config->{$test}) {
1247             return if $vid eq $volid; # do not add duplicates
1248         } else {
1249             $key = $test;
1250         }
1251     }
1252
1253     die "To many unused volume - please delete them first.\n" if !$key;
1254
1255     $config->{$key} = $volid;
1256
1257     return $key;
1258 }
1259
1260 PVE::JSONSchema::register_format('pve-qm-bootdisk', \&verify_bootdisk);
1261 sub verify_bootdisk {
1262     my ($value, $noerr) = @_;
1263
1264     return $value if valid_drivename($value);
1265
1266     return undef if $noerr;
1267
1268     die "invalid boot disk '$value'\n";
1269 }
1270
1271 PVE::JSONSchema::register_format('pve-qm-net', \&verify_net);
1272 sub verify_net {
1273     my ($value, $noerr) = @_;
1274
1275     return $value if parse_net($value);
1276
1277     return undef if $noerr;
1278
1279     die "unable to parse network options\n";
1280 }
1281
1282 PVE::JSONSchema::register_format('pve-qm-drive', \&verify_drive);
1283 sub verify_drive {
1284     my ($value, $noerr) = @_;
1285
1286     return $value if parse_drive(undef, $value);
1287
1288     return undef if $noerr;
1289
1290     die "unable to parse drive options\n";
1291 }
1292
1293 PVE::JSONSchema::register_format('pve-qm-hostpci', \&verify_hostpci);
1294 sub verify_hostpci {
1295     my ($value, $noerr) = @_;
1296
1297     return $value if parse_hostpci($value);
1298
1299     return undef if $noerr;
1300
1301     die "unable to parse pci id\n";
1302 }
1303
1304 PVE::JSONSchema::register_format('pve-qm-watchdog', \&verify_watchdog);
1305 sub verify_watchdog {
1306     my ($value, $noerr) = @_;
1307
1308     return $value if parse_watchdog($value);
1309
1310     return undef if $noerr;
1311
1312     die "unable to parse watchdog options\n";
1313 }
1314
1315 sub parse_watchdog {
1316     my ($value) = @_;
1317
1318     return undef if !$value;
1319
1320     my $res = {};
1321
1322     foreach my $p (split(/,/, $value)) {
1323         next if $p =~ m/^\s*$/;
1324
1325         if ($p =~ m/^(model=)?(i6300esb|ib700)$/) {
1326             $res->{model} = $2;
1327         } elsif ($p =~ m/^(action=)?(reset|shutdown|poweroff|pause|debug|none)$/) {
1328             $res->{action} = $2;
1329         } else {
1330             return undef;
1331         }
1332     }
1333
1334     return $res;
1335 }
1336
1337 PVE::JSONSchema::register_format('pve-qm-startup', \&verify_startup);
1338 sub verify_startup {
1339     my ($value, $noerr) = @_;
1340
1341     return $value if parse_startup($value);
1342
1343     return undef if $noerr;
1344
1345     die "unable to parse startup options\n";
1346 }
1347
1348 sub parse_startup {
1349     my ($value) = @_;
1350
1351     return undef if !$value;
1352
1353     my $res = {};
1354
1355     foreach my $p (split(/,/, $value)) {
1356         next if $p =~ m/^\s*$/;
1357
1358         if ($p =~ m/^(order=)?(\d+)$/) {
1359             $res->{order} = $2;
1360         } elsif ($p =~ m/^up=(\d+)$/) {
1361             $res->{up} = $1;
1362         } elsif ($p =~ m/^down=(\d+)$/) {
1363             $res->{down} = $1;
1364         } else {
1365             return undef;
1366         }
1367     }
1368
1369     return $res;
1370 }
1371
1372 sub parse_usb_device {
1373     my ($value) = @_;
1374
1375     return undef if !$value;
1376
1377     my @dl = split(/,/, $value);
1378     my $found;
1379
1380     my $res = {};
1381     foreach my $v (@dl) {
1382         if ($v =~ m/^host=(0x)?([0-9A-Fa-f]{4}):(0x)?([0-9A-Fa-f]{4})$/) {
1383             $found = 1;
1384             $res->{vendorid} = $2;
1385             $res->{productid} = $4;
1386         } elsif ($v =~ m/^host=(\d+)\-(\d+(\.\d+)*)$/) {
1387             $found = 1;
1388             $res->{hostbus} = $1;
1389             $res->{hostport} = $2;
1390         } else {
1391             return undef;
1392         }
1393     }
1394     return undef if !$found;
1395
1396     return $res;
1397 }
1398
1399 PVE::JSONSchema::register_format('pve-qm-usb-device', \&verify_usb_device);
1400 sub verify_usb_device {
1401     my ($value, $noerr) = @_;
1402
1403     return $value if parse_usb_device($value);
1404
1405     return undef if $noerr;
1406
1407     die "unable to parse usb device\n";
1408 }
1409
1410 # add JSON properties for create and set function
1411 sub json_config_properties {
1412     my $prop = shift;
1413
1414     foreach my $opt (keys %$confdesc) {
1415         next if $opt eq 'parent' || $opt eq 'snaptime' || $opt eq 'vmstate';
1416         $prop->{$opt} = $confdesc->{$opt};
1417     }
1418
1419     return $prop;
1420 }
1421
1422 sub check_type {
1423     my ($key, $value) = @_;
1424
1425     die "unknown setting '$key'\n" if !$confdesc->{$key};
1426
1427     my $type = $confdesc->{$key}->{type};
1428
1429     if (!defined($value)) {
1430         die "got undefined value\n";
1431     }
1432
1433     if ($value =~ m/[\n\r]/) {
1434         die "property contains a line feed\n";
1435     }
1436
1437     if ($type eq 'boolean') {
1438         return 1 if ($value eq '1') || ($value =~ m/^(on|yes|true)$/i);
1439         return 0 if ($value eq '0') || ($value =~ m/^(off|no|false)$/i);
1440         die "type check ('boolean') failed - got '$value'\n";
1441     } elsif ($type eq 'integer') {
1442         return int($1) if $value =~ m/^(\d+)$/;
1443         die "type check ('integer') failed - got '$value'\n";
1444     } elsif ($type eq 'string') {
1445         if (my $fmt = $confdesc->{$key}->{format}) {
1446             if ($fmt eq 'pve-qm-drive') {
1447                 # special case - we need to pass $key to parse_drive()
1448                 my $drive = parse_drive($key, $value);
1449                 return $value if $drive;
1450                 die "unable to parse drive options\n";
1451             }
1452             PVE::JSONSchema::check_format($fmt, $value);
1453             return $value;
1454         }
1455         $value =~ s/^\"(.*)\"$/$1/;
1456         return $value;
1457     } else {
1458         die "internal error"
1459     }
1460 }
1461
1462 sub lock_config_full {
1463     my ($vmid, $timeout, $code, @param) = @_;
1464
1465     my $filename = config_file_lock($vmid);
1466
1467     my $res = lock_file($filename, $timeout, $code, @param);
1468
1469     die $@ if $@;
1470
1471     return $res;
1472 }
1473
1474 sub lock_config {
1475     my ($vmid, $code, @param) = @_;
1476
1477     return lock_config_full($vmid, 10, $code, @param);
1478 }
1479
1480 sub cfs_config_path {
1481     my ($vmid, $node) = @_;
1482
1483     $node = $nodename if !$node;
1484     return "nodes/$node/qemu-server/$vmid.conf";
1485 }
1486
1487 sub check_iommu_support{
1488     #fixme : need to check IOMMU support
1489     #http://www.linux-kvm.org/page/How_to_assign_devices_with_VT-d_in_KVM
1490
1491     my $iommu=1;
1492     return $iommu;
1493
1494 }
1495
1496 sub config_file {
1497     my ($vmid, $node) = @_;
1498
1499     my $cfspath = cfs_config_path($vmid, $node);
1500     return "/etc/pve/$cfspath";
1501 }
1502
1503 sub config_file_lock {
1504     my ($vmid) = @_;
1505
1506     return "$lock_dir/lock-$vmid.conf";
1507 }
1508
1509 sub touch_config {
1510     my ($vmid) = @_;
1511
1512     my $conf = config_file($vmid);
1513     utime undef, undef, $conf;
1514 }
1515
1516 sub destroy_vm {
1517     my ($storecfg, $vmid, $keep_empty_config) = @_;
1518
1519     my $conffile = config_file($vmid);
1520
1521     my $conf = load_config($vmid);
1522
1523     check_lock($conf);
1524
1525     # only remove disks owned by this VM
1526     foreach_drive($conf, sub {
1527         my ($ds, $drive) = @_;
1528
1529         return if drive_is_cdrom($drive);
1530
1531         my $volid = $drive->{file};
1532         return if !$volid || $volid =~ m|^/|;
1533
1534         my ($path, $owner) = PVE::Storage::path($storecfg, $volid);
1535         return if !$path || !$owner || ($owner != $vmid);
1536
1537         PVE::Storage::vdisk_free($storecfg, $volid);
1538     });
1539
1540     if ($keep_empty_config) {
1541         PVE::Tools::file_set_contents($conffile, "memory: 128\n");
1542     } else {
1543         unlink $conffile;
1544     }
1545
1546     # also remove unused disk
1547     eval {
1548         my $dl = PVE::Storage::vdisk_list($storecfg, undef, $vmid);
1549
1550         eval {
1551             PVE::Storage::foreach_volid($dl, sub {
1552                 my ($volid, $sid, $volname, $d) = @_;
1553                 PVE::Storage::vdisk_free($storecfg, $volid);
1554             });
1555         };
1556         warn $@ if $@;
1557
1558     };
1559     warn $@ if $@;
1560 }
1561
1562 sub load_config {
1563     my ($vmid, $node) = @_;
1564
1565     my $cfspath = cfs_config_path($vmid, $node);
1566
1567     my $conf = PVE::Cluster::cfs_read_file($cfspath);
1568
1569     die "no such VM ('$vmid')\n" if !defined($conf);
1570
1571     return $conf;
1572 }
1573
1574 sub parse_vm_config {
1575     my ($filename, $raw) = @_;
1576
1577     return undef if !defined($raw);
1578
1579     my $res = {
1580         digest => Digest::SHA::sha1_hex($raw),
1581         snapshots => {},
1582     };
1583
1584     $filename =~ m|/qemu-server/(\d+)\.conf$|
1585         || die "got strange filename '$filename'";
1586
1587     my $vmid = $1;
1588
1589     my $conf = $res;
1590     my $descr = '';
1591
1592     my @lines = split(/\n/, $raw);
1593     foreach my $line (@lines) {
1594         next if $line =~ m/^\s*$/;
1595
1596         if ($line =~ m/^\[([a-z][a-z0-9_\-]+)\]\s*$/i) {
1597             my $snapname = $1;
1598             $conf->{description} = $descr if $descr;
1599             $descr = '';
1600             $conf = $res->{snapshots}->{$snapname} = {};
1601             next;
1602         }
1603
1604         if ($line =~ m/^\#(.*)\s*$/) {
1605             $descr .= PVE::Tools::decode_text($1) . "\n";
1606             next;
1607         }
1608
1609         if ($line =~ m/^(description):\s*(.*\S)\s*$/) {
1610             $descr .= PVE::Tools::decode_text($2);
1611         } elsif ($line =~ m/snapstate:\s*(prepare|delete)\s*$/) {
1612             $conf->{snapstate} = $1;
1613         } elsif ($line =~ m/^(args):\s*(.*\S)\s*$/) {
1614             my $key = $1;
1615             my $value = $2;
1616             $conf->{$key} = $value;
1617         } elsif ($line =~ m/^([a-z][a-z_]*\d*):\s*(\S+)\s*$/) {
1618             my $key = $1;
1619             my $value = $2;
1620             eval { $value = check_type($key, $value); };
1621             if ($@) {
1622                 warn "vm $vmid - unable to parse value of '$key' - $@";
1623             } else {
1624                 my $fmt = $confdesc->{$key}->{format};
1625                 if ($fmt && $fmt eq 'pve-qm-drive') {
1626                     my $v = parse_drive($key, $value);
1627                     if (my $volid = filename_to_volume_id($vmid, $v->{file}, $v->{media})) {
1628                         $v->{file} = $volid;
1629                         $value = print_drive($vmid, $v);
1630                     } else {
1631                         warn "vm $vmid - unable to parse value of '$key'\n";
1632                         next;
1633                     }
1634                 }
1635
1636                 if ($key eq 'cdrom') {
1637                     $conf->{ide2} = $value;
1638                 } else {
1639                     $conf->{$key} = $value;
1640                 }
1641             }
1642         }
1643     }
1644
1645     $conf->{description} = $descr if $descr;
1646
1647     delete $res->{snapstate}; # just to be sure
1648
1649     return $res;
1650 }
1651
1652 sub write_vm_config {
1653     my ($filename, $conf) = @_;
1654
1655     delete $conf->{snapstate}; # just to be sure
1656
1657     if ($conf->{cdrom}) {
1658         die "option ide2 conflicts with cdrom\n" if $conf->{ide2};
1659         $conf->{ide2} = $conf->{cdrom};
1660         delete $conf->{cdrom};
1661     }
1662
1663     # we do not use 'smp' any longer
1664     if ($conf->{sockets}) {
1665         delete $conf->{smp};
1666     } elsif ($conf->{smp}) {
1667         $conf->{sockets} = $conf->{smp};
1668         delete $conf->{cores};
1669         delete $conf->{smp};
1670     }
1671
1672     my $used_volids = {};
1673
1674     my $cleanup_config = sub {
1675         my ($cref) = @_;
1676
1677         foreach my $key (keys %$cref) {
1678             next if $key eq 'digest' || $key eq 'description' || $key eq 'snapshots' ||
1679                 $key eq 'snapstate';
1680             my $value = $cref->{$key};
1681             eval { $value = check_type($key, $value); };
1682             die "unable to parse value of '$key' - $@" if $@;
1683
1684             $cref->{$key} = $value;
1685
1686             if (valid_drivename($key)) {
1687                 my $drive = PVE::QemuServer::parse_drive($key, $value);
1688                 $used_volids->{$drive->{file}} = 1 if $drive && $drive->{file};
1689             }
1690         }
1691     };
1692
1693     &$cleanup_config($conf);
1694     foreach my $snapname (keys %{$conf->{snapshots}}) {
1695         &$cleanup_config($conf->{snapshots}->{$snapname});
1696     }
1697
1698     # remove 'unusedX' settings if we re-add a volume
1699     foreach my $key (keys %$conf) {
1700         my $value = $conf->{$key};
1701         if ($key =~ m/^unused/ && $used_volids->{$value}) {
1702             delete $conf->{$key};
1703         }
1704     }
1705
1706     my $generate_raw_config = sub {
1707         my ($conf) = @_;
1708
1709         my $raw = '';
1710
1711         # add description as comment to top of file
1712         my $descr = $conf->{description} || '';
1713         foreach my $cl (split(/\n/, $descr)) {
1714             $raw .= '#' .  PVE::Tools::encode_text($cl) . "\n";
1715         }
1716
1717         foreach my $key (sort keys %$conf) {
1718             next if $key eq 'digest' || $key eq 'description' || $key eq 'snapshots';
1719             $raw .= "$key: $conf->{$key}\n";
1720         }
1721         return $raw;
1722     };
1723
1724     my $raw = &$generate_raw_config($conf);
1725     foreach my $snapname (sort keys %{$conf->{snapshots}}) {
1726         $raw .= "\n[$snapname]\n";
1727         $raw .= &$generate_raw_config($conf->{snapshots}->{$snapname});
1728     }
1729
1730     return $raw;
1731 }
1732
1733 sub update_config_nolock {
1734     my ($vmid, $conf, $skiplock) = @_;
1735
1736     check_lock($conf) if !$skiplock;
1737
1738     my $cfspath = cfs_config_path($vmid);
1739
1740     PVE::Cluster::cfs_write_file($cfspath, $conf);
1741 }
1742
1743 sub update_config {
1744     my ($vmid, $conf, $skiplock) = @_;
1745
1746     lock_config($vmid, &update_config_nolock, $conf, $skiplock);
1747 }
1748
1749 sub load_defaults {
1750
1751     my $res = {};
1752
1753     # we use static defaults from our JSON schema configuration
1754     foreach my $key (keys %$confdesc) {
1755         if (defined(my $default = $confdesc->{$key}->{default})) {
1756             $res->{$key} = $default;
1757         }
1758     }
1759
1760     my $conf = PVE::Cluster::cfs_read_file('datacenter.cfg');
1761     $res->{keyboard} = $conf->{keyboard} if $conf->{keyboard};
1762
1763     return $res;
1764 }
1765
1766 sub config_list {
1767     my $vmlist = PVE::Cluster::get_vmlist();
1768     my $res = {};
1769     return $res if !$vmlist || !$vmlist->{ids};
1770     my $ids = $vmlist->{ids};
1771
1772     foreach my $vmid (keys %$ids) {
1773         my $d = $ids->{$vmid};
1774         next if !$d->{node} || $d->{node} ne $nodename;
1775         next if !$d->{type} || $d->{type} ne 'qemu';
1776         $res->{$vmid}->{exists} = 1;
1777     }
1778     return $res;
1779 }
1780
1781 # test if VM uses local resources (to prevent migration)
1782 sub check_local_resources {
1783     my ($conf, $noerr) = @_;
1784
1785     my $loc_res = 0;
1786
1787     $loc_res = 1 if $conf->{hostusb}; # old syntax
1788     $loc_res = 1 if $conf->{hostpci}; # old syntax
1789
1790     foreach my $k (keys %$conf) {
1791         $loc_res = 1 if $k =~ m/^(usb|hostpci|serial|parallel)\d+$/;
1792     }
1793
1794     die "VM uses local resources\n" if $loc_res && !$noerr;
1795
1796     return $loc_res;
1797 }
1798
1799 # check is used storages are available on all nodes (use by migrate)
1800 sub check_storage_availability {
1801     my ($storecfg, $conf, $node) = @_;
1802
1803     foreach_drive($conf, sub {
1804         my ($ds, $drive) = @_;
1805
1806         my $volid = $drive->{file};
1807         return if !$volid;
1808
1809         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
1810         return if !$sid;
1811
1812         # check if storage is available on both nodes
1813         my $scfg = PVE::Storage::storage_check_node($storecfg, $sid);
1814         PVE::Storage::storage_check_node($storecfg, $sid, $node);
1815    });
1816 }
1817
1818 sub check_lock {
1819     my ($conf) = @_;
1820
1821     die "VM is locked ($conf->{lock})\n" if $conf->{lock};
1822 }
1823
1824 sub check_cmdline {
1825     my ($pidfile, $pid) = @_;
1826
1827     my $fh = IO::File->new("/proc/$pid/cmdline", "r");
1828     if (defined($fh)) {
1829         my $line = <$fh>;
1830         $fh->close;
1831         return undef if !$line;
1832         my @param = split(/\0/, $line);
1833
1834         my $cmd = $param[0];
1835         return if !$cmd || ($cmd !~ m|kvm$|);
1836
1837         for (my $i = 0; $i < scalar (@param); $i++) {
1838             my $p = $param[$i];
1839             next if !$p;
1840             if (($p eq '-pidfile') || ($p eq '--pidfile')) {
1841                 my $p = $param[$i+1];
1842                 return 1 if $p && ($p eq $pidfile);
1843                 return undef;
1844             }
1845         }
1846     }
1847     return undef;
1848 }
1849
1850 sub check_running {
1851     my ($vmid, $nocheck, $node) = @_;
1852
1853     my $filename = config_file($vmid, $node);
1854
1855     die "unable to find configuration file for VM $vmid - no such machine\n"
1856         if !$nocheck && ! -f $filename;
1857
1858     my $pidfile = pidfile_name($vmid);
1859
1860     if (my $fd = IO::File->new("<$pidfile")) {
1861         my $st = stat($fd);
1862         my $line = <$fd>;
1863         close($fd);
1864
1865         my $mtime = $st->mtime;
1866         if ($mtime > time()) {
1867             warn "file '$filename' modified in future\n";
1868         }
1869
1870         if ($line =~ m/^(\d+)$/) {
1871             my $pid = $1;
1872             if (check_cmdline($pidfile, $pid)) {
1873                 if (my $pinfo = PVE::ProcFSTools::check_process_running($pid)) {
1874                     return $pid;
1875                 }
1876             }
1877         }
1878     }
1879
1880     return undef;
1881 }
1882
1883 sub vzlist {
1884
1885     my $vzlist = config_list();
1886
1887     my $fd = IO::Dir->new($var_run_tmpdir) || return $vzlist;
1888
1889     while (defined(my $de = $fd->read)) {
1890         next if $de !~ m/^(\d+)\.pid$/;
1891         my $vmid = $1;
1892         next if !defined($vzlist->{$vmid});
1893         if (my $pid = check_running($vmid)) {
1894             $vzlist->{$vmid}->{pid} = $pid;
1895         }
1896     }
1897
1898     return $vzlist;
1899 }
1900
1901 sub disksize {
1902     my ($storecfg, $conf) = @_;
1903
1904     my $bootdisk = $conf->{bootdisk};
1905     return undef if !$bootdisk;
1906     return undef if !valid_drivename($bootdisk);
1907
1908     return undef if !$conf->{$bootdisk};
1909
1910     my $drive = parse_drive($bootdisk, $conf->{$bootdisk});
1911     return undef if !defined($drive);
1912
1913     return undef if drive_is_cdrom($drive);
1914
1915     my $volid = $drive->{file};
1916     return undef if !$volid;
1917
1918     return $drive->{size};
1919 }
1920
1921 my $last_proc_pid_stat;
1922
1923 # get VM status information
1924 # This must be fast and should not block ($full == false)
1925 # We only query KVM using QMP if $full == true (this can be slow)
1926 sub vmstatus {
1927     my ($opt_vmid, $full) = @_;
1928
1929     my $res = {};
1930
1931     my $storecfg = PVE::Storage::config();
1932
1933     my $list = vzlist();
1934     my ($uptime) = PVE::ProcFSTools::read_proc_uptime(1);
1935
1936     my $cpucount = $cpuinfo->{cpus} || 1;
1937
1938     foreach my $vmid (keys %$list) {
1939         next if $opt_vmid && ($vmid ne $opt_vmid);
1940
1941         my $cfspath = cfs_config_path($vmid);
1942         my $conf = PVE::Cluster::cfs_read_file($cfspath) || {};
1943
1944         my $d = {};
1945         $d->{pid} = $list->{$vmid}->{pid};
1946
1947         # fixme: better status?
1948         $d->{status} = $list->{$vmid}->{pid} ? 'running' : 'stopped';
1949
1950         my $size = disksize($storecfg, $conf);
1951         if (defined($size)) {
1952             $d->{disk} = 0; # no info available
1953             $d->{maxdisk} = $size;
1954         } else {
1955             $d->{disk} = 0;
1956             $d->{maxdisk} = 0;
1957         }
1958
1959         $d->{cpus} = ($conf->{sockets} || 1) * ($conf->{cores} || 1);
1960         $d->{cpus} = $cpucount if $d->{cpus} > $cpucount;
1961
1962         $d->{name} = $conf->{name} || "VM $vmid";
1963         $d->{maxmem} = $conf->{memory} ? $conf->{memory}*(1024*1024) : 0;
1964
1965         if ($conf->{balloon}) {
1966             $d->{balloon_min} = $conf->{balloon}*(1024*1024);
1967             $d->{shares} = $conf->{shares} || 1000;
1968         }
1969
1970         $d->{uptime} = 0;
1971         $d->{cpu} = 0;
1972         $d->{mem} = 0;
1973
1974         $d->{netout} = 0;
1975         $d->{netin} = 0;
1976
1977         $d->{diskread} = 0;
1978         $d->{diskwrite} = 0;
1979
1980         $res->{$vmid} = $d;
1981     }
1982
1983     my $netdev = PVE::ProcFSTools::read_proc_net_dev();
1984     foreach my $dev (keys %$netdev) {
1985         next if $dev !~ m/^tap([1-9]\d*)i/;
1986         my $vmid = $1;
1987         my $d = $res->{$vmid};
1988         next if !$d;
1989
1990         $d->{netout} += $netdev->{$dev}->{receive};
1991         $d->{netin} += $netdev->{$dev}->{transmit};
1992     }
1993
1994     my $ctime = gettimeofday;
1995
1996     foreach my $vmid (keys %$list) {
1997
1998         my $d = $res->{$vmid};
1999         my $pid = $d->{pid};
2000         next if !$pid;
2001
2002         my $pstat = PVE::ProcFSTools::read_proc_pid_stat($pid);
2003         next if !$pstat; # not running
2004
2005         my $used = $pstat->{utime} + $pstat->{stime};
2006
2007         $d->{uptime} = int(($uptime - $pstat->{starttime})/$cpuinfo->{user_hz});
2008
2009         if ($pstat->{vsize}) {
2010             $d->{mem} = int(($pstat->{rss}/$pstat->{vsize})*$d->{maxmem});
2011         }
2012
2013         my $old = $last_proc_pid_stat->{$pid};
2014         if (!$old) {
2015             $last_proc_pid_stat->{$pid} = {
2016                 time => $ctime,
2017                 used => $used,
2018                 cpu => 0,
2019             };
2020             next;
2021         }
2022
2023         my $dtime = ($ctime -  $old->{time}) * $cpucount * $cpuinfo->{user_hz};
2024
2025         if ($dtime > 1000) {
2026             my $dutime = $used -  $old->{used};
2027
2028             $d->{cpu} = (($dutime/$dtime)* $cpucount) / $d->{cpus};
2029             $last_proc_pid_stat->{$pid} = {
2030                 time => $ctime,
2031                 used => $used,
2032                 cpu => $d->{cpu},
2033             };
2034         } else {
2035             $d->{cpu} = $old->{cpu};
2036         }
2037     }
2038
2039     return $res if !$full;
2040
2041     my $qmpclient = PVE::QMPClient->new();
2042
2043     my $ballooncb = sub {
2044         my ($vmid, $resp) = @_;
2045
2046         my $info = $resp->{'return'};
2047         return if !$info->{max_mem};
2048
2049         my $d = $res->{$vmid};
2050
2051         # use memory assigned to VM
2052         $d->{maxmem} = $info->{max_mem};
2053         $d->{balloon} = $info->{actual};
2054
2055         if (defined($info->{total_mem}) && defined($info->{free_mem})) {
2056             $d->{mem} = $info->{total_mem} - $info->{free_mem};
2057             $d->{freemem} = $info->{free_mem};
2058         }
2059
2060     };
2061
2062     my $blockstatscb = sub {
2063         my ($vmid, $resp) = @_;
2064         my $data = $resp->{'return'} || [];
2065         my $totalrdbytes = 0;
2066         my $totalwrbytes = 0;
2067         for my $blockstat (@$data) {
2068             $totalrdbytes = $totalrdbytes + $blockstat->{stats}->{rd_bytes};
2069             $totalwrbytes = $totalwrbytes + $blockstat->{stats}->{wr_bytes};
2070         }
2071         $res->{$vmid}->{diskread} = $totalrdbytes;
2072         $res->{$vmid}->{diskwrite} = $totalwrbytes;
2073     };
2074
2075     my $statuscb = sub {
2076         my ($vmid, $resp) = @_;
2077
2078         $qmpclient->queue_cmd($vmid, $blockstatscb, 'query-blockstats');
2079         # this fails if ballon driver is not loaded, so this must be
2080         # the last commnand (following command are aborted if this fails).
2081         $qmpclient->queue_cmd($vmid, $ballooncb, 'query-balloon');
2082
2083         my $status = 'unknown';
2084         if (!defined($status = $resp->{'return'}->{status})) {
2085             warn "unable to get VM status\n";
2086             return;
2087         }
2088
2089         $res->{$vmid}->{qmpstatus} = $resp->{'return'}->{status};
2090     };
2091
2092     foreach my $vmid (keys %$list) {
2093         next if $opt_vmid && ($vmid ne $opt_vmid);
2094         next if !$res->{$vmid}->{pid}; # not running
2095         $qmpclient->queue_cmd($vmid, $statuscb, 'query-status');
2096     }
2097
2098     $qmpclient->queue_execute();
2099
2100     foreach my $vmid (keys %$list) {
2101         next if $opt_vmid && ($vmid ne $opt_vmid);
2102         $res->{$vmid}->{qmpstatus} = $res->{$vmid}->{status} if !$res->{$vmid}->{qmpstatus};
2103     }
2104
2105     return $res;
2106 }
2107
2108 sub foreach_drive {
2109     my ($conf, $func) = @_;
2110
2111     foreach my $ds (keys %$conf) {
2112         next if !valid_drivename($ds);
2113
2114         my $drive = parse_drive($ds, $conf->{$ds});
2115         next if !$drive;
2116
2117         &$func($ds, $drive);
2118     }
2119 }
2120
2121 sub foreach_volid {
2122     my ($conf, $func) = @_;
2123
2124     my $volhash = {};
2125
2126     my $test_volid = sub {
2127         my ($volid, $is_cdrom) = @_;
2128
2129         return if !$volid;
2130
2131         $volhash->{$volid} = $is_cdrom || 0;
2132     };
2133
2134     PVE::QemuServer::foreach_drive($conf, sub {
2135         my ($ds, $drive) = @_;
2136         &$test_volid($drive->{file}, drive_is_cdrom($drive));
2137     });
2138
2139     foreach my $snapname (keys %{$conf->{snapshots}}) {
2140         my $snap = $conf->{snapshots}->{$snapname};
2141         &$test_volid($snap->{vmstate}, 0);
2142         PVE::QemuServer::foreach_drive($snap, sub {
2143             my ($ds, $drive) = @_;
2144             &$test_volid($drive->{file}, drive_is_cdrom($drive));
2145         });
2146     }
2147
2148     foreach my $volid (keys %$volhash) {
2149         &$func($volid, $volhash->{$volid});
2150     }
2151 }
2152
2153 sub config_to_command {
2154     my ($storecfg, $vmid, $conf, $defaults) = @_;
2155
2156     my $cmd = [];
2157     my $globalFlags = [];
2158     my $machineFlags = [];
2159     my $rtcFlags = [];
2160     my $devices = [];
2161     my $pciaddr = '';
2162     my $bridges = {};
2163     my $kvmver = kvm_user_version();
2164     my $vernum = 0; # unknown
2165     if ($kvmver =~ m/^(\d+)\.(\d+)$/) {
2166         $vernum = $1*1000000+$2*1000;
2167     } elsif ($kvmver =~ m/^(\d+)\.(\d+)\.(\d+)$/) {
2168         $vernum = $1*1000000+$2*1000+$3;
2169     }
2170
2171     die "detected old qemu-kvm binary ($kvmver)\n" if $vernum < 15000;
2172
2173     my $have_ovz = -f '/proc/vz/vestat';
2174
2175     push @$cmd, '/usr/bin/kvm';
2176
2177     push @$cmd, '-id', $vmid;
2178
2179     my $use_virtio = 0;
2180
2181     my $qmpsocket = qmp_socket($vmid);
2182     push @$cmd, '-chardev', "socket,id=qmp,path=$qmpsocket,server,nowait";
2183     push @$cmd, '-mon', "chardev=qmp,mode=control";
2184
2185     my $socket = vnc_socket($vmid);
2186     push @$cmd,  '-vnc', "unix:$socket,x509,password";
2187
2188     push @$cmd, '-pidfile' , pidfile_name($vmid);
2189
2190     push @$cmd, '-daemonize';
2191
2192     my $use_usb2 = 0;
2193     for (my $i = 0; $i < $MAX_USB_DEVICES; $i++)  {
2194         next if !$conf->{"usb$i"};
2195         $use_usb2 = 1;
2196     }
2197     # include usb device config
2198     push @$devices, '-readconfig', '/usr/share/qemu-server/pve-usb.cfg' if $use_usb2;
2199
2200     # enable absolute mouse coordinates (needed by vnc)
2201     my $tablet = defined($conf->{tablet}) ? $conf->{tablet} : $defaults->{tablet};
2202     if ($tablet) {
2203         if ($use_usb2) {
2204             push @$devices, '-device', 'usb-tablet,bus=ehci.0,port=6';
2205         } else {
2206             push @$devices, '-usbdevice', 'tablet';
2207         }
2208     }
2209
2210     # host pci devices
2211     for (my $i = 0; $i < $MAX_HOSTPCI_DEVICES; $i++)  {
2212           my $d = parse_hostpci($conf->{"hostpci$i"});
2213           next if !$d;
2214           $pciaddr = print_pci_addr("hostpci$i", $bridges);
2215           push @$devices, '-device', "pci-assign,host=$d->{pciid},id=hostpci$i$pciaddr";
2216     }
2217
2218     # usb devices
2219     for (my $i = 0; $i < $MAX_USB_DEVICES; $i++)  {
2220         my $d = parse_usb_device($conf->{"usb$i"});
2221         next if !$d;
2222         if ($d->{vendorid} && $d->{productid}) {
2223             push @$devices, '-device', "usb-host,vendorid=0x$d->{vendorid},productid=0x$d->{productid}";
2224         } elsif (defined($d->{hostbus}) && defined($d->{hostport})) {
2225             push @$devices, '-device', "usb-host,hostbus=$d->{hostbus},hostport=$d->{hostport}";
2226         }
2227     }
2228
2229     # serial devices
2230     for (my $i = 0; $i < $MAX_SERIAL_PORTS; $i++)  {
2231         if (my $path = $conf->{"serial$i"}) {
2232             die "no such serial device\n" if ! -c $path;
2233             push @$devices, '-chardev', "tty,id=serial$i,path=$path";
2234             push @$devices, '-device', "isa-serial,chardev=serial$i";
2235         }
2236     }
2237
2238     # parallel devices
2239     for (my $i = 0; $i < $MAX_PARALLEL_PORTS; $i++)  {
2240         if (my $path = $conf->{"parallel$i"}) {
2241             die "no such parallel device\n" if ! -c $path;
2242             push @$devices, '-chardev', "parport,id=parallel$i,path=$path";
2243             push @$devices, '-device', "isa-parallel,chardev=parallel$i";
2244         }
2245     }
2246
2247     my $vmname = $conf->{name} || "vm$vmid";
2248
2249     push @$cmd, '-name', $vmname;
2250
2251     my $sockets = 1;
2252     $sockets = $conf->{smp} if $conf->{smp}; # old style - no longer iused
2253     $sockets = $conf->{sockets} if  $conf->{sockets};
2254
2255     my $cores = $conf->{cores} || 1;
2256
2257     push @$cmd, '-smp', "sockets=$sockets,cores=$cores";
2258
2259     push @$cmd, '-cpu', $conf->{cpu} if $conf->{cpu};
2260
2261     push @$cmd, '-nodefaults';
2262
2263     my $bootorder = $conf->{boot} || $confdesc->{boot}->{default};
2264
2265     my $bootindex_hash = {};
2266     my $i = 1;
2267     foreach my $o (split(//, $bootorder)) {
2268         $bootindex_hash->{$o} = $i*100;
2269         $i++;
2270     }
2271
2272     push @$cmd, '-boot', "menu=on";
2273
2274     push @$cmd, '-no-acpi' if defined($conf->{acpi}) && $conf->{acpi} == 0;
2275
2276     push @$cmd, '-no-reboot' if  defined($conf->{reboot}) && $conf->{reboot} == 0;
2277
2278     my $vga = $conf->{vga};
2279     if (!$vga) {
2280         if ($conf->{ostype} && ($conf->{ostype} eq 'win8' || $conf->{ostype} eq 'win7' || $conf->{ostype} eq 'w2k8')) {
2281             $vga = 'std';
2282         } else {
2283             $vga = 'cirrus';
2284         }
2285     }
2286
2287     push @$cmd, '-vga', $vga if $vga; # for kvm 77 and later
2288
2289     # time drift fix
2290     my $tdf = defined($conf->{tdf}) ? $conf->{tdf} : $defaults->{tdf};
2291
2292     my $nokvm = defined($conf->{kvm}) && $conf->{kvm} == 0 ? 1 : 0;
2293     my $useLocaltime = $conf->{localtime};
2294
2295     if (my $ost = $conf->{ostype}) {
2296         # other, wxp, w2k, w2k3, w2k8, wvista, win7, win8, l24, l26
2297
2298         if ($ost =~ m/^w/) { # windows
2299             $useLocaltime = 1 if !defined($conf->{localtime});
2300
2301             # use time drift fix when acpi is enabled
2302             if (!(defined($conf->{acpi}) && $conf->{acpi} == 0)) {
2303                 $tdf = 1 if !defined($conf->{tdf});
2304             }
2305         }
2306
2307         if ($ost eq 'win7' || $ost eq 'win8' || $ost eq 'w2k8' ||
2308             $ost eq 'wvista') {
2309             push @$globalFlags, 'kvm-pit.lost_tick_policy=discard';
2310             push @$cmd, '-no-hpet';
2311         }
2312     }
2313
2314     push @$rtcFlags, 'driftfix=slew' if $tdf;
2315
2316     if ($nokvm) {
2317         push @$machineFlags, 'accel=tcg';
2318     } else {
2319         die "No accelerator found!\n" if !$cpuinfo->{hvm};
2320     }
2321
2322     if ($conf->{startdate}) {
2323         push @$rtcFlags, "base=$conf->{startdate}";
2324     } elsif ($useLocaltime) {
2325         push @$rtcFlags, 'base=localtime';
2326     }
2327
2328     push @$cmd, '-S' if $conf->{freeze};
2329
2330     # set keyboard layout
2331     my $kb = $conf->{keyboard} || $defaults->{keyboard};
2332     push @$cmd, '-k', $kb if $kb;
2333
2334     # enable sound
2335     #my $soundhw = $conf->{soundhw} || $defaults->{soundhw};
2336     #push @$cmd, '-soundhw', 'es1370';
2337     #push @$cmd, '-soundhw', $soundhw if $soundhw;
2338
2339     if($conf->{agent}) {
2340         my $qgasocket = qga_socket($vmid);
2341         my $pciaddr = print_pci_addr("qga0", $bridges);
2342         push @$devices, '-chardev', "socket,path=$qgasocket,server,nowait,id=qga0";
2343         push @$devices, '-device', "virtio-serial,id=qga0$pciaddr";
2344         push @$devices, '-device', 'virtserialport,chardev=qga0,name=org.qemu.guest_agent.0';
2345     }
2346
2347     # enable balloon by default, unless explicitly disabled
2348     if (!defined($conf->{balloon}) || $conf->{balloon}) {
2349         $pciaddr = print_pci_addr("balloon0", $bridges);
2350         push @$devices, '-device', "virtio-balloon-pci,id=balloon0$pciaddr";
2351     }
2352
2353     if ($conf->{watchdog}) {
2354         my $wdopts = parse_watchdog($conf->{watchdog});
2355         $pciaddr = print_pci_addr("watchdog", $bridges);
2356         my $watchdog = $wdopts->{model} || 'i6300esb';
2357         push @$devices, '-device', "$watchdog$pciaddr";
2358         push @$devices, '-watchdog-action', $wdopts->{action} if $wdopts->{action};
2359     }
2360
2361     my $vollist = [];
2362     my $scsicontroller = {};
2363     my $ahcicontroller = {};
2364     my $scsihw = defined($conf->{scsihw}) ? $conf->{scsihw} : $defaults->{scsihw};
2365
2366     foreach_drive($conf, sub {
2367         my ($ds, $drive) = @_;
2368
2369         if (PVE::Storage::parse_volume_id($drive->{file}, 1)) {
2370             push @$vollist, $drive->{file};
2371         }
2372
2373         $use_virtio = 1 if $ds =~ m/^virtio/;
2374
2375         if (drive_is_cdrom ($drive)) {
2376             if ($bootindex_hash->{d}) {
2377                 $drive->{bootindex} = $bootindex_hash->{d};
2378                 $bootindex_hash->{d} += 1;
2379             }
2380         } else {
2381             if ($bootindex_hash->{c}) {
2382                 $drive->{bootindex} = $bootindex_hash->{c} if $conf->{bootdisk} && ($conf->{bootdisk} eq $ds);
2383                 $bootindex_hash->{c} += 1;
2384             }
2385         }
2386
2387         if ($drive->{interface} eq 'scsi') {
2388
2389             my $maxdev = ($scsihw ne 'lsi') ? 256 : 7;
2390             my $controller = int($drive->{index} / $maxdev);
2391             $pciaddr = print_pci_addr("scsihw$controller", $bridges);
2392             push @$devices, '-device', "$scsihw,id=scsihw$controller$pciaddr" if !$scsicontroller->{$controller};
2393             $scsicontroller->{$controller}=1;
2394         }
2395
2396         if ($drive->{interface} eq 'sata') {
2397            my $controller = int($drive->{index} / $MAX_SATA_DISKS);
2398            $pciaddr = print_pci_addr("ahci$controller", $bridges);
2399            push @$devices, '-device', "ahci,id=ahci$controller,multifunction=on$pciaddr" if !$ahcicontroller->{$controller};
2400            $ahcicontroller->{$controller}=1;
2401         }
2402
2403         push @$devices, '-drive',print_drive_full($storecfg, $vmid, $drive);
2404         push @$devices, '-device',print_drivedevice_full($storecfg, $conf, $vmid, $drive, $bridges);
2405     });
2406
2407     push @$cmd, '-m', $conf->{memory} || $defaults->{memory};
2408
2409     for (my $i = 0; $i < $MAX_NETS; $i++) {
2410          next if !$conf->{"net$i"};
2411          my $d = parse_net($conf->{"net$i"});
2412          next if !$d;
2413
2414          $use_virtio = 1 if $d->{model} eq 'virtio';
2415
2416          if ($bootindex_hash->{n}) {
2417             $d->{bootindex} = $bootindex_hash->{n};
2418             $bootindex_hash->{n} += 1;
2419          }
2420
2421          my $netdevfull = print_netdev_full($vmid,$conf,$d,"net$i");
2422          push @$devices, '-netdev', $netdevfull;
2423
2424          my $netdevicefull = print_netdevice_full($vmid,$conf,$d,"net$i",$bridges);
2425          push @$devices, '-device', $netdevicefull;
2426     }
2427
2428     #bridges
2429     while (my ($k, $v) = each %$bridges) {
2430         $pciaddr = print_pci_addr("pci.$k");
2431         unshift @$devices, '-device', "pci-bridge,id=pci.$k,chassis_nr=$k$pciaddr" if $k > 0;
2432     }
2433
2434
2435     # hack: virtio with fairsched is unreliable, so we do not use fairsched
2436     # when the VM uses virtio devices.
2437     if (!$use_virtio && $have_ovz) {
2438
2439         my $cpuunits = defined($conf->{cpuunits}) ?
2440             $conf->{cpuunits} : $defaults->{cpuunits};
2441
2442         push @$cmd, '-cpuunits', $cpuunits if $cpuunits;
2443
2444         # fixme: cpulimit is currently ignored
2445         #push @$cmd, '-cpulimit', $conf->{cpulimit} if $conf->{cpulimit};
2446     }
2447
2448     # add custom args
2449     if ($conf->{args}) {
2450         my $aa = PVE::Tools::split_args($conf->{args});
2451         push @$cmd, @$aa;
2452     }
2453
2454     push @$cmd, @$devices;
2455     push @$cmd, '-rtc', join(',', @$rtcFlags)
2456         if scalar(@$rtcFlags);
2457     push @$cmd, '-machine', join(',', @$machineFlags)
2458         if scalar(@$machineFlags);
2459     push @$cmd, '-global', join(',', @$globalFlags)
2460         if scalar(@$globalFlags);
2461
2462     return wantarray ? ($cmd, $vollist) : $cmd;
2463 }
2464
2465 sub vnc_socket {
2466     my ($vmid) = @_;
2467     return "${var_run_tmpdir}/$vmid.vnc";
2468 }
2469
2470 sub qmp_socket {
2471     my ($vmid) = @_;
2472     return "${var_run_tmpdir}/$vmid.qmp";
2473 }
2474
2475 sub qga_socket {
2476     my ($vmid) = @_;
2477     return "${var_run_tmpdir}/$vmid.qga";
2478 }
2479
2480 sub pidfile_name {
2481     my ($vmid) = @_;
2482     return "${var_run_tmpdir}/$vmid.pid";
2483 }
2484
2485 sub next_migrate_port {
2486
2487     for (my $p = 60000; $p < 60010; $p++) {
2488
2489         my $sock = IO::Socket::INET->new(Listen => 5,
2490                                          LocalAddr => 'localhost',
2491                                          LocalPort => $p,
2492                                          ReuseAddr => 1,
2493                                          Proto     => 0);
2494
2495         if ($sock) {
2496             close($sock);
2497             return $p;
2498         }
2499     }
2500
2501     die "unable to find free migration port";
2502 }
2503
2504 sub vm_devices_list {
2505     my ($vmid) = @_;
2506
2507     my $res = vm_mon_cmd($vmid, 'query-pci');
2508
2509     my $devices = {};
2510     foreach my $pcibus (@$res) {
2511         foreach my $device (@{$pcibus->{devices}}) {
2512             next if !$device->{'qdev_id'};
2513             $devices->{$device->{'qdev_id'}} = $device;
2514         }
2515     }
2516
2517     return $devices;
2518 }
2519
2520 sub vm_deviceplug {
2521     my ($storecfg, $conf, $vmid, $deviceid, $device) = @_;
2522
2523     return 1 if !check_running($vmid) || !$conf->{hotplug};
2524
2525     my $devices_list = vm_devices_list($vmid);
2526     return 1 if defined($devices_list->{$deviceid});
2527
2528     qemu_bridgeadd($storecfg, $conf, $vmid, $deviceid); #add bridge if we need it for the device
2529
2530     if ($deviceid =~ m/^(virtio)(\d+)$/) {
2531         return undef if !qemu_driveadd($storecfg, $vmid, $device);
2532         my $devicefull = print_drivedevice_full($storecfg, $conf, $vmid, $device);
2533         qemu_deviceadd($vmid, $devicefull);
2534         if(!qemu_deviceaddverify($vmid, $deviceid)) {
2535            qemu_drivedel($vmid, $deviceid);
2536            return undef;
2537         }
2538     }
2539
2540     if ($deviceid =~ m/^(scsihw)(\d+)$/) {
2541         my $scsihw = defined($conf->{scsihw}) ? $conf->{scsihw} : "lsi";
2542         my $pciaddr = print_pci_addr($deviceid);
2543         my $devicefull = "$scsihw,id=$deviceid$pciaddr";
2544         qemu_deviceadd($vmid, $devicefull);
2545         return undef if(!qemu_deviceaddverify($vmid, $deviceid));
2546     }
2547
2548     if ($deviceid =~ m/^(scsi)(\d+)$/) {
2549         return 1 if ($conf->{scsihw} && $conf->{scsihw} ne 'lsi'); #virtio-scsi not yet support hotplug
2550         return undef if !qemu_findorcreatescsihw($storecfg,$conf, $vmid, $device);
2551         return undef if !qemu_driveadd($storecfg, $vmid, $device);
2552         my $devicefull = print_drivedevice_full($storecfg, $conf, $vmid, $device);
2553         if(!qemu_deviceadd($vmid, $devicefull)) {
2554            qemu_drivedel($vmid, $deviceid);
2555            return undef;
2556         }
2557     }
2558
2559     if ($deviceid =~ m/^(net)(\d+)$/) {
2560         return undef if !qemu_netdevadd($vmid, $conf, $device, $deviceid);
2561         my $netdevicefull = print_netdevice_full($vmid, $conf, $device, $deviceid);
2562         qemu_deviceadd($vmid, $netdevicefull);
2563         if(!qemu_deviceaddverify($vmid, $deviceid)) {
2564            qemu_netdevdel($vmid, $deviceid);
2565            return undef;
2566         }
2567     }
2568
2569     if ($deviceid =~ m/^(pci\.)(\d+)$/) {
2570         my $bridgeid = $2;
2571         my $pciaddr = print_pci_addr($deviceid);
2572         my $devicefull = "pci-bridge,id=pci.$bridgeid,chassis_nr=$bridgeid$pciaddr";
2573         qemu_deviceadd($vmid, $devicefull);
2574         return undef if !qemu_deviceaddverify($vmid, $deviceid);
2575     }
2576
2577     return 1;
2578 }
2579
2580 sub vm_deviceunplug {
2581     my ($vmid, $conf, $deviceid) = @_;
2582
2583     return 1 if !check_running ($vmid) || !$conf->{hotplug};
2584
2585     my $devices_list = vm_devices_list($vmid);
2586     return 1 if !defined($devices_list->{$deviceid});
2587
2588     die "can't unplug bootdisk" if $conf->{bootdisk} && $conf->{bootdisk} eq $deviceid;
2589
2590     if ($deviceid =~ m/^(virtio)(\d+)$/) {
2591         return undef if !qemu_drivedel($vmid, $deviceid);
2592         qemu_devicedel($vmid, $deviceid);
2593         return undef if !qemu_devicedelverify($vmid, $deviceid);
2594     }
2595
2596     if ($deviceid =~ m/^(lsi)(\d+)$/) {
2597         return undef if !qemu_devicedel($vmid, $deviceid);
2598     }
2599
2600     if ($deviceid =~ m/^(scsi)(\d+)$/) {
2601         return undef if !qemu_devicedel($vmid, $deviceid);
2602         return undef if !qemu_drivedel($vmid, $deviceid);
2603     }
2604
2605     if ($deviceid =~ m/^(net)(\d+)$/) {
2606         return undef if !qemu_netdevdel($vmid, $deviceid);
2607         qemu_devicedel($vmid, $deviceid);
2608         return undef if !qemu_devicedelverify($vmid, $deviceid);
2609     }
2610
2611     return 1;
2612 }
2613
2614 sub qemu_deviceadd {
2615     my ($vmid, $devicefull) = @_;
2616
2617     my $ret = vm_human_monitor_command($vmid, "device_add $devicefull");
2618     $ret =~ s/^\s+//;
2619     # Otherwise, if the command succeeds, no output is sent. So any non-empty string shows an error
2620     return 1 if $ret eq "";
2621     syslog("err", "error on hotplug device : $ret");
2622     return undef;
2623
2624 }
2625
2626 sub qemu_devicedel {
2627     my($vmid, $deviceid) = @_;
2628
2629     my $ret = vm_human_monitor_command($vmid, "device_del $deviceid");
2630     $ret =~ s/^\s+//;
2631     return 1 if $ret eq "";
2632     syslog("err", "detaching device $deviceid failed : $ret");
2633     return undef;
2634 }
2635
2636 sub qemu_driveadd {
2637     my($storecfg, $vmid, $device) = @_;
2638
2639     my $drive = print_drive_full($storecfg, $vmid, $device);
2640     my $ret = vm_human_monitor_command($vmid, "drive_add auto $drive");
2641     # If the command succeeds qemu prints: "OK"
2642     if ($ret !~ m/OK/s) {
2643         syslog("err", "adding drive failed: $ret");
2644         return undef;
2645     }
2646     return 1;
2647 }
2648
2649 sub qemu_drivedel {
2650     my($vmid, $deviceid) = @_;
2651
2652     my $ret = vm_human_monitor_command($vmid, "drive_del drive-$deviceid");
2653     $ret =~ s/^\s+//;
2654     if ($ret =~ m/Device \'.*?\' not found/s) {
2655         # NB: device not found errors mean the drive was auto-deleted and we ignore the error
2656     }
2657     elsif ($ret ne "") {
2658       syslog("err", "deleting drive $deviceid failed : $ret");
2659       return undef;
2660     }
2661     return 1;
2662 }
2663
2664 sub qemu_deviceaddverify {
2665     my ($vmid,$deviceid) = @_;
2666
2667     for (my $i = 0; $i <= 5; $i++) {
2668          my $devices_list = vm_devices_list($vmid);
2669          return 1 if defined($devices_list->{$deviceid});
2670          sleep 1;
2671     }
2672     syslog("err", "error on hotplug device $deviceid");
2673     return undef;
2674 }
2675
2676
2677 sub qemu_devicedelverify {
2678     my ($vmid,$deviceid) = @_;
2679
2680     #need to verify the device is correctly remove as device_del is async and empty return is not reliable
2681     for (my $i = 0; $i <= 5; $i++) {
2682          my $devices_list = vm_devices_list($vmid);
2683          return 1 if !defined($devices_list->{$deviceid});
2684          sleep 1;
2685     }
2686     syslog("err", "error on hot-unplugging device $deviceid");
2687     return undef;
2688 }
2689
2690 sub qemu_findorcreatescsihw {
2691     my ($storecfg, $conf, $vmid, $device) = @_;
2692
2693     my $maxdev = ($conf->{scsihw} && $conf->{scsihw} ne 'lsi') ? 256 : 7;
2694     my $controller = int($device->{index} / $maxdev);
2695     my $scsihwid="scsihw$controller";
2696     my $devices_list = vm_devices_list($vmid);
2697
2698     if(!defined($devices_list->{$scsihwid})) {
2699        return undef if !vm_deviceplug($storecfg, $conf, $vmid, $scsihwid);
2700     }
2701     return 1;
2702 }
2703
2704 sub qemu_bridgeadd {
2705     my ($storecfg, $conf, $vmid, $device) = @_;
2706
2707     my $bridges = {};
2708     my $bridgeid = undef;
2709     print_pci_addr($device, $bridges);
2710
2711     while (my ($k, $v) = each %$bridges) {
2712         $bridgeid = $k;
2713     }
2714     return if $bridgeid < 1;
2715     my $bridge = "pci.$bridgeid";
2716     my $devices_list = vm_devices_list($vmid);
2717
2718     if(!defined($devices_list->{$bridge})) {
2719         return undef if !vm_deviceplug($storecfg, $conf, $vmid, $bridge);
2720     }
2721     return 1;
2722 }
2723
2724 sub qemu_netdevadd {
2725     my ($vmid, $conf, $device, $deviceid) = @_;
2726
2727     my $netdev = print_netdev_full($vmid, $conf, $device, $deviceid);
2728     my $ret = vm_human_monitor_command($vmid, "netdev_add $netdev");
2729     $ret =~ s/^\s+//;
2730
2731     #if the command succeeds, no output is sent. So any non-empty string shows an error
2732     return 1 if $ret eq "";
2733     syslog("err", "adding netdev failed: $ret");
2734     return undef;
2735 }
2736
2737 sub qemu_netdevdel {
2738     my ($vmid, $deviceid) = @_;
2739
2740     my $ret = vm_human_monitor_command($vmid, "netdev_del $deviceid");
2741     $ret =~ s/^\s+//;
2742     #if the command succeeds, no output is sent. So any non-empty string shows an error
2743     return 1 if $ret eq "";
2744     syslog("err", "deleting netdev failed: $ret");
2745     return undef;
2746 }
2747
2748 sub qemu_block_set_io_throttle {
2749     my ($vmid, $deviceid, $bps, $bps_rd, $bps_wr, $iops, $iops_rd, $iops_wr) = @_;
2750
2751     return if !check_running($vmid) ;
2752
2753     $bps = 0 if !$bps;
2754     $bps_rd = 0 if !$bps_rd;
2755     $bps_wr = 0 if !$bps_wr;
2756     $iops = 0 if !$iops;
2757     $iops_rd = 0 if !$iops_rd;
2758     $iops_wr = 0 if !$iops_wr;
2759
2760     vm_mon_cmd($vmid, "block_set_io_throttle", device => $deviceid, bps => int($bps), bps_rd => int($bps_rd), bps_wr => int($bps_wr), iops => int($iops), iops_rd => int($iops_rd), iops_wr => int($iops_wr));
2761
2762 }
2763
2764 # old code, only used to shutdown old VM after update
2765 sub __read_avail {
2766     my ($fh, $timeout) = @_;
2767
2768     my $sel = new IO::Select;
2769     $sel->add($fh);
2770
2771     my $res = '';
2772     my $buf;
2773
2774     my @ready;
2775     while (scalar (@ready = $sel->can_read($timeout))) {
2776         my $count;
2777         if ($count = $fh->sysread($buf, 8192)) {
2778             if ($buf =~ /^(.*)\(qemu\) $/s) {
2779                 $res .= $1;
2780                 last;
2781             } else {
2782                 $res .= $buf;
2783             }
2784         } else {
2785             if (!defined($count)) {
2786                 die "$!\n";
2787             }
2788             last;
2789         }
2790     }
2791
2792     die "monitor read timeout\n" if !scalar(@ready);
2793
2794     return $res;
2795 }
2796
2797 # old code, only used to shutdown old VM after update
2798 sub vm_monitor_command {
2799     my ($vmid, $cmdstr, $nocheck) = @_;
2800
2801     my $res;
2802
2803     eval {
2804         die "VM $vmid not running\n" if !check_running($vmid, $nocheck);
2805
2806         my $sname = "${var_run_tmpdir}/$vmid.mon";
2807
2808         my $sock = IO::Socket::UNIX->new( Peer => $sname ) ||
2809             die "unable to connect to VM $vmid socket - $!\n";
2810
2811         my $timeout = 3;
2812
2813         # hack: migrate sometime blocks the monitor (when migrate_downtime
2814         # is set)
2815         if ($cmdstr =~ m/^(info\s+migrate|migrate\s)/) {
2816             $timeout = 60*60; # 1 hour
2817         }
2818
2819         # read banner;
2820         my $data = __read_avail($sock, $timeout);
2821
2822         if ($data !~ m/^QEMU\s+(\S+)\s+monitor\s/) {
2823             die "got unexpected qemu monitor banner\n";
2824         }
2825
2826         my $sel = new IO::Select;
2827         $sel->add($sock);
2828
2829         if (!scalar(my @ready = $sel->can_write($timeout))) {
2830             die "monitor write error - timeout";
2831         }
2832
2833         my $fullcmd = "$cmdstr\r";
2834
2835         # syslog('info', "VM $vmid monitor command: $cmdstr");
2836
2837         my $b;
2838         if (!($b = $sock->syswrite($fullcmd)) || ($b != length($fullcmd))) {
2839             die "monitor write error - $!";
2840         }
2841
2842         return if ($cmdstr eq 'q') || ($cmdstr eq 'quit');
2843
2844         $timeout = 20;
2845
2846         if ($cmdstr =~ m/^(info\s+migrate|migrate\s)/) {
2847             $timeout = 60*60; # 1 hour
2848         } elsif ($cmdstr =~ m/^(eject|change)/) {
2849             $timeout = 60; # note: cdrom mount command is slow
2850         }
2851         if ($res = __read_avail($sock, $timeout)) {
2852
2853             my @lines = split("\r?\n", $res);
2854
2855             shift @lines if $lines[0] !~ m/^unknown command/; # skip echo
2856
2857             $res = join("\n", @lines);
2858             $res .= "\n";
2859         }
2860     };
2861
2862     my $err = $@;
2863
2864     if ($err) {
2865         syslog("err", "VM $vmid monitor command failed - $err");
2866         die $err;
2867     }
2868
2869     return $res;
2870 }
2871
2872 sub qemu_block_resize {
2873     my ($vmid, $deviceid, $storecfg, $volid, $size) = @_;
2874
2875     my $running = PVE::QemuServer::check_running($vmid);
2876
2877     return if !PVE::Storage::volume_resize($storecfg, $volid, $size, $running);
2878
2879     return if !$running;
2880
2881     vm_mon_cmd($vmid, "block_resize", device => $deviceid, size => int($size));
2882
2883 }
2884
2885 sub qemu_volume_snapshot {
2886     my ($vmid, $deviceid, $storecfg, $volid, $snap) = @_;
2887
2888     my $running = PVE::QemuServer::check_running($vmid);
2889
2890     return if !PVE::Storage::volume_snapshot($storecfg, $volid, $snap, $running);
2891
2892     return if !$running;
2893
2894     vm_mon_cmd($vmid, "snapshot-drive", device => $deviceid, name => $snap);
2895
2896 }
2897
2898 sub qemu_volume_snapshot_delete {
2899     my ($vmid, $deviceid, $storecfg, $volid, $snap) = @_;
2900
2901     my $running = PVE::QemuServer::check_running($vmid);
2902
2903     return if !PVE::Storage::volume_snapshot_delete($storecfg, $volid, $snap, $running);
2904
2905     return if !$running;
2906
2907     vm_mon_cmd($vmid, "delete-drive-snapshot", device => $deviceid, name => $snap);
2908 }
2909
2910 sub qga_freezefs {
2911     my ($vmid) = @_;
2912
2913     #need to impplement call to qemu-ga
2914 }
2915
2916 sub qga_unfreezefs {
2917     my ($vmid) = @_;
2918
2919     #need to impplement call to qemu-ga
2920 }
2921
2922 sub vm_start {
2923     my ($storecfg, $vmid, $statefile, $skiplock, $migratedfrom, $paused) = @_;
2924
2925     lock_config($vmid, sub {
2926         my $conf = load_config($vmid, $migratedfrom);
2927
2928         check_lock($conf) if !$skiplock;
2929
2930         die "VM $vmid already running\n" if check_running($vmid, undef, $migratedfrom);
2931
2932         my $defaults = load_defaults();
2933
2934         # set environment variable useful inside network script
2935         $ENV{PVE_MIGRATED_FROM} = $migratedfrom if $migratedfrom;
2936
2937         my ($cmd, $vollist) = config_to_command($storecfg, $vmid, $conf, $defaults);
2938
2939         my $migrate_port = 0;
2940
2941         if ($statefile) {
2942             if ($statefile eq 'tcp') {
2943                 $migrate_port = next_migrate_port();
2944                 my $migrate_uri = "tcp:localhost:${migrate_port}";
2945                 push @$cmd, '-incoming', $migrate_uri;
2946                 push @$cmd, '-S';
2947             } else {
2948                 push @$cmd, '-loadstate', $statefile;
2949             }
2950         } elsif ($paused) {
2951             push @$cmd, '-S';
2952         }
2953
2954         # host pci devices
2955         for (my $i = 0; $i < $MAX_HOSTPCI_DEVICES; $i++)  {
2956           my $d = parse_hostpci($conf->{"hostpci$i"});
2957           next if !$d;
2958           my $info = pci_device_info("0000:$d->{pciid}");
2959           die "IOMMU not present\n" if !check_iommu_support();
2960           die "no pci device info for device '$d->{pciid}'\n" if !$info;
2961           die "can't unbind pci device '$d->{pciid}'\n" if !pci_dev_bind_to_stub($info);
2962           die "can't reset pci device '$d->{pciid}'\n" if !pci_dev_reset($info);
2963         }
2964
2965         PVE::Storage::activate_volumes($storecfg, $vollist);
2966
2967         eval  { run_command($cmd, timeout => $statefile ? undef : 30,
2968                     umask => 0077); };
2969         my $err = $@;
2970         die "start failed: $err" if $err;
2971
2972         print "migration listens on port $migrate_port\n" if $migrate_port;
2973
2974         if ($statefile && $statefile ne 'tcp')  {
2975             eval { vm_mon_cmd_nocheck($vmid, "cont"); };
2976             warn $@ if $@;
2977         }
2978
2979         if($migratedfrom) {
2980             my $capabilities = {};
2981             $capabilities->{capability} =  "xbzrle";
2982             $capabilities->{state} = JSON::true;
2983             eval { PVE::QemuServer::vm_mon_cmd_nocheck($vmid, "migrate-set-capabilities", capabilities => [$capabilities]); };
2984         }
2985
2986         # fixme: how do we handle that on migration?
2987
2988         if (!defined($conf->{balloon}) || $conf->{balloon}) {
2989             vm_mon_cmd_nocheck($vmid, "balloon", value => $conf->{balloon}*1024*1024)
2990                 if $conf->{balloon};
2991             vm_mon_cmd_nocheck($vmid, 'qom-set',
2992                        path => "machine/peripheral/balloon0",
2993                        property => "stats-polling-interval",
2994                        value => 2);
2995         }
2996     });
2997 }
2998
2999 sub vm_mon_cmd {
3000     my ($vmid, $execute, %params) = @_;
3001
3002     my $cmd = { execute => $execute, arguments => \%params };
3003     vm_qmp_command($vmid, $cmd);
3004 }
3005
3006 sub vm_mon_cmd_nocheck {
3007     my ($vmid, $execute, %params) = @_;
3008
3009     my $cmd = { execute => $execute, arguments => \%params };
3010     vm_qmp_command($vmid, $cmd, 1);
3011 }
3012
3013 sub vm_qmp_command {
3014     my ($vmid, $cmd, $nocheck) = @_;
3015
3016     my $res;
3017
3018     my $timeout;
3019     if ($cmd->{arguments} && $cmd->{arguments}->{timeout}) {
3020         $timeout = $cmd->{arguments}->{timeout};
3021         delete $cmd->{arguments}->{timeout};
3022     }
3023
3024     eval {
3025         die "VM $vmid not running\n" if !check_running($vmid, $nocheck);
3026         my $sname = PVE::QemuServer::qmp_socket($vmid);
3027         if (-e $sname) {
3028             my $qmpclient = PVE::QMPClient->new();
3029
3030             $res = $qmpclient->cmd($vmid, $cmd, $timeout);
3031         } elsif (-e "${var_run_tmpdir}/$vmid.mon") {
3032             die "can't execute complex command on old monitor - stop/start your vm to fix the problem\n"
3033                 if scalar(%{$cmd->{arguments}});
3034             vm_monitor_command($vmid, $cmd->{execute}, $nocheck);
3035         } else {
3036             die "unable to open monitor socket\n";
3037         }
3038     };
3039     if (my $err = $@) {
3040         syslog("err", "VM $vmid qmp command failed - $err");
3041         die $err;
3042     }
3043
3044     return $res;
3045 }
3046
3047 sub vm_human_monitor_command {
3048     my ($vmid, $cmdline) = @_;
3049
3050     my $res;
3051
3052     my $cmd = {
3053         execute => 'human-monitor-command',
3054         arguments => { 'command-line' => $cmdline},
3055     };
3056
3057     return vm_qmp_command($vmid, $cmd);
3058 }
3059
3060 sub vm_commandline {
3061     my ($storecfg, $vmid) = @_;
3062
3063     my $conf = load_config($vmid);
3064
3065     my $defaults = load_defaults();
3066
3067     my $cmd = config_to_command($storecfg, $vmid, $conf, $defaults);
3068
3069     return join(' ', @$cmd);
3070 }
3071
3072 sub vm_reset {
3073     my ($vmid, $skiplock) = @_;
3074
3075     lock_config($vmid, sub {
3076
3077         my $conf = load_config($vmid);
3078
3079         check_lock($conf) if !$skiplock;
3080
3081         vm_mon_cmd($vmid, "system_reset");
3082     });
3083 }
3084
3085 sub get_vm_volumes {
3086     my ($conf) = @_;
3087
3088     my $vollist = [];
3089     foreach_volid($conf, sub {
3090         my ($volid, $is_cdrom) = @_;
3091
3092         return if $volid =~ m|^/|;
3093
3094         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
3095         return if !$sid;
3096
3097         push @$vollist, $volid;
3098     });
3099
3100     return $vollist;
3101 }
3102
3103 sub vm_stop_cleanup {
3104     my ($storecfg, $vmid, $conf, $keepActive) = @_;
3105
3106     eval {
3107         fairsched_rmnod($vmid); # try to destroy group
3108
3109         if (!$keepActive) {
3110             my $vollist = get_vm_volumes($conf);
3111             PVE::Storage::deactivate_volumes($storecfg, $vollist);
3112         }
3113
3114         foreach my $ext (qw(mon qmp pid vnc qga)) {
3115             unlink "/var/run/qemu-server/${vmid}.$ext";
3116         }
3117     };
3118     warn $@ if $@; # avoid errors - just warn
3119 }
3120
3121 # Note: use $nockeck to skip tests if VM configuration file exists.
3122 # We need that when migration VMs to other nodes (files already moved)
3123 # Note: we set $keepActive in vzdump stop mode - volumes need to stay active
3124 sub vm_stop {
3125     my ($storecfg, $vmid, $skiplock, $nocheck, $timeout, $shutdown, $force, $keepActive, $migratedfrom) = @_;
3126
3127     $force = 1 if !defined($force) && !$shutdown;
3128
3129     if ($migratedfrom){
3130         my $pid = check_running($vmid, $nocheck, $migratedfrom);
3131         kill 15, $pid if $pid;
3132         my $conf = load_config($vmid, $migratedfrom);
3133         vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive);
3134         return;
3135     }
3136
3137     lock_config($vmid, sub {
3138
3139         my $pid = check_running($vmid, $nocheck);
3140         return if !$pid;
3141
3142         my $conf;
3143         if (!$nocheck) {
3144             $conf = load_config($vmid);
3145             check_lock($conf) if !$skiplock;
3146             if (!defined($timeout) && $shutdown && $conf->{startup}) {
3147                 my $opts = parse_startup($conf->{startup});
3148                 $timeout = $opts->{down} if $opts->{down};
3149             }
3150         }
3151
3152         $timeout = 60 if !defined($timeout);
3153
3154         eval {
3155             if ($shutdown) {
3156                 $nocheck ? vm_mon_cmd_nocheck($vmid, "system_powerdown") : vm_mon_cmd($vmid, "system_powerdown");
3157
3158             } else {
3159                 $nocheck ? vm_mon_cmd_nocheck($vmid, "quit") : vm_mon_cmd($vmid, "quit");
3160             }
3161         };
3162         my $err = $@;
3163
3164         if (!$err) {
3165             my $count = 0;
3166             while (($count < $timeout) && check_running($vmid, $nocheck)) {
3167                 $count++;
3168                 sleep 1;
3169             }
3170
3171             if ($count >= $timeout) {
3172                 if ($force) {
3173                     warn "VM still running - terminating now with SIGTERM\n";
3174                     kill 15, $pid;
3175                 } else {
3176                     die "VM quit/powerdown failed - got timeout\n";
3177                 }
3178             } else {
3179                 vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive) if $conf;
3180                 return;
3181             }
3182         } else {
3183             if ($force) {
3184                 warn "VM quit/powerdown failed - terminating now with SIGTERM\n";
3185                 kill 15, $pid;
3186             } else {
3187                 die "VM quit/powerdown failed\n";
3188             }
3189         }
3190
3191         # wait again
3192         $timeout = 10;
3193
3194         my $count = 0;
3195         while (($count < $timeout) && check_running($vmid, $nocheck)) {
3196             $count++;
3197             sleep 1;
3198         }
3199
3200         if ($count >= $timeout) {
3201             warn "VM still running - terminating now with SIGKILL\n";
3202             kill 9, $pid;
3203             sleep 1;
3204         }
3205
3206         vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive) if $conf;
3207    });
3208 }
3209
3210 sub vm_suspend {
3211     my ($vmid, $skiplock) = @_;
3212
3213     lock_config($vmid, sub {
3214
3215         my $conf = load_config($vmid);
3216
3217         check_lock($conf) if !$skiplock;
3218
3219         vm_mon_cmd($vmid, "stop");
3220     });
3221 }
3222
3223 sub vm_resume {
3224     my ($vmid, $skiplock) = @_;
3225
3226     lock_config($vmid, sub {
3227
3228         my $conf = load_config($vmid);
3229
3230         check_lock($conf) if !$skiplock;
3231
3232         vm_mon_cmd($vmid, "cont");
3233     });
3234 }
3235
3236 sub vm_sendkey {
3237     my ($vmid, $skiplock, $key) = @_;
3238
3239     lock_config($vmid, sub {
3240
3241         my $conf = load_config($vmid);
3242
3243         # there is no qmp command, so we use the human monitor command
3244         vm_human_monitor_command($vmid, "sendkey $key");
3245     });
3246 }
3247
3248 sub vm_destroy {
3249     my ($storecfg, $vmid, $skiplock) = @_;
3250
3251     lock_config($vmid, sub {
3252
3253         my $conf = load_config($vmid);
3254
3255         check_lock($conf) if !$skiplock;
3256
3257         if (!check_running($vmid)) {
3258             fairsched_rmnod($vmid); # try to destroy group
3259             destroy_vm($storecfg, $vmid);
3260         } else {
3261             die "VM $vmid is running - destroy failed\n";
3262         }
3263     });
3264 }
3265
3266 # pci helpers
3267
3268 sub file_write {
3269     my ($filename, $buf) = @_;
3270
3271     my $fh = IO::File->new($filename, "w");
3272     return undef if !$fh;
3273
3274     my $res = print $fh $buf;
3275
3276     $fh->close();
3277
3278     return $res;
3279 }
3280
3281 sub pci_device_info {
3282     my ($name) = @_;
3283
3284     my $res;
3285
3286     return undef if $name !~ m/^([a-f0-9]{4}):([a-f0-9]{2}):([a-f0-9]{2})\.([a-f0-9])$/;
3287     my ($domain, $bus, $slot, $func) = ($1, $2, $3, $4);
3288
3289     my $irq = file_read_firstline("$pcisysfs/devices/$name/irq");
3290     return undef if !defined($irq) || $irq !~ m/^\d+$/;
3291
3292     my $vendor = file_read_firstline("$pcisysfs/devices/$name/vendor");
3293     return undef if !defined($vendor) || $vendor !~ s/^0x//;
3294
3295     my $product = file_read_firstline("$pcisysfs/devices/$name/device");
3296     return undef if !defined($product) || $product !~ s/^0x//;
3297
3298     $res = {
3299         name => $name,
3300         vendor => $vendor,
3301         product => $product,
3302         domain => $domain,
3303         bus => $bus,
3304         slot => $slot,
3305         func => $func,
3306         irq => $irq,
3307         has_fl_reset => -f "$pcisysfs/devices/$name/reset" || 0,
3308     };
3309
3310     return $res;
3311 }
3312
3313 sub pci_dev_reset {
3314     my ($dev) = @_;
3315
3316     my $name = $dev->{name};
3317
3318     my $fn = "$pcisysfs/devices/$name/reset";
3319
3320     return file_write($fn, "1");
3321 }
3322
3323 sub pci_dev_bind_to_stub {
3324     my ($dev) = @_;
3325
3326     my $name = $dev->{name};
3327
3328     my $testdir = "$pcisysfs/drivers/pci-stub/$name";
3329     return 1 if -d $testdir;
3330
3331     my $data = "$dev->{vendor} $dev->{product}";
3332     return undef if !file_write("$pcisysfs/drivers/pci-stub/new_id", $data);
3333
3334     my $fn = "$pcisysfs/devices/$name/driver/unbind";
3335     if (!file_write($fn, $name)) {
3336         return undef if -f $fn;
3337     }
3338
3339     $fn = "$pcisysfs/drivers/pci-stub/bind";
3340     if (! -d $testdir) {
3341         return undef if !file_write($fn, $name);
3342     }
3343
3344     return -d $testdir;
3345 }
3346
3347 sub print_pci_addr {
3348     my ($id, $bridges) = @_;
3349
3350     my $res = '';
3351     my $devices = {
3352         #addr1 : ide,parallel,serial (motherboard)
3353         #addr2 : first videocard
3354         balloon0 => { bus => 0, addr => 3 },
3355         watchdog => { bus => 0, addr => 4 },
3356         scsihw0 => { bus => 0, addr => 5 },
3357         scsihw1 => { bus => 0, addr => 6 },
3358         ahci0 => { bus => 0, addr => 7 },
3359         qga0 => { bus => 0, addr => 8 },
3360         virtio0 => { bus => 0, addr => 10 },
3361         virtio1 => { bus => 0, addr => 11 },
3362         virtio2 => { bus => 0, addr => 12 },
3363         virtio3 => { bus => 0, addr => 13 },
3364         virtio4 => { bus => 0, addr => 14 },
3365         virtio5 => { bus => 0, addr => 15 },
3366         hostpci0 => { bus => 0, addr => 16 },
3367         hostpci1 => { bus => 0, addr => 17 },
3368         net0 => { bus => 0, addr => 18 },
3369         net1 => { bus => 0, addr => 19 },
3370         net2 => { bus => 0, addr => 20 },
3371         net3 => { bus => 0, addr => 21 },
3372         net4 => { bus => 0, addr => 22 },
3373         net5 => { bus => 0, addr => 23 },
3374         #addr29 : usb-host (pve-usb.cfg)
3375         'pci.1' => { bus => 0, addr => 30 },
3376         'pci.2' => { bus => 0, addr => 31 },
3377         'net6' => { bus => 1, addr => 1 },
3378         'net7' => { bus => 1, addr => 2 },
3379         'net8' => { bus => 1, addr => 3 },
3380         'net9' => { bus => 1, addr => 4 },
3381         'net10' => { bus => 1, addr => 5 },
3382         'net11' => { bus => 1, addr => 6 },
3383         'net12' => { bus => 1, addr => 7 },
3384         'net13' => { bus => 1, addr => 8 },
3385         'net14' => { bus => 1, addr => 9 },
3386         'net15' => { bus => 1, addr => 10 },
3387         'net16' => { bus => 1, addr => 11 },
3388         'net17' => { bus => 1, addr => 12 },
3389         'net18' => { bus => 1, addr => 13 },
3390         'net19' => { bus => 1, addr => 14 },
3391         'net20' => { bus => 1, addr => 15 },
3392         'net21' => { bus => 1, addr => 16 },
3393         'net22' => { bus => 1, addr => 17 },
3394         'net23' => { bus => 1, addr => 18 },
3395         'net24' => { bus => 1, addr => 19 },
3396         'net25' => { bus => 1, addr => 20 },
3397         'net26' => { bus => 1, addr => 21 },
3398         'net27' => { bus => 1, addr => 22 },
3399         'net28' => { bus => 1, addr => 23 },
3400         'net29' => { bus => 1, addr => 24 },
3401         'net30' => { bus => 1, addr => 25 },
3402         'net31' => { bus => 1, addr => 26 },
3403         'virtio6' => { bus => 2, addr => 1 },
3404         'virtio7' => { bus => 2, addr => 2 },
3405         'virtio8' => { bus => 2, addr => 3 },
3406         'virtio9' => { bus => 2, addr => 4 },
3407         'virtio10' => { bus => 2, addr => 5 },
3408         'virtio11' => { bus => 2, addr => 6 },
3409         'virtio12' => { bus => 2, addr => 7 },
3410         'virtio13' => { bus => 2, addr => 8 },
3411         'virtio14' => { bus => 2, addr => 9 },
3412         'virtio15' => { bus => 2, addr => 10 },
3413     };
3414
3415     if (defined($devices->{$id}->{bus}) && defined($devices->{$id}->{addr})) {
3416            my $addr = sprintf("0x%x", $devices->{$id}->{addr});
3417            my $bus = $devices->{$id}->{bus};
3418            $res = ",bus=pci.$bus,addr=$addr";
3419            $bridges->{$bus} = 1 if $bridges;
3420     }
3421     return $res;
3422
3423 }
3424
3425 # vzdump restore implementaion
3426
3427 sub archive_read_firstfile {
3428     my $archive = shift;
3429
3430     die "ERROR: file '$archive' does not exist\n" if ! -f $archive;
3431
3432     # try to detect archive type first
3433     my $pid = open (TMP, "tar tf '$archive'|") ||
3434         die "unable to open file '$archive'\n";
3435     my $firstfile = <TMP>;
3436     kill 15, $pid;
3437     close TMP;
3438
3439     die "ERROR: archive contaions no data\n" if !$firstfile;
3440     chomp $firstfile;
3441
3442     return $firstfile;
3443 }
3444
3445 sub restore_cleanup {
3446     my $statfile = shift;
3447
3448     print STDERR "starting cleanup\n";
3449
3450     if (my $fd = IO::File->new($statfile, "r")) {
3451         while (defined(my $line = <$fd>)) {
3452             if ($line =~ m/vzdump:([^\s:]*):(\S+)$/) {
3453                 my $volid = $2;
3454                 eval {
3455                     if ($volid =~ m|^/|) {
3456                         unlink $volid || die 'unlink failed\n';
3457                     } else {
3458                         my $cfg = cfs_read_file('storage.cfg');
3459                         PVE::Storage::vdisk_free($cfg, $volid);
3460                     }
3461                     print STDERR "temporary volume '$volid' sucessfuly removed\n";
3462                 };
3463                 print STDERR "unable to cleanup '$volid' - $@" if $@;
3464             } else {
3465                 print STDERR "unable to parse line in statfile - $line";
3466             }
3467         }
3468         $fd->close();
3469     }
3470 }
3471
3472 sub restore_archive {
3473     my ($archive, $vmid, $user, $opts) = @_;
3474
3475     my $format = $opts->{format};
3476     my $comp;
3477
3478     if ($archive =~ m/\.tgz$/ || $archive =~ m/\.tar\.gz$/) {
3479         $format = 'tar' if !$format;
3480         $comp = 'gzip';
3481     } elsif ($archive =~ m/\.tar$/) {
3482         $format = 'tar' if !$format;
3483     } elsif ($archive =~ m/.tar.lzo$/) {
3484         $format = 'tar' if !$format;
3485         $comp = 'lzop';
3486     } elsif ($archive =~ m/\.vma$/) {
3487         $format = 'vma' if !$format;
3488     } elsif ($archive =~ m/\.vma\.gz$/) {
3489         $format = 'vma' if !$format;
3490         $comp = 'gzip';
3491     } elsif ($archive =~ m/\.vma\.lzo$/) {
3492         $format = 'vma' if !$format;
3493         $comp = 'lzop';
3494     } else {
3495         $format = 'vma' if !$format; # default
3496     }
3497
3498     # try to detect archive format
3499     if ($format eq 'tar') {
3500         return restore_tar_archive($archive, $vmid, $user, $opts);
3501     } else {
3502         return restore_vma_archive($archive, $vmid, $user, $opts, $comp);
3503     }
3504 }
3505
3506 sub restore_update_config_line {
3507     my ($outfd, $cookie, $vmid, $map, $line, $unique) = @_;
3508
3509     return if $line =~ m/^\#qmdump\#/;
3510     return if $line =~ m/^\#vzdump\#/;
3511     return if $line =~ m/^lock:/;
3512     return if $line =~ m/^unused\d+:/;
3513     return if $line =~ m/^parent:/;
3514
3515     if (($line =~ m/^(vlan(\d+)):\s*(\S+)\s*$/)) {
3516         # try to convert old 1.X settings
3517         my ($id, $ind, $ethcfg) = ($1, $2, $3);
3518         foreach my $devconfig (PVE::Tools::split_list($ethcfg)) {
3519             my ($model, $macaddr) = split(/\=/, $devconfig);
3520             $macaddr = PVE::Tools::random_ether_addr() if !$macaddr || $unique;
3521             my $net = {
3522                 model => $model,
3523                 bridge => "vmbr$ind",
3524                 macaddr => $macaddr,
3525             };
3526             my $netstr = print_net($net);
3527
3528             print $outfd "net$cookie->{netcount}: $netstr\n";
3529             $cookie->{netcount}++;
3530         }
3531     } elsif (($line =~ m/^(net\d+):\s*(\S+)\s*$/) && $unique) {
3532         my ($id, $netstr) = ($1, $2);
3533         my $net = parse_net($netstr);
3534         $net->{macaddr} = PVE::Tools::random_ether_addr() if $net->{macaddr};
3535         $netstr = print_net($net);
3536         print $outfd "$id: $netstr\n";
3537     } elsif ($line =~ m/^((ide|scsi|virtio|sata)\d+):\s*(\S+)\s*$/) {
3538         my $virtdev = $1;
3539         my $value = $2;
3540         if ($line =~ m/backup=no/) {
3541             print $outfd "#$line";
3542         } elsif ($virtdev && $map->{$virtdev}) {
3543             my $di = PVE::QemuServer::parse_drive($virtdev, $value);
3544             $di->{file} = $map->{$virtdev};
3545             $value = PVE::QemuServer::print_drive($vmid, $di);
3546             print $outfd "$virtdev: $value\n";
3547         } else {
3548             print $outfd $line;
3549         }
3550     } else {
3551         print $outfd $line;
3552     }
3553 }
3554
3555 sub scan_volids {
3556     my ($cfg, $vmid) = @_;
3557
3558     my $info = PVE::Storage::vdisk_list($cfg, undef, $vmid);
3559
3560     my $volid_hash = {};
3561     foreach my $storeid (keys %$info) {
3562         foreach my $item (@{$info->{$storeid}}) {
3563             next if !($item->{volid} && $item->{size});
3564             $volid_hash->{$item->{volid}} = $item;
3565         }
3566     }
3567
3568     return $volid_hash;
3569 }
3570
3571 sub update_disksize {
3572     my ($vmid, $conf, $volid_hash) = @_;
3573
3574     my $changes;
3575
3576     my $used = {};
3577
3578     # update size info
3579     foreach my $opt (keys %$conf) {
3580         if (PVE::QemuServer::valid_drivename($opt)) {
3581             my $drive = PVE::QemuServer::parse_drive($opt, $conf->{$opt});
3582             my $volid = $drive->{file};
3583             next if !$volid;
3584
3585             $used->{$volid} = 1;
3586
3587             next if PVE::QemuServer::drive_is_cdrom($drive);
3588             next if !$volid_hash->{$volid};
3589
3590             $drive->{size} = $volid_hash->{$volid}->{size};
3591             $changes = 1;
3592             $conf->{$opt} = PVE::QemuServer::print_drive($vmid, $drive);
3593         }
3594     }
3595
3596     foreach my $volid (sort keys %$volid_hash) {
3597         next if $volid =~ m/vm-$vmid-state-/;
3598         next if $used->{$volid};
3599         $changes = 1;
3600         PVE::QemuServer::add_unused_volume($conf, $volid);
3601     }
3602
3603     return $changes;
3604 }
3605
3606 sub rescan {
3607     my ($vmid, $nolock) = @_;
3608
3609     my $cfg = PVE::Cluster::cfs_read_file("storage.cfg");
3610
3611     my $volid_hash = scan_volids($cfg, $vmid);
3612
3613     my $updatefn =  sub {
3614         my ($vmid) = @_;
3615
3616         my $conf = PVE::QemuServer::load_config($vmid);
3617
3618         PVE::QemuServer::check_lock($conf);
3619
3620         my $changes = PVE::QemuServer::update_disksize($vmid, $conf, $volid_hash);
3621
3622         PVE::QemuServer::update_config_nolock($vmid, $conf, 1) if $changes;
3623     };
3624
3625     if (defined($vmid)) {
3626         if ($nolock) {
3627             &$updatefn($vmid);
3628         } else {
3629             PVE::QemuServer::lock_config($vmid, $updatefn, $vmid);
3630         }
3631     } else {
3632         my $vmlist = config_list();
3633         foreach my $vmid (keys %$vmlist) {
3634             if ($nolock) {
3635                 &$updatefn($vmid);
3636             } else {
3637                 PVE::QemuServer::lock_config($vmid, $updatefn, $vmid);
3638             }
3639         }
3640     }
3641 }
3642
3643 sub restore_vma_archive {
3644     my ($archive, $vmid, $user, $opts, $comp) = @_;
3645
3646     my $input = $archive eq '-' ? "<&STDIN" : undef;
3647     my $readfrom = $archive;
3648
3649     my $uncomp = '';
3650     if ($comp) {
3651         $readfrom = '-';
3652         my $qarchive = PVE::Tools::shellquote($archive);
3653         if ($comp eq 'gzip') {
3654             $uncomp = "zcat $qarchive|";
3655         } elsif ($comp eq 'lzop') {
3656             $uncomp = "lzop -d -c $qarchive|";
3657         } else {
3658             die "unknown compression method '$comp'\n";
3659         }
3660
3661     }
3662
3663     my $tmpdir = "/var/tmp/vzdumptmp$$";
3664     rmtree $tmpdir;
3665
3666     # disable interrupts (always do cleanups)
3667     local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = sub {
3668         warn "got interrupt - ignored\n";
3669     };
3670
3671     my $mapfifo = "/var/tmp/vzdumptmp$$.fifo";
3672     POSIX::mkfifo($mapfifo, 0600);
3673     my $fifofh;
3674
3675     my $openfifo = sub {
3676         open($fifofh, '>', $mapfifo) || die $!;
3677     };
3678
3679     my $cmd = "${uncomp}vma extract -v -r $mapfifo $readfrom $tmpdir";
3680
3681     my $oldtimeout;
3682     my $timeout = 5;
3683
3684     my $devinfo = {};
3685
3686     my $rpcenv = PVE::RPCEnvironment::get();
3687
3688     my $conffile = PVE::QemuServer::config_file($vmid);
3689     my $tmpfn = "$conffile.$$.tmp";
3690
3691     my $print_devmap = sub {
3692         my $virtdev_hash = {};
3693
3694         my $cfgfn = "$tmpdir/qemu-server.conf";
3695
3696         # we can read the config - that is already extracted
3697         my $fh = IO::File->new($cfgfn, "r") ||
3698             "unable to read qemu-server.conf - $!\n";
3699
3700         while (defined(my $line = <$fh>)) {
3701             if ($line =~ m/^\#qmdump\#map:(\S+):(\S+):(\S*):(\S*):$/) {
3702                 my ($virtdev, $devname, $storeid, $format) = ($1, $2, $3, $4);
3703                 die "archive does not contain data for drive '$virtdev'\n"
3704                     if !$devinfo->{$devname};
3705                 if (defined($opts->{storage})) {
3706                     $storeid = $opts->{storage} || 'local';
3707                 } elsif (!$storeid) {
3708                     $storeid = 'local';
3709                 }
3710                 $format = 'raw' if !$format;
3711                 $devinfo->{$devname}->{devname} = $devname;
3712                 $devinfo->{$devname}->{virtdev} = $virtdev;
3713                 $devinfo->{$devname}->{format} = $format;
3714                 $devinfo->{$devname}->{storeid} = $storeid;
3715
3716                 # check permission on storage
3717                 my $pool = $opts->{pool}; # todo: do we need that?
3718                 if ($user ne 'root@pam') {
3719                     $rpcenv->check($user, "/storage/$storeid", ['Datastore.AllocateSpace']);
3720                 }
3721
3722                 $virtdev_hash->{$virtdev} = $devinfo->{$devname};
3723             }
3724         }
3725
3726         foreach my $devname (keys %$devinfo) {
3727             die "found no device mapping information for device '$devname'\n"
3728                 if !$devinfo->{$devname}->{virtdev};
3729         }
3730
3731         my $map = {};
3732         my $cfg = cfs_read_file('storage.cfg');
3733         foreach my $virtdev (sort keys %$virtdev_hash) {
3734             my $d = $virtdev_hash->{$virtdev};
3735             my $alloc_size = int(($d->{size} + 1024 - 1)/1024);
3736             my $scfg = PVE::Storage::storage_config($cfg, $d->{storeid});
3737             my $volid = PVE::Storage::vdisk_alloc($cfg, $d->{storeid}, $vmid,
3738                                                   $d->{format}, undef, $alloc_size);
3739             print STDERR "new volume ID is '$volid'\n";
3740             $d->{volid} = $volid;
3741             my $path = PVE::Storage::path($cfg, $volid);
3742
3743             my $write_zeros = 1;
3744             # fixme: what other storages types initialize volumes with zero?
3745             if ($scfg->{type} eq 'dir' || $scfg->{type} eq 'nfs') {
3746                 $write_zeros = 0;
3747             }
3748
3749             print $fifofh "${write_zeros}:$d->{devname}=$path\n";
3750
3751             print "map '$d->{devname}' to '$path' (write zeros = ${write_zeros})\n";
3752             $map->{$virtdev} = $volid;
3753         }
3754
3755         $fh->seek(0, 0) || die "seek failed - $!\n";
3756
3757         my $outfd = new IO::File ($tmpfn, "w") ||
3758             die "unable to write config for VM $vmid\n";
3759
3760         my $cookie = { netcount => 0 };
3761         while (defined(my $line = <$fh>)) {
3762             restore_update_config_line($outfd, $cookie, $vmid, $map, $line, $opts->{unique});
3763         }
3764
3765         $fh->close();
3766         $outfd->close();
3767     };
3768
3769     eval {
3770         # enable interrupts
3771         local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = $SIG{PIPE} = sub {
3772             die "interrupted by signal\n";
3773         };
3774         local $SIG{ALRM} = sub { die "got timeout\n"; };
3775
3776         $oldtimeout = alarm($timeout);
3777
3778         my $parser = sub {
3779             my $line = shift;
3780
3781             print "$line\n";
3782
3783             if ($line =~ m/^DEV:\sdev_id=(\d+)\ssize:\s(\d+)\sdevname:\s(\S+)$/) {
3784                 my ($dev_id, $size, $devname) = ($1, $2, $3);
3785                 $devinfo->{$devname} = { size => $size, dev_id => $dev_id };
3786             } elsif ($line =~ m/^CTIME: /) {
3787                 &$print_devmap();
3788                 print $fifofh "done\n";
3789                 my $tmp = $oldtimeout || 0;
3790                 $oldtimeout = undef;
3791                 alarm($tmp);
3792                 close($fifofh);
3793             }
3794         };
3795
3796         print "restore vma archive: $cmd\n";
3797         run_command($cmd, input => $input, outfunc => $parser, afterfork => $openfifo);
3798     };
3799     my $err = $@;
3800
3801     alarm($oldtimeout) if $oldtimeout;
3802
3803     unlink $mapfifo;
3804
3805     if ($err) {
3806         rmtree $tmpdir;
3807         unlink $tmpfn;
3808
3809         my $cfg = cfs_read_file('storage.cfg');
3810         foreach my $devname (keys %$devinfo) {
3811             my $volid = $devinfo->{$devname}->{volid};
3812             next if !$volid;
3813             eval {
3814                 if ($volid =~ m|^/|) {
3815                     unlink $volid || die 'unlink failed\n';
3816                 } else {
3817                     PVE::Storage::vdisk_free($cfg, $volid);
3818                 }
3819                 print STDERR "temporary volume '$volid' sucessfuly removed\n";
3820             };
3821             print STDERR "unable to cleanup '$volid' - $@" if $@;
3822         }
3823         die $err;
3824     }
3825
3826     rmtree $tmpdir;
3827
3828     rename $tmpfn, $conffile ||
3829         die "unable to commit configuration file '$conffile'\n";
3830
3831     eval { rescan($vmid, 1); };
3832     warn $@ if $@;
3833 }
3834
3835 sub restore_tar_archive {
3836     my ($archive, $vmid, $user, $opts) = @_;
3837
3838     if ($archive ne '-') {
3839         my $firstfile = archive_read_firstfile($archive);
3840         die "ERROR: file '$archive' dos not lock like a QemuServer vzdump backup\n"
3841             if $firstfile ne 'qemu-server.conf';
3842     }
3843
3844     my $tocmd = "/usr/lib/qemu-server/qmextract";
3845
3846     $tocmd .= " --storage " . PVE::Tools::shellquote($opts->{storage}) if $opts->{storage};
3847     $tocmd .= " --pool " . PVE::Tools::shellquote($opts->{pool}) if $opts->{pool};
3848     $tocmd .= ' --prealloc' if $opts->{prealloc};
3849     $tocmd .= ' --info' if $opts->{info};
3850
3851     # tar option "xf" does not autodetect compression when read from STDIN,
3852     # so we pipe to zcat
3853     my $cmd = "zcat -f|tar xf " . PVE::Tools::shellquote($archive) . " " .
3854         PVE::Tools::shellquote("--to-command=$tocmd");
3855
3856     my $tmpdir = "/var/tmp/vzdumptmp$$";
3857     mkpath $tmpdir;
3858
3859     local $ENV{VZDUMP_TMPDIR} = $tmpdir;
3860     local $ENV{VZDUMP_VMID} = $vmid;
3861     local $ENV{VZDUMP_USER} = $user;
3862
3863     my $conffile = PVE::QemuServer::config_file($vmid);
3864     my $tmpfn = "$conffile.$$.tmp";
3865
3866     # disable interrupts (always do cleanups)
3867     local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = sub {
3868         print STDERR "got interrupt - ignored\n";
3869     };
3870
3871     eval {
3872         # enable interrupts
3873         local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = $SIG{PIPE} = sub {
3874             die "interrupted by signal\n";
3875         };
3876
3877         if ($archive eq '-') {
3878             print "extracting archive from STDIN\n";
3879             run_command($cmd, input => "<&STDIN");
3880         } else {
3881             print "extracting archive '$archive'\n";
3882             run_command($cmd);
3883         }
3884
3885         return if $opts->{info};
3886
3887         # read new mapping
3888         my $map = {};
3889         my $statfile = "$tmpdir/qmrestore.stat";
3890         if (my $fd = IO::File->new($statfile, "r")) {
3891             while (defined (my $line = <$fd>)) {
3892                 if ($line =~ m/vzdump:([^\s:]*):(\S+)$/) {
3893                     $map->{$1} = $2 if $1;
3894                 } else {
3895                     print STDERR "unable to parse line in statfile - $line\n";
3896                 }
3897             }
3898             $fd->close();
3899         }
3900
3901         my $confsrc = "$tmpdir/qemu-server.conf";
3902
3903         my $srcfd = new IO::File($confsrc, "r") ||
3904             die "unable to open file '$confsrc'\n";
3905
3906         my $outfd = new IO::File ($tmpfn, "w") ||
3907             die "unable to write config for VM $vmid\n";
3908
3909         my $cookie = { netcount => 0 };
3910         while (defined (my $line = <$srcfd>)) {
3911             restore_update_config_line($outfd, $cookie, $vmid, $map, $line, $opts->{unique});
3912         }
3913
3914         $srcfd->close();
3915         $outfd->close();
3916     };
3917     my $err = $@;
3918
3919     if ($err) {
3920
3921         unlink $tmpfn;
3922
3923         restore_cleanup("$tmpdir/qmrestore.stat") if !$opts->{info};
3924
3925         die $err;
3926     }
3927
3928     rmtree $tmpdir;
3929
3930     rename $tmpfn, $conffile ||
3931         die "unable to commit configuration file '$conffile'\n";
3932
3933     eval { rescan($vmid, 1); };
3934     warn $@ if $@;
3935 };
3936
3937
3938 # Internal snapshots
3939
3940 # NOTE: Snapshot create/delete involves several non-atomic
3941 # action, and can take a long time.
3942 # So we try to avoid locking the file and use 'lock' variable
3943 # inside the config file instead.
3944
3945 my $snapshot_copy_config = sub {
3946     my ($source, $dest) = @_;
3947
3948     foreach my $k (keys %$source) {
3949         next if $k eq 'snapshots';
3950         next if $k eq 'snapstate';
3951         next if $k eq 'snaptime';
3952         next if $k eq 'vmstate';
3953         next if $k eq 'lock';
3954         next if $k eq 'digest';
3955         next if $k eq 'description';
3956         next if $k =~ m/^unused\d+$/;
3957
3958         $dest->{$k} = $source->{$k};
3959     }
3960 };
3961
3962 my $snapshot_apply_config = sub {
3963     my ($conf, $snap) = @_;
3964
3965     # copy snapshot list
3966     my $newconf = {
3967         snapshots => $conf->{snapshots},
3968     };
3969
3970     # keep description and list of unused disks
3971     foreach my $k (keys %$conf) {
3972         next if !($k =~ m/^unused\d+$/ || $k eq 'description');
3973         $newconf->{$k} = $conf->{$k};
3974     }
3975
3976     &$snapshot_copy_config($snap, $newconf);
3977
3978     return $newconf;
3979 };
3980
3981 sub foreach_writable_storage {
3982     my ($conf, $func) = @_;
3983
3984     my $sidhash = {};
3985
3986     foreach my $ds (keys %$conf) {
3987         next if !valid_drivename($ds);
3988
3989         my $drive = parse_drive($ds, $conf->{$ds});
3990         next if !$drive;
3991         next if drive_is_cdrom($drive);
3992
3993         my $volid = $drive->{file};
3994
3995         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
3996         $sidhash->{$sid} = $sid if $sid;
3997     }
3998
3999     foreach my $sid (sort keys %$sidhash) {
4000         &$func($sid);
4001     }
4002 }
4003
4004 my $alloc_vmstate_volid = sub {
4005     my ($storecfg, $vmid, $conf, $snapname) = @_;
4006
4007     # Note: we try to be smart when selecting a $target storage
4008
4009     my $target;
4010
4011     # search shared storage first
4012     foreach_writable_storage($conf, sub {
4013         my ($sid) = @_;
4014         my $scfg = PVE::Storage::storage_config($storecfg, $sid);
4015         return if !$scfg->{shared};
4016
4017         $target = $sid if !$target || $scfg->{path}; # prefer file based storage
4018     });
4019
4020     if (!$target) {
4021         # now search local storage
4022         foreach_writable_storage($conf, sub {
4023             my ($sid) = @_;
4024             my $scfg = PVE::Storage::storage_config($storecfg, $sid);
4025             return if $scfg->{shared};
4026
4027             $target = $sid if !$target || $scfg->{path}; # prefer file based storage;
4028         });
4029     }
4030
4031     $target = 'local' if !$target;
4032
4033     my $driver_state_size = 500; # assume 32MB is enough to safe all driver state;
4034     # we abort live save after $conf->{memory}, so we need at max twice that space
4035     my $size = $conf->{memory}*2 + $driver_state_size;
4036
4037     my $name = "vm-$vmid-state-$snapname";
4038     my $scfg = PVE::Storage::storage_config($storecfg, $target);
4039     $name .= ".raw" if $scfg->{path}; # add filename extension for file base storage
4040     my $volid = PVE::Storage::vdisk_alloc($storecfg, $target, $vmid, 'raw', $name, $size*1024);
4041
4042     return $volid;
4043 };
4044
4045 my $snapshot_prepare = sub {
4046     my ($vmid, $snapname, $save_vmstate, $comment) = @_;
4047
4048     my $snap;
4049
4050     my $updatefn =  sub {
4051
4052         my $conf = load_config($vmid);
4053
4054         check_lock($conf);
4055
4056         $conf->{lock} = 'snapshot';
4057
4058         die "snapshot name '$snapname' already used\n"
4059             if defined($conf->{snapshots}->{$snapname});
4060
4061         my $storecfg = PVE::Storage::config();
4062         die "snapshot feature is not available" if !has_feature('snapshot', $conf, $storecfg);
4063
4064         $snap = $conf->{snapshots}->{$snapname} = {};
4065
4066         if ($save_vmstate && check_running($vmid)) {
4067             $snap->{vmstate} = &$alloc_vmstate_volid($storecfg, $vmid, $conf, $snapname);
4068         }
4069
4070         &$snapshot_copy_config($conf, $snap);
4071
4072         $snap->{snapstate} = "prepare";
4073         $snap->{snaptime} = time();
4074         $snap->{description} = $comment if $comment;
4075
4076         update_config_nolock($vmid, $conf, 1);
4077     };
4078
4079     lock_config($vmid, $updatefn);
4080
4081     return $snap;
4082 };
4083
4084 my $snapshot_commit = sub {
4085     my ($vmid, $snapname) = @_;
4086
4087     my $updatefn = sub {
4088
4089         my $conf = load_config($vmid);
4090
4091         die "missing snapshot lock\n"
4092             if !($conf->{lock} && $conf->{lock} eq 'snapshot');
4093
4094         my $snap = $conf->{snapshots}->{$snapname};
4095
4096         die "snapshot '$snapname' does not exist\n" if !defined($snap);
4097
4098         die "wrong snapshot state\n"
4099             if !($snap->{snapstate} && $snap->{snapstate} eq "prepare");
4100
4101         delete $snap->{snapstate};
4102         delete $conf->{lock};
4103
4104         my $newconf = &$snapshot_apply_config($conf, $snap);
4105
4106         $newconf->{parent} = $snapname;
4107
4108         update_config_nolock($vmid, $newconf, 1);
4109     };
4110
4111     lock_config($vmid, $updatefn);
4112 };
4113
4114 sub snapshot_rollback {
4115     my ($vmid, $snapname) = @_;
4116
4117     my $snap;
4118
4119     my $prepare = 1;
4120
4121     my $storecfg = PVE::Storage::config();
4122
4123     my $updatefn = sub {
4124
4125         my $conf = load_config($vmid);
4126
4127         $snap = $conf->{snapshots}->{$snapname};
4128
4129         die "snapshot '$snapname' does not exist\n" if !defined($snap);
4130
4131         die "unable to rollback to incomplete snapshot (snapstate = $snap->{snapstate})\n"
4132             if $snap->{snapstate};
4133
4134         if ($prepare) {
4135             check_lock($conf);
4136             vm_stop($storecfg, $vmid, undef, undef, 5, undef, undef);
4137         }
4138
4139         die "unable to rollback vm $vmid: vm is running\n"
4140             if check_running($vmid);
4141
4142         if ($prepare) {
4143             $conf->{lock} = 'rollback';
4144         } else {
4145             die "got wrong lock\n" if !($conf->{lock} && $conf->{lock} eq 'rollback');
4146             delete $conf->{lock};
4147         }
4148
4149         if (!$prepare) {
4150             # copy snapshot config to current config
4151             $conf = &$snapshot_apply_config($conf, $snap);
4152             $conf->{parent} = $snapname;
4153         }
4154
4155         update_config_nolock($vmid, $conf, 1);
4156
4157         if (!$prepare && $snap->{vmstate}) {
4158             my $statefile = PVE::Storage::path($storecfg, $snap->{vmstate});
4159             vm_start($storecfg, $vmid, $statefile);
4160         }
4161     };
4162
4163     lock_config($vmid, $updatefn);
4164
4165     foreach_drive($snap, sub {
4166         my ($ds, $drive) = @_;
4167
4168         return if drive_is_cdrom($drive);
4169
4170         my $volid = $drive->{file};
4171         my $device = "drive-$ds";
4172
4173         PVE::Storage::volume_snapshot_rollback($storecfg, $volid, $snapname);
4174     });
4175
4176     $prepare = 0;
4177     lock_config($vmid, $updatefn);
4178 }
4179
4180 my $savevm_wait = sub {
4181     my ($vmid) = @_;
4182
4183     for(;;) {
4184         my $stat = PVE::QemuServer::vm_mon_cmd_nocheck($vmid, "query-savevm");
4185         if (!$stat->{status}) {
4186             die "savevm not active\n";
4187         } elsif ($stat->{status} eq 'active') {
4188             sleep(1);
4189             next;
4190         } elsif ($stat->{status} eq 'completed') {
4191             last;
4192         } else {
4193             die "query-savevm returned status '$stat->{status}'\n";
4194         }
4195     }
4196 };
4197
4198 sub snapshot_create {
4199     my ($vmid, $snapname, $save_vmstate, $freezefs, $comment) = @_;
4200
4201     my $snap = &$snapshot_prepare($vmid, $snapname, $save_vmstate, $comment);
4202
4203     $freezefs = $save_vmstate = 0 if !$snap->{vmstate}; # vm is not running
4204
4205     my $drivehash = {};
4206
4207     my $running = check_running($vmid);
4208
4209     eval {
4210         # create internal snapshots of all drives
4211
4212         my $storecfg = PVE::Storage::config();
4213
4214         if ($running) {
4215             if ($snap->{vmstate}) {
4216                 my $path = PVE::Storage::path($storecfg, $snap->{vmstate});
4217                 vm_mon_cmd($vmid, "savevm-start", statefile => $path);
4218                 &$savevm_wait($vmid);
4219             } else {
4220                 vm_mon_cmd($vmid, "savevm-start");
4221             }
4222         };
4223
4224         qga_freezefs($vmid) if $running && $freezefs;
4225
4226         foreach_drive($snap, sub {
4227             my ($ds, $drive) = @_;
4228
4229             return if drive_is_cdrom($drive);
4230
4231             my $volid = $drive->{file};
4232             my $device = "drive-$ds";
4233
4234             qemu_volume_snapshot($vmid, $device, $storecfg, $volid, $snapname);
4235             $drivehash->{$ds} = 1;
4236        });
4237     };
4238     my $err = $@;
4239
4240     eval { gqa_unfreezefs($vmid) if $running && $freezefs; };
4241     warn $@ if $@;
4242
4243     eval { vm_mon_cmd($vmid, "savevm-end") if $running; };
4244     warn $@ if $@;
4245
4246     if ($err) {
4247         warn "snapshot create failed: starting cleanup\n";
4248         eval { snapshot_delete($vmid, $snapname, 0, $drivehash); };
4249         warn $@ if $@;
4250         die $err;
4251     }
4252
4253     &$snapshot_commit($vmid, $snapname);
4254 }
4255
4256 # Note: $drivehash is only set when called from snapshot_create.
4257 sub snapshot_delete {
4258     my ($vmid, $snapname, $force, $drivehash) = @_;
4259
4260     my $prepare = 1;
4261
4262     my $snap;
4263     my $unused = [];
4264
4265     my $unlink_parent = sub {
4266         my ($confref, $new_parent) = @_;
4267
4268         if ($confref->{parent} && $confref->{parent} eq $snapname) {
4269             if ($new_parent) {
4270                 $confref->{parent} = $new_parent;
4271             } else {
4272                 delete $confref->{parent};
4273             }
4274         }
4275     };
4276
4277     my $updatefn =  sub {
4278         my ($remove_drive) = @_;
4279
4280         my $conf = load_config($vmid);
4281
4282         check_lock($conf) if !$drivehash;
4283
4284         $snap = $conf->{snapshots}->{$snapname};
4285
4286         die "snapshot '$snapname' does not exist\n" if !defined($snap);
4287
4288         # remove parent refs
4289         &$unlink_parent($conf, $snap->{parent});
4290         foreach my $sn (keys %{$conf->{snapshots}}) {
4291             next if $sn eq $snapname;
4292             &$unlink_parent($conf->{snapshots}->{$sn}, $snap->{parent});
4293         }
4294
4295         if ($remove_drive) {
4296             if ($remove_drive eq 'vmstate') {
4297                 delete $snap->{$remove_drive};
4298             } else {
4299                 my $drive = parse_drive($remove_drive, $snap->{$remove_drive});
4300                 my $volid = $drive->{file};
4301                 delete $snap->{$remove_drive};
4302                 add_unused_volume($conf, $volid);
4303             }
4304         }
4305
4306         if ($prepare) {
4307             $snap->{snapstate} = 'delete';
4308         } else {
4309             delete $conf->{snapshots}->{$snapname};
4310             delete $conf->{lock} if $drivehash;
4311             foreach my $volid (@$unused) {
4312                 add_unused_volume($conf, $volid);
4313             }
4314         }
4315
4316         update_config_nolock($vmid, $conf, 1);
4317     };
4318
4319     lock_config($vmid, $updatefn);
4320
4321     # now remove vmstate file
4322
4323     my $storecfg = PVE::Storage::config();
4324
4325     if ($snap->{vmstate}) {
4326         eval {  PVE::Storage::vdisk_free($storecfg, $snap->{vmstate}); };
4327         if (my $err = $@) {
4328             die $err if !$force;
4329             warn $err;
4330         }
4331         # save changes (remove vmstate from snapshot)
4332         lock_config($vmid, $updatefn, 'vmstate') if !$force;
4333     };
4334
4335     # now remove all internal snapshots
4336     foreach_drive($snap, sub {
4337         my ($ds, $drive) = @_;
4338
4339         return if drive_is_cdrom($drive);
4340
4341         my $volid = $drive->{file};
4342         my $device = "drive-$ds";
4343
4344         if (!$drivehash || $drivehash->{$ds}) {
4345             eval { qemu_volume_snapshot_delete($vmid, $device, $storecfg, $volid, $snapname); };
4346             if (my $err = $@) {
4347                 die $err if !$force;
4348                 warn $err;
4349             }
4350         }
4351
4352         # save changes (remove drive fron snapshot)
4353         lock_config($vmid, $updatefn, $ds) if !$force;
4354         push @$unused, $volid;
4355     });
4356
4357     # now cleanup config
4358     $prepare = 0;
4359     lock_config($vmid, $updatefn);
4360 }
4361
4362 sub has_feature{
4363     my ($feature, $conf, $storecfg, $snapname, $running) = @_;
4364
4365     my $err = undef;
4366     foreach_drive($conf, sub {
4367         my ($ds, $drive) = @_;
4368
4369         return if drive_is_cdrom($drive);
4370         my $volid = $drive->{file};
4371         $err = 1 if !PVE::Storage::volume_has_feature($storecfg, $feature, $volid, $snapname, $running);
4372     });
4373
4374     return 1 if !$err;
4375 }
4376 1;