PVE/QemuServer.pm

   1 package PVE::QemuServer;
   2
   3 use strict;
   4 use warnings;
   5 use POSIX;
   6 use IO::Handle;
   7 use IO::Select;
   8 use IO::File;
   9 use IO::Dir;
  10 use IO::Socket::UNIX;
  11 use File::Basename;
  12 use File::Path;
  13 use File::stat;
  14 use Getopt::Long;
  15 use Digest::SHA;
  16 use Fcntl ':flock';
  17 use Cwd 'abs_path';
  18 use IPC::Open3;
  19 use JSON;
  20 use Fcntl;
  21 use PVE::SafeSyslog;
  22 use Storable qw(dclone);
  23 use PVE::Exception qw(raise raise_param_exc);
  24 use PVE::Storage;
  25 use PVE::Tools qw(run_command lock_file lock_file_full file_read_firstline dir_glob_foreach);
  26 use PVE::JSONSchema qw(get_standard_option);
  27 use PVE::Cluster qw(cfs_register_file cfs_read_file cfs_write_file cfs_lock_file);
  28 use PVE::INotify;
  29 use PVE::ProcFSTools;
  30 use PVE::QMPClient;
  31 use PVE::RPCEnvironment;
  32 use Time::HiRes qw(gettimeofday);
  33
  34 my $cpuinfo = PVE::ProcFSTools::read_cpuinfo();
  35
  36 # Note about locking: we use flock on the config file protect
  37 # against concurent actions.
  38 # Aditionaly, we have a 'lock' setting in the config file. This
  39 # can be set to 'migrate', 'backup', 'snapshot' or 'rollback'. Most actions are not
  40 # allowed when such lock is set. But you can ignore this kind of
  41 # lock with the --skiplock flag.
  42
  43 cfs_register_file('/qemu-server/',
  44                   \&parse_vm_config,
  45                   \&write_vm_config);
  46
  47 PVE::JSONSchema::register_standard_option('skiplock', {
  48     description => "Ignore locks - only root is allowed to use this option.",
  49     type => 'boolean',
  50     optional => 1,
  51 });
  52
  53 PVE::JSONSchema::register_standard_option('pve-qm-stateuri', {
  54     description => "Some command save/restore state from this location.",
  55     type => 'string',
  56     maxLength => 128,
  57     optional => 1,
  58 });
  59
  60 PVE::JSONSchema::register_standard_option('pve-snapshot-name', {
  61     description => "The name of the snapshot.",
  62     type => 'string', format => 'pve-configid',
  63     maxLength => 40,
  64 });
  65
  66 #no warnings 'redefine';
  67
  68 unless(defined(&_VZSYSCALLS_H_)) {
  69     eval 'sub _VZSYSCALLS_H_ () {1;}' unless defined(&_VZSYSCALLS_H_);
  70     require 'sys/syscall.ph';
  71     if(defined(&__x86_64__)) {
  72         eval 'sub __NR_fairsched_vcpus () {499;}' unless defined(&__NR_fairsched_vcpus);
  73         eval 'sub __NR_fairsched_mknod () {504;}' unless defined(&__NR_fairsched_mknod);
  74         eval 'sub __NR_fairsched_rmnod () {505;}' unless defined(&__NR_fairsched_rmnod);
  75         eval 'sub __NR_fairsched_chwt () {506;}' unless defined(&__NR_fairsched_chwt);
  76         eval 'sub __NR_fairsched_mvpr () {507;}' unless defined(&__NR_fairsched_mvpr);
  77         eval 'sub __NR_fairsched_rate () {508;}' unless defined(&__NR_fairsched_rate);
  78         eval 'sub __NR_setluid () {501;}' unless defined(&__NR_setluid);
  79         eval 'sub __NR_setublimit () {502;}' unless defined(&__NR_setublimit);
  80     }
  81     elsif(defined( &__i386__) ) {
  82         eval 'sub __NR_fairsched_mknod () {500;}' unless defined(&__NR_fairsched_mknod);
  83         eval 'sub __NR_fairsched_rmnod () {501;}' unless defined(&__NR_fairsched_rmnod);
  84         eval 'sub __NR_fairsched_chwt () {502;}' unless defined(&__NR_fairsched_chwt);
  85         eval 'sub __NR_fairsched_mvpr () {503;}' unless defined(&__NR_fairsched_mvpr);
  86         eval 'sub __NR_fairsched_rate () {504;}' unless defined(&__NR_fairsched_rate);
  87         eval 'sub __NR_fairsched_vcpus () {505;}' unless defined(&__NR_fairsched_vcpus);
  88         eval 'sub __NR_setluid () {511;}' unless defined(&__NR_setluid);
  89         eval 'sub __NR_setublimit () {512;}' unless defined(&__NR_setublimit);
  90     } else {
  91         die("no fairsched syscall for this arch");
  92     }
  93     require 'asm/ioctl.ph';
  94     eval 'sub KVM_GET_API_VERSION () { &_IO(0xAE, 0x);}' unless defined(&KVM_GET_API_VERSION);
  95 }
  96
  97 sub fairsched_mknod {
  98     my ($parent, $weight, $desired) = @_;
  99
 100     return syscall(&__NR_fairsched_mknod, int($parent), int($weight), int($desired));
 101 }
 102
 103 sub fairsched_rmnod {
 104     my ($id) = @_;
 105
 106     return syscall(&__NR_fairsched_rmnod, int($id));
 107 }
 108
 109 sub fairsched_mvpr {
 110     my ($pid, $newid) = @_;
 111
 112     return syscall(&__NR_fairsched_mvpr, int($pid), int($newid));
 113 }
 114
 115 sub fairsched_vcpus {
 116     my ($id, $vcpus) = @_;
 117
 118     return syscall(&__NR_fairsched_vcpus, int($id), int($vcpus));
 119 }
 120
 121 sub fairsched_rate {
 122     my ($id, $op, $rate) = @_;
 123
 124     return syscall(&__NR_fairsched_rate, int($id), int($op), int($rate));
 125 }
 126
 127 use constant FAIRSCHED_SET_RATE  => 0;
 128 use constant FAIRSCHED_DROP_RATE => 1;
 129 use constant FAIRSCHED_GET_RATE  => 2;
 130
 131 sub fairsched_cpulimit {
 132     my ($id, $limit) = @_;
 133
 134     my $cpulim1024 = int($limit * 1024 / 100);
 135     my $op = $cpulim1024 ? FAIRSCHED_SET_RATE : FAIRSCHED_DROP_RATE;
 136
 137     return fairsched_rate($id, $op, $cpulim1024);
 138 }
 139
 140 my $nodename = PVE::INotify::nodename();
 141
 142 mkdir "/etc/pve/nodes/$nodename";
 143 my $confdir = "/etc/pve/nodes/$nodename/qemu-server";
 144 mkdir $confdir;
 145
 146 my $var_run_tmpdir = "/var/run/qemu-server";
 147 mkdir $var_run_tmpdir;
 148
 149 my $lock_dir = "/var/lock/qemu-server";
 150 mkdir $lock_dir;
 151
 152 my $pcisysfs = "/sys/bus/pci";
 153
 154 my $confdesc = {
 155     iothread => {
 156         optional => 1,
 157         type => 'boolean',
 158         description => "Enable iothread dataplane.",
 159         default => 0,
 160     },
 161     onboot => {
 162         optional => 1,
 163         type => 'boolean',
 164         description => "Specifies whether a VM will be started during system bootup.",
 165         default => 0,
 166     },
 167     autostart => {
 168         optional => 1,
 169         type => 'boolean',
 170         description => "Automatic restart after crash (currently ignored).",
 171         default => 0,
 172     },
 173     hotplug => {
 174         optional => 1,
 175         type => 'string', format => 'pve-hotplug-features',
 176         description => "Selectively enable hotplug features. This is a comma separated list of hotplug features: 'network', 'disk', 'cpu', 'memory' and 'usb'. Use '0' to disable hotplug completely. Value '1' is an alias for the default 'network,disk,usb'.",
 177         default => 'network,disk,usb',
 178     },
 179     reboot => {
 180         optional => 1,
 181         type => 'boolean',
 182         description => "Allow reboot. If set to '0' the VM exit on reboot.",
 183         default => 1,
 184     },
 185     lock => {
 186         optional => 1,
 187         type => 'string',
 188         description => "Lock/unlock the VM.",
 189         enum => [qw(migrate backup snapshot rollback)],
 190     },
 191     cpulimit => {
 192         optional => 1,
 193         type => 'integer',
 194         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.",
 195         minimum => 0,
 196         default => 0,
 197     },
 198     cpuunits => {
 199         optional => 1,
 200         type => 'integer',
 201         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.",
 202         minimum => 0,
 203         maximum => 500000,
 204         default => 1000,
 205     },
 206     memory => {
 207         optional => 1,
 208         type => 'integer',
 209         description => "Amount of RAM for the VM in MB. This is the maximum available memory when you use the balloon device.",
 210         minimum => 16,
 211         default => 512,
 212     },
 213     balloon => {
 214         optional => 1,
 215         type => 'integer',
 216         description => "Amount of target RAM for the VM in MB. Using zero disables the ballon driver.",
 217         minimum => 0,
 218     },
 219     shares => {
 220         optional => 1,
 221         type => 'integer',
 222         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",
 223         minimum => 0,
 224         maximum => 50000,
 225         default => 1000,
 226     },
 227     keyboard => {
 228         optional => 1,
 229         type => 'string',
 230         description => "Keybord layout for vnc server. Default is read from the datacenter configuration file.",
 231         enum => PVE::Tools::kvmkeymaplist(),
 232         default => 'en-us',
 233     },
 234     name => {
 235         optional => 1,
 236         type => 'string', format => 'dns-name',
 237         description => "Set a name for the VM. Only used on the configuration web interface.",
 238     },
 239     scsihw => {
 240         optional => 1,
 241         type => 'string',
 242         description => "scsi controller model",
 243         enum => [qw(lsi lsi53c810 virtio-scsi-pci megasas pvscsi)],
 244         default => 'lsi',
 245     },
 246     description => {
 247         optional => 1,
 248         type => 'string',
 249         description => "Description for the VM. Only used on the configuration web interface. This is saved as comment inside the configuration file.",
 250     },
 251     ostype => {
 252         optional => 1,
 253         type => 'string',
 254         enum => [qw(other wxp w2k w2k3 w2k8 wvista win7 win8 l24 l26 solaris)],
 255         description => <<EODESC,
 256 Used to enable special optimization/features for specific
 257 operating systems:
 258
 259 other  => unspecified OS
 260 wxp    => Microsoft Windows XP
 261 w2k    => Microsoft Windows 2000
 262 w2k3   => Microsoft Windows 2003
 263 w2k8   => Microsoft Windows 2008
 264 wvista => Microsoft Windows Vista
 265 win7   => Microsoft Windows 7
 266 win8   => Microsoft Windows 8/2012
 267 l24    => Linux 2.4 Kernel
 268 l26    => Linux 2.6/3.X Kernel
 269 solaris => solaris/opensolaris/openindiania kernel
 270
 271 other|l24|l26|solaris                       ... no special behaviour
 272 wxp|w2k|w2k3|w2k8|wvista|win7|win8  ... use --localtime switch
 273 EODESC
 274     },
 275     boot => {
 276         optional => 1,
 277         type => 'string',
 278         description => "Boot on floppy (a), hard disk (c), CD-ROM (d), or network (n).",
 279         pattern => '[acdn]{1,4}',
 280         default => 'cdn',
 281     },
 282     bootdisk => {
 283         optional => 1,
 284         type => 'string', format => 'pve-qm-bootdisk',
 285         description => "Enable booting from specified disk.",
 286         pattern => '(ide|sata|scsi|virtio)\d+',
 287     },
 288     smp => {
 289         optional => 1,
 290         type => 'integer',
 291         description => "The number of CPUs. Please use option -sockets instead.",
 292         minimum => 1,
 293         default => 1,
 294     },
 295     sockets => {
 296         optional => 1,
 297         type => 'integer',
 298         description => "The number of CPU sockets.",
 299         minimum => 1,
 300         default => 1,
 301     },
 302     cores => {
 303         optional => 1,
 304         type => 'integer',
 305         description => "The number of cores per socket.",
 306         minimum => 1,
 307         default => 1,
 308     },
 309     numa => {
 310         optional => 1,
 311         type => 'boolean',
 312         description => "Enable/disable Numa.",
 313         default => 0,
 314     },
 315     vcpus => {
 316         optional => 1,
 317         type => 'integer',
 318         description => "Number of hotplugged vcpus.",
 319         minimum => 1,
 320         default => 0,
 321     },
 322     acpi => {
 323         optional => 1,
 324         type => 'boolean',
 325         description => "Enable/disable ACPI.",
 326         default => 1,
 327     },
 328     agent => {
 329         optional => 1,
 330         type => 'boolean',
 331         description => "Enable/disable Qemu GuestAgent.",
 332         default => 0,
 333     },
 334     kvm => {
 335         optional => 1,
 336         type => 'boolean',
 337         description => "Enable/disable KVM hardware virtualization.",
 338         default => 1,
 339     },
 340     tdf => {
 341         optional => 1,
 342         type => 'boolean',
 343         description => "Enable/disable time drift fix.",
 344         default => 0,
 345     },
 346     localtime => {
 347         optional => 1,
 348         type => 'boolean',
 349         description => "Set the real time clock to local time. This is enabled by default if ostype indicates a Microsoft OS.",
 350     },
 351     freeze => {
 352         optional => 1,
 353         type => 'boolean',
 354         description => "Freeze CPU at startup (use 'c' monitor command to start execution).",
 355     },
 356     vga => {
 357         optional => 1,
 358         type => 'string',
 359         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. Option 'qxl' enables the SPICE display sever. You can also run without any graphic card using a serial devive as terminal.",
 360         enum => [qw(std cirrus vmware qxl serial0 serial1 serial2 serial3 qxl2 qxl3 qxl4)],
 361     },
 362     watchdog => {
 363         optional => 1,
 364         type => 'string', format => 'pve-qm-watchdog',
 365         typetext => '[[model=]i6300esb|ib700] [,[action=]reset|shutdown|poweroff|pause|debug|none]',
 366         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)",
 367     },
 368     startdate => {
 369         optional => 1,
 370         type => 'string',
 371         typetext => "(now | YYYY-MM-DD | YYYY-MM-DDTHH:MM:SS)",
 372         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'.",
 373         pattern => '(now|\d{4}-\d{1,2}-\d{1,2}(T\d{1,2}:\d{1,2}:\d{1,2})?)',
 374         default => 'now',
 375     },
 376     startup => {
 377         optional => 1,
 378         type => 'string', format => 'pve-qm-startup',
 379         typetext => '[[order=]\d+] [,up=\d+] [,down=\d+] ',
 380         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.",
 381     },
 382     template => {
 383         optional => 1,
 384         type => 'boolean',
 385         description => "Enable/disable Template.",
 386         default => 0,
 387     },
 388     args => {
 389         optional => 1,
 390         type => 'string',
 391         description => <<EODESCR,
 392 Note: this option is for experts only. It allows you to pass arbitrary arguments to kvm, for example:
 393
 394 args: -no-reboot -no-hpet
 395 EODESCR
 396     },
 397     tablet => {
 398         optional => 1,
 399         type => 'boolean',
 400         default => 1,
 401         description => "Enable/disable the usb tablet device. This device is usually needed to allow absolute mouse positioning with VNC. 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. This is turned of by default if you use spice (vga=qxl).",
 402     },
 403     migrate_speed => {
 404         optional => 1,
 405         type => 'integer',
 406         description => "Set maximum speed (in MB/s) for migrations. Value 0 is no limit.",
 407         minimum => 0,
 408         default => 0,
 409     },
 410     migrate_downtime => {
 411         optional => 1,
 412         type => 'number',
 413         description => "Set maximum tolerated downtime (in seconds) for migrations.",
 414         minimum => 0,
 415         default => 0.1,
 416     },
 417     cdrom => {
 418         optional => 1,
 419         type => 'string', format => 'pve-qm-drive',
 420         typetext => 'volume',
 421         description => "This is an alias for option -ide2",
 422     },
 423     cpu => {
 424         optional => 1,
 425         description => "Emulated CPU type.",
 426         type => 'string',
 427         enum => [ qw(486 athlon pentium pentium2 pentium3 coreduo core2duo kvm32 kvm64 qemu32 qemu64 phenom Conroe Penryn Nehalem Westmere SandyBridge Haswell Broadwell Opteron_G1 Opteron_G2 Opteron_G3 Opteron_G4 Opteron_G5 host) ],
 428         default => 'kvm64',
 429     },
 430     parent => get_standard_option('pve-snapshot-name', {
 431         optional => 1,
 432         description => "Parent snapshot name. This is used internally, and should not be modified.",
 433     }),
 434     snaptime => {
 435         optional => 1,
 436         description => "Timestamp for snapshots.",
 437         type => 'integer',
 438         minimum => 0,
 439     },
 440     vmstate => {
 441         optional => 1,
 442         type => 'string', format => 'pve-volume-id',
 443         description => "Reference to a volume which stores the VM state. This is used internally for snapshots.",
 444     },
 445     machine => {
 446         description => "Specific the Qemu machine type.",
 447         type => 'string',
 448         pattern => '(pc|pc(-i440fx)?-\d+\.\d+|q35|pc-q35-\d+\.\d+)',
 449         maxLength => 40,
 450         optional => 1,
 451     },
 452     smbios1 => {
 453         description => "Specify SMBIOS type 1 fields.",
 454         type => 'string', format => 'pve-qm-smbios1',
 455         typetext => "[manufacturer=str][,product=str][,version=str][,serial=str] [,uuid=uuid][,sku=str][,family=str]",
 456         maxLength => 256,
 457         optional => 1,
 458     },
 459 };
 460
 461 # what about other qemu settings ?
 462 #cpu => 'string',
 463 #machine => 'string',
 464 #fda => 'file',
 465 #fdb => 'file',
 466 #mtdblock => 'file',
 467 #sd => 'file',
 468 #pflash => 'file',
 469 #snapshot => 'bool',
 470 #bootp => 'file',
 471 ##tftp => 'dir',
 472 ##smb => 'dir',
 473 #kernel => 'file',
 474 #append => 'string',
 475 #initrd => 'file',
 476 ##soundhw => 'string',
 477
 478 while (my ($k, $v) = each %$confdesc) {
 479     PVE::JSONSchema::register_standard_option("pve-qm-$k", $v);
 480 }
 481
 482 my $MAX_IDE_DISKS = 4;
 483 my $MAX_SCSI_DISKS = 14;
 484 my $MAX_VIRTIO_DISKS = 16;
 485 my $MAX_SATA_DISKS = 6;
 486 my $MAX_USB_DEVICES = 5;
 487 my $MAX_NETS = 32;
 488 my $MAX_UNUSED_DISKS = 8;
 489 my $MAX_HOSTPCI_DEVICES = 4;
 490 my $MAX_SERIAL_PORTS = 4;
 491 my $MAX_PARALLEL_PORTS = 3;
 492 my $MAX_NUMA = 8;
 493 my $MAX_MEM = 4194304;
 494 my $STATICMEM = 1024;
 495
 496 my $numadesc = {
 497     optional => 1,
 498     type => 'string', format => 'pve-qm-numanode',
 499     typetext => "cpus=<id[-id],memory=<mb>[[,hostnodes=<id[-id]>] [,policy=<preferred|bind|interleave>]]",
 500     description => "numa topology",
 501 };
 502 PVE::JSONSchema::register_standard_option("pve-qm-numanode", $numadesc);
 503
 504 for (my $i = 0; $i < $MAX_NUMA; $i++)  {
 505     $confdesc->{"numa$i"} = $numadesc;
 506 }
 507
 508 my $nic_model_list = ['rtl8139', 'ne2k_pci', 'e1000',  'pcnet',  'virtio',
 509                       'ne2k_isa', 'i82551', 'i82557b', 'i82559er', 'vmxnet3',
 510                       'e1000-82540em', 'e1000-82544gc', 'e1000-82545em'];
 511 my $nic_model_list_txt = join(' ', sort @$nic_model_list);
 512
 513 my $netdesc = {
 514     optional => 1,
 515     type => 'string', format => 'pve-qm-net',
 516     typetext => "MODEL=XX:XX:XX:XX:XX:XX [,bridge=<dev>][,queues=<nbqueues>][,rate=<mbps>] [,tag=<vlanid>][,firewall=0|1],link_down=0|1]",
 517     description => <<EODESCR,
 518 Specify network devices.
 519
 520 MODEL is one of: $nic_model_list_txt
 521
 522 XX:XX:XX:XX:XX:XX should be an unique MAC address. This is
 523 automatically generated if not specified.
 524
 525 The bridge parameter can be used to automatically add the interface to a bridge device. The Proxmox VE standard bridge is called 'vmbr0'.
 526
 527 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'.
 528
 529 If you specify no bridge, we create a kvm 'user' (NATed) network device, which provides DHCP and DNS services. The following addresses are used:
 530
 531 10.0.2.2   Gateway
 532 10.0.2.3   DNS Server
 533 10.0.2.4   SMB Server
 534
 535 The DHCP server assign addresses to the guest starting from 10.0.2.15.
 536
 537 EODESCR
 538 };
 539 PVE::JSONSchema::register_standard_option("pve-qm-net", $netdesc);
 540
 541 for (my $i = 0; $i < $MAX_NETS; $i++)  {
 542     $confdesc->{"net$i"} = $netdesc;
 543 }
 544
 545 my $drivename_hash;
 546
 547 my $idedesc = {
 548     optional => 1,
 549     type => 'string', format => 'pve-qm-drive',
 550     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads] [,discard=ignore|on]',
 551     description => "Use volume as IDE hard disk or CD-ROM (n is 0 to " .($MAX_IDE_DISKS -1) . ").",
 552 };
 553 PVE::JSONSchema::register_standard_option("pve-qm-ide", $idedesc);
 554
 555 my $scsidesc = {
 556     optional => 1,
 557     type => 'string', format => 'pve-qm-drive',
 558     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads] [,discard=ignore|on]',
 559     description => "Use volume as SCSI hard disk or CD-ROM (n is 0 to " . ($MAX_SCSI_DISKS - 1) . ").",
 560 };
 561 PVE::JSONSchema::register_standard_option("pve-qm-scsi", $scsidesc);
 562
 563 my $satadesc = {
 564     optional => 1,
 565     type => 'string', format => 'pve-qm-drive',
 566     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads]  [,discard=ignore|on]',
 567     description => "Use volume as SATA hard disk or CD-ROM (n is 0 to " . ($MAX_SATA_DISKS - 1). ").",
 568 };
 569 PVE::JSONSchema::register_standard_option("pve-qm-sata", $satadesc);
 570
 571 my $virtiodesc = {
 572     optional => 1,
 573     type => 'string', format => 'pve-qm-drive',
 574     typetext => '[volume=]volume,] [,media=cdrom|disk] [,cyls=c,heads=h,secs=s[,trans=t]] [,snapshot=on|off] [,cache=none|writethrough|writeback|unsafe|directsync] [,format=f] [,backup=yes|no] [,rerror=ignore|report|stop] [,werror=enospc|ignore|report|stop] [,aio=native|threads]  [,discard=ignore|on]',
 575     description => "Use volume as VIRTIO hard disk (n is 0 to " . ($MAX_VIRTIO_DISKS - 1) . ").",
 576 };
 577 PVE::JSONSchema::register_standard_option("pve-qm-virtio", $virtiodesc);
 578
 579 my $usbdesc = {
 580     optional => 1,
 581     type => 'string', format => 'pve-qm-usb-device',
 582     typetext => 'host=HOSTUSBDEVICE|spice',
 583     description => <<EODESCR,
 584 Configure an USB device (n is 0 to 4). This can be used to
 585 pass-through usb devices to the guest. HOSTUSBDEVICE syntax is:
 586
 587 'bus-port(.port)*' (decimal numbers) or
 588 'vendor_id:product_id' (hexadeciaml numbers)
 589
 590 You can use the 'lsusb -t' command to list existing usb devices.
 591
 592 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 593
 594 The value 'spice' can be used to add a usb redirection devices for spice.
 595
 596 EODESCR
 597 };
 598 PVE::JSONSchema::register_standard_option("pve-qm-usb", $usbdesc);
 599
 600 my $hostpcidesc = {
 601         optional => 1,
 602         type => 'string', format => 'pve-qm-hostpci',
 603         typetext => "[host=]HOSTPCIDEVICE [,driver=kvm|vfio] [,rombar=on|off] [,pcie=0|1] [,x-vga=on|off]",
 604         description => <<EODESCR,
 605 Map host pci devices. HOSTPCIDEVICE syntax is:
 606
 607 'bus:dev.func' (hexadecimal numbers)
 608
 609 You can us the 'lspci' command to list existing pci devices.
 610
 611 The 'rombar' option determines whether or not the device's ROM will be visible in the guest's memory map (default is 'on').
 612
 613 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 614
 615 Experimental: user reported problems with this option.
 616 EODESCR
 617 };
 618 PVE::JSONSchema::register_standard_option("pve-qm-hostpci", $hostpcidesc);
 619
 620 my $serialdesc = {
 621         optional => 1,
 622         type => 'string',
 623         pattern => '(/dev/.+|socket)',
 624         description =>  <<EODESCR,
 625 Create a serial device inside the VM (n is 0 to 3), and pass through a host serial device (i.e. /dev/ttyS0), or create a unix socket on the host side (use 'qm terminal' to open a terminal connection).
 626
 627 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 628
 629 Experimental: user reported problems with this option.
 630 EODESCR
 631 };
 632
 633 my $paralleldesc= {
 634         optional => 1,
 635         type => 'string',
 636         pattern => '/dev/parport\d+|/dev/usb/lp\d+',
 637         description =>  <<EODESCR,
 638 Map host parallel devices (n is 0 to 2).
 639
 640 Note: This option allows direct access to host hardware. So it is no longer possible to migrate such machines - use with special care.
 641
 642 Experimental: user reported problems with this option.
 643 EODESCR
 644 };
 645
 646 for (my $i = 0; $i < $MAX_PARALLEL_PORTS; $i++)  {
 647     $confdesc->{"parallel$i"} = $paralleldesc;
 648 }
 649
 650 for (my $i = 0; $i < $MAX_SERIAL_PORTS; $i++)  {
 651     $confdesc->{"serial$i"} = $serialdesc;
 652 }
 653
 654 for (my $i = 0; $i < $MAX_HOSTPCI_DEVICES; $i++)  {
 655     $confdesc->{"hostpci$i"} = $hostpcidesc;
 656 }
 657
 658 for (my $i = 0; $i < $MAX_IDE_DISKS; $i++)  {
 659     $drivename_hash->{"ide$i"} = 1;
 660     $confdesc->{"ide$i"} = $idedesc;
 661 }
 662
 663 for (my $i = 0; $i < $MAX_SATA_DISKS; $i++)  {
 664     $drivename_hash->{"sata$i"} = 1;
 665     $confdesc->{"sata$i"} = $satadesc;
 666 }
 667
 668 for (my $i = 0; $i < $MAX_SCSI_DISKS; $i++)  {
 669     $drivename_hash->{"scsi$i"} = 1;
 670     $confdesc->{"scsi$i"} = $scsidesc ;
 671 }
 672
 673 for (my $i = 0; $i < $MAX_VIRTIO_DISKS; $i++)  {
 674     $drivename_hash->{"virtio$i"} = 1;
 675     $confdesc->{"virtio$i"} = $virtiodesc;
 676 }
 677
 678 for (my $i = 0; $i < $MAX_USB_DEVICES; $i++)  {
 679     $confdesc->{"usb$i"} = $usbdesc;
 680 }
 681
 682 my $unuseddesc = {
 683     optional => 1,
 684     type => 'string', format => 'pve-volume-id',
 685     description => "Reference to unused volumes.",
 686 };
 687
 688 for (my $i = 0; $i < $MAX_UNUSED_DISKS; $i++)  {
 689     $confdesc->{"unused$i"} = $unuseddesc;
 690 }
 691
 692 my $kvm_api_version = 0;
 693
 694 sub kvm_version {
 695
 696     return $kvm_api_version if $kvm_api_version;
 697
 698     my $fh = IO::File->new("</dev/kvm") ||
 699         return 0;
 700
 701     if (my $v = $fh->ioctl(KVM_GET_API_VERSION(), 0)) {
 702         $kvm_api_version = $v;
 703     }
 704
 705     $fh->close();
 706
 707     return  $kvm_api_version;
 708 }
 709
 710 my $kvm_user_version;
 711
 712 sub kvm_user_version {
 713
 714     return $kvm_user_version if $kvm_user_version;
 715
 716     $kvm_user_version = 'unknown';
 717
 718     my $tmp = `kvm -help 2>/dev/null`;
 719
 720     if ($tmp =~ m/^QEMU( PC)? emulator version (\d+\.\d+(\.\d+)?)[,\s]/) {
 721         $kvm_user_version = $2;
 722     }
 723
 724     return $kvm_user_version;
 725
 726 }
 727
 728 my $kernel_has_vhost_net = -c '/dev/vhost-net';
 729
 730 sub disknames {
 731     # order is important - used to autoselect boot disk
 732     return ((map { "ide$_" } (0 .. ($MAX_IDE_DISKS - 1))),
 733             (map { "scsi$_" } (0 .. ($MAX_SCSI_DISKS - 1))),
 734             (map { "virtio$_" } (0 .. ($MAX_VIRTIO_DISKS - 1))),
 735             (map { "sata$_" } (0 .. ($MAX_SATA_DISKS - 1))));
 736 }
 737
 738 sub valid_drivename {
 739     my $dev = shift;
 740
 741     return defined($drivename_hash->{$dev});
 742 }
 743
 744 sub option_exists {
 745     my $key = shift;
 746     return defined($confdesc->{$key});
 747 }
 748
 749 sub nic_models {
 750     return $nic_model_list;
 751 }
 752
 753 sub os_list_description {
 754
 755     return {
 756         other => 'Other',
 757         wxp => 'Windows XP',
 758         w2k => 'Windows 2000',
 759         w2k3 =>, 'Windows 2003',
 760         w2k8 => 'Windows 2008',
 761         wvista => 'Windows Vista',
 762         win7 => 'Windows 7',
 763         win8 => 'Windows 8/2012',
 764         l24 => 'Linux 2.4',
 765         l26 => 'Linux 2.6',
 766     };
 767 }
 768
 769 my $cdrom_path;
 770
 771 sub get_cdrom_path {
 772
 773     return  $cdrom_path if $cdrom_path;
 774
 775     return $cdrom_path = "/dev/cdrom" if -l "/dev/cdrom";
 776     return $cdrom_path = "/dev/cdrom1" if -l "/dev/cdrom1";
 777     return $cdrom_path = "/dev/cdrom2" if -l "/dev/cdrom2";
 778 }
 779
 780 sub get_iso_path {
 781     my ($storecfg, $vmid, $cdrom) = @_;
 782
 783     if ($cdrom eq 'cdrom') {
 784         return get_cdrom_path();
 785     } elsif ($cdrom eq 'none') {
 786         return '';
 787     } elsif ($cdrom =~ m|^/|) {
 788         return $cdrom;
 789     } else {
 790         return PVE::Storage::path($storecfg, $cdrom);
 791     }
 792 }
 793
 794 # try to convert old style file names to volume IDs
 795 sub filename_to_volume_id {
 796     my ($vmid, $file, $media) = @_;
 797
 798     if (!($file eq 'none' || $file eq 'cdrom' ||
 799           $file =~ m|^/dev/.+| || $file =~ m/^([^:]+):(.+)$/)) {
 800
 801         return undef if $file =~ m|/|;
 802
 803         if ($media && $media eq 'cdrom') {
 804             $file = "local:iso/$file";
 805         } else {
 806             $file = "local:$vmid/$file";
 807         }
 808     }
 809
 810     return $file;
 811 }
 812
 813 sub verify_media_type {
 814     my ($opt, $vtype, $media) = @_;
 815
 816     return if !$media;
 817
 818     my $etype;
 819     if ($media eq 'disk') {
 820         $etype = 'images';
 821     } elsif ($media eq 'cdrom') {
 822         $etype = 'iso';
 823     } else {
 824         die "internal error";
 825     }
 826
 827     return if ($vtype eq $etype);
 828
 829     raise_param_exc({ $opt => "unexpected media type ($vtype != $etype)" });
 830 }
 831
 832 sub cleanup_drive_path {
 833     my ($opt, $storecfg, $drive) = @_;
 834
 835     # try to convert filesystem paths to volume IDs
 836
 837     if (($drive->{file} !~ m/^(cdrom|none)$/) &&
 838         ($drive->{file} !~ m|^/dev/.+|) &&
 839         ($drive->{file} !~ m/^([^:]+):(.+)$/) &&
 840         ($drive->{file} !~ m/^\d+$/)) {
 841         my ($vtype, $volid) = PVE::Storage::path_to_volume_id($storecfg, $drive->{file});
 842         raise_param_exc({ $opt => "unable to associate path '$drive->{file}' to any storage"}) if !$vtype;
 843         $drive->{media} = 'cdrom' if !$drive->{media} && $vtype eq 'iso';
 844         verify_media_type($opt, $vtype, $drive->{media});
 845         $drive->{file} = $volid;
 846     }
 847
 848     $drive->{media} = 'cdrom' if !$drive->{media} && $drive->{file} =~ m/^(cdrom|none)$/;
 849 }
 850
 851 sub create_conf_nolock {
 852     my ($vmid, $settings) = @_;
 853
 854     my $filename = config_file($vmid);
 855
 856     die "configuration file '$filename' already exists\n" if -f $filename;
 857
 858     my $defaults = load_defaults();
 859
 860     $settings->{name} = "vm$vmid" if !$settings->{name};
 861     $settings->{memory} = $defaults->{memory} if !$settings->{memory};
 862
 863     my $data = '';
 864     foreach my $opt (keys %$settings) {
 865         next if !$confdesc->{$opt};
 866
 867         my $value = $settings->{$opt};
 868         next if !$value;
 869
 870         $data .= "$opt: $value\n";
 871     }
 872
 873     PVE::Tools::file_set_contents($filename, $data);
 874 }
 875
 876 sub parse_hotplug_features {
 877     my ($data) = @_;
 878
 879     my $res = {};
 880
 881     return $res if $data eq '0';
 882
 883     $data = $confdesc->{hotplug}->{default} if $data eq '1';
 884
 885     foreach my $feature (PVE::Tools::split_list($data)) {
 886         if ($feature =~ m/^(network|disk|cpu|memory|usb)$/) {
 887             $res->{$1} = 1;
 888         } else {
 889             warn "ignoring unknown hotplug feature '$feature'\n";
 890         }
 891     }
 892     return $res;
 893 }
 894
 895 PVE::JSONSchema::register_format('pve-hotplug-features', \&pve_verify_hotplug_features);
 896 sub pve_verify_hotplug_features {
 897     my ($value, $noerr) = @_;
 898
 899     return $value if parse_hotplug_features($value);
 900
 901     return undef if $noerr;
 902
 903     die "unable to parse hotplug option\n";
 904 }
 905
 906 my $parse_size = sub {
 907     my ($value) = @_;
 908
 909     return undef if $value !~ m/^(\d+(\.\d+)?)([KMG])?$/;
 910     my ($size, $unit) = ($1, $3);
 911     if ($unit) {
 912         if ($unit eq 'K') {
 913             $size = $size * 1024;
 914         } elsif ($unit eq 'M') {
 915             $size = $size * 1024 * 1024;
 916         } elsif ($unit eq 'G') {
 917             $size = $size * 1024 * 1024 * 1024;
 918         }
 919     }
 920     return int($size);
 921 };
 922
 923 my $format_size = sub {
 924     my ($size) = @_;
 925
 926     $size = int($size);
 927
 928     my $kb = int($size/1024);
 929     return $size if $kb*1024 != $size;
 930
 931     my $mb = int($kb/1024);
 932     return "${kb}K" if $mb*1024 != $kb;
 933
 934     my $gb = int($mb/1024);
 935     return "${mb}M" if $gb*1024 != $mb;
 936
 937     return "${gb}G";
 938 };
 939
 940 # ideX = [volume=]volume-id[,media=d][,cyls=c,heads=h,secs=s[,trans=t]]
 941 #        [,snapshot=on|off][,cache=on|off][,format=f][,backup=yes|no]
 942 #        [,rerror=ignore|report|stop][,werror=enospc|ignore|report|stop]
 943 #        [,aio=native|threads][,discard=ignore|on]
 944
 945 sub parse_drive {
 946     my ($key, $data) = @_;
 947
 948     my $res = {};
 949
 950     # $key may be undefined - used to verify JSON parameters
 951     if (!defined($key)) {
 952         $res->{interface} = 'unknown'; # should not harm when used to verify parameters
 953         $res->{index} = 0;
 954     } elsif ($key =~ m/^([^\d]+)(\d+)$/) {
 955         $res->{interface} = $1;
 956         $res->{index} = $2;
 957     } else {
 958         return undef;
 959     }
 960
 961     foreach my $p (split (/,/, $data)) {
 962         next if $p =~ m/^\s*$/;
 963
 964         if ($p =~ m/^(file|volume|cyls|heads|secs|trans|media|snapshot|cache|format|rerror|werror|backup|aio|bps|mbps|mbps_max|bps_rd|mbps_rd|mbps_rd_max|bps_wr|mbps_wr|mbps_wr_max|iops|iops_max|iops_rd|iops_rd_max|iops_wr|iops_wr_max|size|discard)=(.+)$/) {
 965             my ($k, $v) = ($1, $2);
 966
 967             $k = 'file' if $k eq 'volume';
 968
 969             return undef if defined $res->{$k};
 970
 971             if ($k eq 'bps' || $k eq 'bps_rd' || $k eq 'bps_wr') {
 972                 return undef if !$v || $v !~ m/^\d+/;
 973                 $k = "m$k";
 974                 $v = sprintf("%.3f", $v / (1024*1024));
 975             }
 976             $res->{$k} = $v;
 977         } else {
 978             if (!$res->{file} && $p !~ m/=/) {
 979                 $res->{file} = $p;
 980             } else {
 981                 return undef;
 982             }
 983         }
 984     }
 985
 986     return undef if !$res->{file};
 987
 988     if($res->{file} =~ m/\.(raw|cow|qcow|qcow2|vmdk|cloop)$/){
 989         $res->{format} = $1;
 990     }
 991
 992     return undef if $res->{cache} &&
 993         $res->{cache} !~ m/^(off|none|writethrough|writeback|unsafe|directsync)$/;
 994     return undef if $res->{snapshot} && $res->{snapshot} !~ m/^(on|off)$/;
 995     return undef if $res->{cyls} && $res->{cyls} !~ m/^\d+$/;
 996     return undef if $res->{heads} && $res->{heads} !~ m/^\d+$/;
 997     return undef if $res->{secs} && $res->{secs} !~ m/^\d+$/;
 998     return undef if $res->{media} && $res->{media} !~ m/^(disk|cdrom)$/;
 999     return undef if $res->{trans} && $res->{trans} !~ m/^(none|lba|auto)$/;
1000     return undef if $res->{format} && $res->{format} !~ m/^(raw|cow|qcow|qcow2|vmdk|cloop)$/;
1001     return undef if $res->{rerror} && $res->{rerror} !~ m/^(ignore|report|stop)$/;
1002     return undef if $res->{werror} && $res->{werror} !~ m/^(enospc|ignore|report|stop)$/;
1003     return undef if $res->{backup} && $res->{backup} !~ m/^(yes|no)$/;
1004     return undef if $res->{aio} && $res->{aio} !~ m/^(native|threads)$/;
1005     return undef if $res->{discard} && $res->{discard} !~ m/^(ignore|on)$/;
1006
1007     return undef if $res->{mbps_rd} && $res->{mbps};
1008     return undef if $res->{mbps_wr} && $res->{mbps};
1009
1010     return undef if $res->{mbps} && $res->{mbps} !~ m/^\d+(\.\d+)?$/;
1011     return undef if $res->{mbps_max} && $res->{mbps_max} !~ m/^\d+(\.\d+)?$/;
1012     return undef if $res->{mbps_rd} && $res->{mbps_rd} !~ m/^\d+(\.\d+)?$/;
1013     return undef if $res->{mbps_rd_max} && $res->{mbps_rd_max} !~ m/^\d+(\.\d+)?$/;
1014     return undef if $res->{mbps_wr} && $res->{mbps_wr} !~ m/^\d+(\.\d+)?$/;
1015     return undef if $res->{mbps_wr_max} && $res->{mbps_wr_max} !~ m/^\d+(\.\d+)?$/;
1016
1017     return undef if $res->{iops_rd} && $res->{iops};
1018     return undef if $res->{iops_wr} && $res->{iops};
1019
1020
1021     return undef if $res->{iops} && $res->{iops} !~ m/^\d+$/;
1022     return undef if $res->{iops_max} && $res->{iops_max} !~ m/^\d+$/;
1023     return undef if $res->{iops_rd} && $res->{iops_rd} !~ m/^\d+$/;
1024     return undef if $res->{iops_rd_max} && $res->{iops_rd_max} !~ m/^\d+$/;
1025     return undef if $res->{iops_wr} && $res->{iops_wr} !~ m/^\d+$/;
1026     return undef if $res->{iops_wr_max} && $res->{iops_wr_max} !~ m/^\d+$/;
1027
1028
1029     if ($res->{size}) {
1030         return undef if !defined($res->{size} = &$parse_size($res->{size}));
1031     }
1032
1033     if ($res->{media} && ($res->{media} eq 'cdrom')) {
1034         return undef if $res->{snapshot} || $res->{trans} || $res->{format};
1035         return undef if $res->{heads} || $res->{secs} || $res->{cyls};
1036         return undef if $res->{interface} eq 'virtio';
1037     }
1038
1039     # rerror does not work with scsi drives
1040     if ($res->{rerror}) {
1041         return undef if $res->{interface} eq 'scsi';
1042     }
1043
1044     return $res;
1045 }
1046
1047 my @qemu_drive_options = qw(heads secs cyls trans media format cache snapshot rerror werror aio discard iops iops_rd iops_wr iops_max iops_rd_max iops_wr_max);
1048
1049 sub print_drive {
1050     my ($vmid, $drive) = @_;
1051
1052     my $opts = '';
1053     foreach my $o (@qemu_drive_options, 'mbps', 'mbps_rd', 'mbps_wr', 'mbps_max', 'mbps_rd_max', 'mbps_wr_max', 'backup') {
1054         $opts .= ",$o=$drive->{$o}" if $drive->{$o};
1055     }
1056
1057     if ($drive->{size}) {
1058         $opts .= ",size=" . &$format_size($drive->{size});
1059     }
1060
1061     return "$drive->{file}$opts";
1062 }
1063
1064 sub scsi_inquiry {
1065     my($fh, $noerr) = @_;
1066
1067     my $SG_IO = 0x2285;
1068     my $SG_GET_VERSION_NUM = 0x2282;
1069
1070     my $versionbuf = "\x00" x 8;
1071     my $ret = ioctl($fh, $SG_GET_VERSION_NUM, $versionbuf);
1072     if (!$ret) {
1073         die "scsi ioctl SG_GET_VERSION_NUM failoed - $!\n" if !$noerr;
1074         return undef;
1075     }
1076     my $version = unpack("I", $versionbuf);
1077     if ($version < 30000) {
1078         die "scsi generic interface too old\n"  if !$noerr;
1079         return undef;
1080     }
1081
1082     my $buf = "\x00" x 36;
1083     my $sensebuf = "\x00" x 8;
1084     my $cmd = pack("C x3 C x1", 0x12, 36);
1085
1086     # see /usr/include/scsi/sg.h
1087     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";
1088
1089     my $packet = pack($sg_io_hdr_t, ord('S'), -3, length($cmd),
1090                       length($sensebuf), 0, length($buf), $buf,
1091                       $cmd, $sensebuf, 6000);
1092
1093     $ret = ioctl($fh, $SG_IO, $packet);
1094     if (!$ret) {
1095         die "scsi ioctl SG_IO failed - $!\n" if !$noerr;
1096         return undef;
1097     }
1098
1099     my @res = unpack($sg_io_hdr_t, $packet);
1100     if ($res[17] || $res[18]) {
1101         die "scsi ioctl SG_IO status error - $!\n" if !$noerr;
1102         return undef;
1103     }
1104
1105     my $res = {};
1106     (my $byte0, my $byte1, $res->{vendor},
1107      $res->{product}, $res->{revision}) = unpack("C C x6 A8 A16 A4", $buf);
1108
1109     $res->{removable} = $byte1 & 128 ? 1 : 0;
1110     $res->{type} = $byte0 & 31;
1111
1112     return $res;
1113 }
1114
1115 sub path_is_scsi {
1116     my ($path) = @_;
1117
1118     my $fh = IO::File->new("+<$path") || return undef;
1119     my $res = scsi_inquiry($fh, 1);
1120     close($fh);
1121
1122     return $res;
1123 }
1124
1125 sub machine_type_is_q35 {
1126     my ($conf) = @_;
1127
1128     return $conf->{machine} && ($conf->{machine} =~ m/q35/) ? 1 : 0;
1129 }
1130
1131 sub print_tabletdevice_full {
1132     my ($conf) = @_;
1133
1134     my $q35 = machine_type_is_q35($conf);
1135
1136     # we use uhci for old VMs because tablet driver was buggy in older qemu
1137     my $usbbus = $q35 ? "ehci" : "uhci";
1138
1139     return "usb-tablet,id=tablet,bus=$usbbus.0,port=1";
1140 }
1141
1142 sub print_drivedevice_full {
1143     my ($storecfg, $conf, $vmid, $drive, $bridges) = @_;
1144
1145     my $device = '';
1146     my $maxdev = 0;
1147
1148     if ($drive->{interface} eq 'virtio') {
1149         my $pciaddr = print_pci_addr("$drive->{interface}$drive->{index}", $bridges);
1150         $device = "virtio-blk-pci,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}$pciaddr";
1151         $device .= ",iothread=iothread0" if $conf->{iothread};
1152     } elsif ($drive->{interface} eq 'scsi') {
1153         $maxdev = ($conf->{scsihw} && ($conf->{scsihw} !~ m/^lsi/)) ? 256 : 7;
1154         my $controller = int($drive->{index} / $maxdev);
1155         my $unit = $drive->{index} % $maxdev;
1156         my $devicetype = 'hd';
1157         my $path = '';
1158         if (drive_is_cdrom($drive)) {
1159               $devicetype = 'cd';
1160         } else {
1161               if ($drive->{file} =~ m|^/|) {
1162                   $path = $drive->{file};
1163               } else {
1164                   $path = PVE::Storage::path($storecfg, $drive->{file});
1165               }
1166
1167               if($path =~ m/^iscsi\:\/\//){
1168                   $devicetype = 'generic';
1169               } else {
1170                   if (my $info = path_is_scsi($path)) {
1171                       if ($info->{type} == 0) {
1172                           $devicetype = 'block';
1173                       } elsif ($info->{type} == 1) { # tape
1174                           $devicetype = 'generic';
1175                       }
1176                   }
1177               }
1178          }
1179
1180         if (!$conf->{scsihw} || ($conf->{scsihw} =~ m/^lsi/)){
1181             $device = "scsi-$devicetype,bus=scsihw$controller.0,scsi-id=$unit,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}";
1182         } else {
1183             $device = "scsi-$devicetype,bus=scsihw$controller.0,channel=0,scsi-id=0,lun=$drive->{index},drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}";
1184         }
1185
1186     } elsif ($drive->{interface} eq 'ide'){
1187         $maxdev = 2;
1188         my $controller = int($drive->{index} / $maxdev);
1189         my $unit = $drive->{index} % $maxdev;
1190         my $devicetype = ($drive->{media} && $drive->{media} eq 'cdrom') ? "cd" : "hd";
1191
1192         $device = "ide-$devicetype,bus=ide.$controller,unit=$unit,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}";
1193     } elsif ($drive->{interface} eq 'sata'){
1194         my $controller = int($drive->{index} / $MAX_SATA_DISKS);
1195         my $unit = $drive->{index} % $MAX_SATA_DISKS;
1196         $device = "ide-drive,bus=ahci$controller.$unit,drive=drive-$drive->{interface}$drive->{index},id=$drive->{interface}$drive->{index}";
1197     } elsif ($drive->{interface} eq 'usb') {
1198         die "implement me";
1199         #  -device ide-drive,bus=ide.1,unit=0,drive=drive-ide0-1-0,id=ide0-1-0
1200     } else {
1201         die "unsupported interface type";
1202     }
1203
1204     $device .= ",bootindex=$drive->{bootindex}" if $drive->{bootindex};
1205
1206     return $device;
1207 }
1208
1209 sub get_initiator_name {
1210     my $initiator;
1211
1212     my $fh = IO::File->new('/etc/iscsi/initiatorname.iscsi') || return undef;
1213     while (defined(my $line = <$fh>)) {
1214         next if $line !~ m/^\s*InitiatorName\s*=\s*([\.\-:\w]+)/;
1215         $initiator = $1;
1216         last;
1217     }
1218     $fh->close();
1219
1220     return $initiator;
1221 }
1222
1223 sub print_drive_full {
1224     my ($storecfg, $vmid, $drive) = @_;
1225
1226     my $opts = '';
1227     foreach my $o (@qemu_drive_options) {
1228         next if $o eq 'bootindex';
1229         $opts .= ",$o=$drive->{$o}" if $drive->{$o};
1230     }
1231
1232     foreach my $o (qw(bps bps_rd bps_wr)) {
1233         my $v = $drive->{"m$o"};
1234         $opts .= ",$o=" . int($v*1024*1024) if $v;
1235     }
1236
1237     # use linux-aio by default (qemu default is threads)
1238     $opts .= ",aio=native" if !$drive->{aio};
1239
1240     my $path;
1241     my $volid = $drive->{file};
1242     if (drive_is_cdrom($drive)) {
1243         $path = get_iso_path($storecfg, $vmid, $volid);
1244     } else {
1245         if ($volid =~ m|^/|) {
1246             $path = $volid;
1247         } else {
1248             $path = PVE::Storage::path($storecfg, $volid);
1249         }
1250     }
1251
1252     $opts .= ",cache=none" if !$drive->{cache} && !drive_is_cdrom($drive);
1253
1254     my $detectzeroes = $drive->{discard} ? "unmap" : "on";
1255     $opts .= ",detect-zeroes=$detectzeroes" if !drive_is_cdrom($drive);
1256
1257     my $pathinfo = $path ? "file=$path," : '';
1258
1259     return "${pathinfo}if=none,id=drive-$drive->{interface}$drive->{index}$opts";
1260 }
1261
1262 sub print_netdevice_full {
1263     my ($vmid, $conf, $net, $netid, $bridges) = @_;
1264
1265     my $bootorder = $conf->{boot} || $confdesc->{boot}->{default};
1266
1267     my $device = $net->{model};
1268     if ($net->{model} eq 'virtio') {
1269          $device = 'virtio-net-pci';
1270      };
1271
1272     my $pciaddr = print_pci_addr("$netid", $bridges);
1273     my $tmpstr = "$device,mac=$net->{macaddr},netdev=$netid$pciaddr,id=$netid";
1274     if ($net->{queues} && $net->{queues} > 1 && $net->{model} eq 'virtio'){
1275         #Consider we have N queues, the number of vectors needed is 2*N + 2 (plus one config interrupt and control vq)
1276         my $vectors = $net->{queues} * 2 + 2;
1277         $tmpstr .= ",vectors=$vectors,mq=on";
1278     }
1279     $tmpstr .= ",bootindex=$net->{bootindex}" if $net->{bootindex} ;
1280     return $tmpstr;
1281 }
1282
1283 sub print_netdev_full {
1284     my ($vmid, $conf, $net, $netid) = @_;
1285
1286     my $i = '';
1287     if ($netid =~ m/^net(\d+)$/) {
1288         $i = int($1);
1289     }
1290
1291     die "got strange net id '$i'\n" if $i >= ${MAX_NETS};
1292
1293     my $ifname = "tap${vmid}i$i";
1294
1295     # kvm uses TUNSETIFF ioctl, and that limits ifname length
1296     die "interface name '$ifname' is too long (max 15 character)\n"
1297         if length($ifname) >= 16;
1298
1299     my $vhostparam = '';
1300     $vhostparam = ',vhost=on' if $kernel_has_vhost_net && $net->{model} eq 'virtio';
1301
1302     my $vmname = $conf->{name} || "vm$vmid";
1303
1304     my $netdev = "";
1305
1306     if ($net->{bridge}) {
1307         $netdev = "type=tap,id=$netid,ifname=${ifname},script=/var/lib/qemu-server/pve-bridge,downscript=/var/lib/qemu-server/pve-bridgedown$vhostparam";
1308     } else {
1309         $netdev = "type=user,id=$netid,hostname=$vmname";
1310     }
1311
1312     $netdev .= ",queues=$net->{queues}" if ($net->{queues} && $net->{model} eq 'virtio');
1313
1314     return $netdev;
1315 }
1316
1317 sub drive_is_cdrom {
1318     my ($drive) = @_;
1319
1320     return $drive && $drive->{media} && ($drive->{media} eq 'cdrom');
1321
1322 }
1323
1324 sub parse_numa {
1325     my ($data) = @_;
1326
1327     my $res = {};
1328
1329     foreach my $kvp (split(/,/, $data)) {
1330
1331         if ($kvp =~ m/^memory=(\S+)$/) {
1332             $res->{memory} = $1;
1333         } elsif ($kvp =~ m/^policy=(preferred|bind|interleave)$/) {
1334             $res->{policy} = $1;
1335         } elsif ($kvp =~ m/^cpus=(\d+)(-(\d+))?$/) {
1336             $res->{cpus}->{start} = $1;
1337             $res->{cpus}->{end} = $3;
1338         } elsif ($kvp =~ m/^hostnodes=(\d+)(-(\d+))?$/) {
1339             $res->{hostnodes}->{start} = $1;
1340             $res->{hostnodes}->{end} = $3;
1341         } else {
1342             return undef;
1343         }
1344     }
1345
1346     return $res;
1347 }
1348
1349 sub parse_hostpci {
1350     my ($value) = @_;
1351
1352     return undef if !$value;
1353
1354
1355     my @list = split(/,/, $value);
1356     my $found;
1357
1358     my $res = {};
1359     foreach my $kv (@list) {
1360
1361         if ($kv =~ m/^(host=)?([a-f0-9]{2}:[a-f0-9]{2})(\.([a-f0-9]))?$/) {
1362             $found = 1;
1363             if(defined($4)){
1364                 push @{$res->{pciid}}, { id => $2 , function => $4};
1365
1366             }else{
1367                 my $pcidevices = lspci($2);
1368                 $res->{pciid} = $pcidevices->{$2};
1369             }
1370         } elsif ($kv =~ m/^driver=(kvm|vfio)$/) {
1371             $res->{driver} = $1;
1372         } elsif ($kv =~ m/^rombar=(on|off)$/) {
1373             $res->{rombar} = $1;
1374         } elsif ($kv =~ m/^x-vga=(on|off)$/) {
1375             $res->{'x-vga'} = $1;
1376         } elsif ($kv =~ m/^pcie=(\d+)$/) {
1377             $res->{pcie} = 1 if $1 == 1;
1378         } else {
1379             warn "unknown hostpci setting '$kv'\n";
1380         }
1381     }
1382
1383     return undef if !$found;
1384
1385     return $res;
1386 }
1387
1388 # netX: e1000=XX:XX:XX:XX:XX:XX,bridge=vmbr0,rate=<mbps>
1389 sub parse_net {
1390     my ($data) = @_;
1391
1392     my $res = {};
1393
1394     foreach my $kvp (split(/,/, $data)) {
1395
1396         if ($kvp =~ m/^(ne2k_pci|e1000|e1000-82540em|e1000-82544gc|e1000-82545em|rtl8139|pcnet|virtio|ne2k_isa|i82551|i82557b|i82559er|vmxnet3)(=([0-9a-f]{2}(:[0-9a-f]{2}){5}))?$/i) {
1397             my $model = lc($1);
1398             my $mac = defined($3) ? uc($3) : PVE::Tools::random_ether_addr();
1399             $res->{model} = $model;
1400             $res->{macaddr} = $mac;
1401         } elsif ($kvp =~ m/^bridge=(\S+)$/) {
1402             $res->{bridge} = $1;
1403         } elsif ($kvp =~ m/^queues=(\d+)$/) {
1404             $res->{queues} = $1;
1405         } elsif ($kvp =~ m/^rate=(\d+(\.\d+)?)$/) {
1406             $res->{rate} = $1;
1407         } elsif ($kvp =~ m/^tag=(\d+)$/) {
1408             $res->{tag} = $1;
1409         } elsif ($kvp =~ m/^firewall=([01])$/) {
1410             $res->{firewall} = $1;
1411         } elsif ($kvp =~ m/^link_down=([01])$/) {
1412             $res->{link_down} = $1;
1413         } else {
1414             return undef;
1415         }
1416
1417     }
1418
1419     return undef if !$res->{model};
1420
1421     return $res;
1422 }
1423
1424 sub print_net {
1425     my $net = shift;
1426
1427     my $res = "$net->{model}";
1428     $res .= "=$net->{macaddr}" if $net->{macaddr};
1429     $res .= ",bridge=$net->{bridge}" if $net->{bridge};
1430     $res .= ",rate=$net->{rate}" if $net->{rate};
1431     $res .= ",tag=$net->{tag}" if $net->{tag};
1432     $res .= ",firewall=1" if $net->{firewall};
1433     $res .= ",link_down=1" if $net->{link_down};
1434     $res .= ",queues=$net->{queues}" if $net->{queues};
1435
1436     return $res;
1437 }
1438
1439 sub add_random_macs {
1440     my ($settings) = @_;
1441
1442     foreach my $opt (keys %$settings) {
1443         next if $opt !~ m/^net(\d+)$/;
1444         my $net = parse_net($settings->{$opt});
1445         next if !$net;
1446         $settings->{$opt} = print_net($net);
1447     }
1448 }
1449
1450 sub add_unused_volume {
1451     my ($config, $volid) = @_;
1452
1453     my $key;
1454     for (my $ind = $MAX_UNUSED_DISKS - 1; $ind >= 0; $ind--) {
1455         my $test = "unused$ind";
1456         if (my $vid = $config->{$test}) {
1457             return if $vid eq $volid; # do not add duplicates
1458         } else {
1459             $key = $test;
1460         }
1461     }
1462
1463     die "To many unused volume - please delete them first.\n" if !$key;
1464
1465     $config->{$key} = $volid;
1466
1467     return $key;
1468 }
1469
1470 sub vm_is_volid_owner {
1471     my ($storecfg, $vmid, $volid) = @_;
1472
1473     if ($volid !~  m|^/|) {
1474         my ($path, $owner);
1475         eval { ($path, $owner) = PVE::Storage::path($storecfg, $volid); };
1476         if ($owner && ($owner == $vmid)) {
1477             return 1;
1478         }
1479     }
1480
1481     return undef;
1482 }
1483
1484 sub vmconfig_delete_pending_option {
1485     my ($conf, $key) = @_;
1486
1487     delete $conf->{pending}->{$key};
1488     my $pending_delete_hash = { $key => 1 };
1489     foreach my $opt (PVE::Tools::split_list($conf->{pending}->{delete})) {
1490         $pending_delete_hash->{$opt} = 1;
1491     }
1492     $conf->{pending}->{delete} = join(',', keys %$pending_delete_hash);
1493 }
1494
1495 sub vmconfig_undelete_pending_option {
1496     my ($conf, $key) = @_;
1497
1498     my $pending_delete_hash = {};
1499     foreach my $opt (PVE::Tools::split_list($conf->{pending}->{delete})) {
1500         $pending_delete_hash->{$opt} = 1;
1501     }
1502     delete $pending_delete_hash->{$key};
1503
1504     my @keylist = keys %$pending_delete_hash;
1505     if (scalar(@keylist)) {
1506         $conf->{pending}->{delete} = join(',', @keylist);
1507     } else {
1508         delete $conf->{pending}->{delete};
1509     }
1510 }
1511
1512 sub vmconfig_register_unused_drive {
1513     my ($storecfg, $vmid, $conf, $drive) = @_;
1514
1515     if (!drive_is_cdrom($drive)) {
1516         my $volid = $drive->{file};
1517         if (vm_is_volid_owner($storecfg, $vmid, $volid)) {
1518             add_unused_volume($conf, $volid, $vmid);
1519         }
1520     }
1521 }
1522
1523 sub vmconfig_cleanup_pending {
1524     my ($conf) = @_;
1525
1526     # remove pending changes when nothing changed
1527     my $changes;
1528     foreach my $opt (keys %{$conf->{pending}}) {
1529         if (defined($conf->{$opt}) && ($conf->{pending}->{$opt} eq  $conf->{$opt})) {
1530             $changes = 1;
1531             delete $conf->{pending}->{$opt};
1532         }
1533     }
1534
1535     # remove delete if option is not set
1536     my $pending_delete_hash = {};
1537     foreach my $opt (PVE::Tools::split_list($conf->{pending}->{delete})) {
1538         if (defined($conf->{$opt})) {
1539             $pending_delete_hash->{$opt} = 1;
1540         } else {
1541             $changes = 1;
1542         }
1543     }
1544
1545     my @keylist = keys %$pending_delete_hash;
1546     if (scalar(@keylist)) {
1547         $conf->{pending}->{delete} = join(',', @keylist);
1548     } else {
1549         delete $conf->{pending}->{delete};
1550     }
1551
1552     return $changes;
1553 }
1554
1555 my $valid_smbios1_options = {
1556     manufacturer => '\S+',
1557     product => '\S+',
1558     version => '\S+',
1559     serial => '\S+',
1560     uuid => '[a-fA-F0-9]{8}(?:-[a-fA-F0-9]{4}){3}-[a-fA-F0-9]{12}',
1561     sku => '\S+',
1562     family => '\S+',
1563 };
1564
1565 # smbios: [manufacturer=str][,product=str][,version=str][,serial=str][,uuid=uuid][,sku=str][,family=str]
1566 sub parse_smbios1 {
1567     my ($data) = @_;
1568
1569     my $res = {};
1570
1571     foreach my $kvp (split(/,/, $data)) {
1572         return undef if $kvp !~ m/^(\S+)=(.+)$/;
1573         my ($k, $v) = split(/=/, $kvp);
1574         return undef if !defined($k) || !defined($v);
1575         return undef if !$valid_smbios1_options->{$k};
1576         return undef if $v !~ m/^$valid_smbios1_options->{$k}$/;
1577         $res->{$k} = $v;
1578     }
1579
1580     return $res;
1581 }
1582
1583 sub print_smbios1 {
1584     my ($smbios1) = @_;
1585
1586     my $data = '';
1587     foreach my $k (keys %$smbios1) {
1588         next if !defined($smbios1->{$k});
1589         next if !$valid_smbios1_options->{$k};
1590         $data .= ',' if $data;
1591         $data .= "$k=$smbios1->{$k}";
1592     }
1593     return $data;
1594 }
1595
1596 PVE::JSONSchema::register_format('pve-qm-smbios1', \&verify_smbios1);
1597 sub verify_smbios1 {
1598     my ($value, $noerr) = @_;
1599
1600     return $value if parse_smbios1($value);
1601
1602     return undef if $noerr;
1603
1604     die "unable to parse smbios (type 1) options\n";
1605 }
1606
1607 PVE::JSONSchema::register_format('pve-qm-bootdisk', \&verify_bootdisk);
1608 sub verify_bootdisk {
1609     my ($value, $noerr) = @_;
1610
1611     return $value if valid_drivename($value);
1612
1613     return undef if $noerr;
1614
1615     die "invalid boot disk '$value'\n";
1616 }
1617
1618 PVE::JSONSchema::register_format('pve-qm-numanode', \&verify_numa);
1619 sub verify_numa {
1620     my ($value, $noerr) = @_;
1621
1622     return $value if parse_numa($value);
1623
1624     return undef if $noerr;
1625
1626     die "unable to parse numa options\n";
1627 }
1628
1629 PVE::JSONSchema::register_format('pve-qm-net', \&verify_net);
1630 sub verify_net {
1631     my ($value, $noerr) = @_;
1632
1633     return $value if parse_net($value);
1634
1635     return undef if $noerr;
1636
1637     die "unable to parse network options\n";
1638 }
1639
1640 PVE::JSONSchema::register_format('pve-qm-drive', \&verify_drive);
1641 sub verify_drive {
1642     my ($value, $noerr) = @_;
1643
1644     return $value if parse_drive(undef, $value);
1645
1646     return undef if $noerr;
1647
1648     die "unable to parse drive options\n";
1649 }
1650
1651 PVE::JSONSchema::register_format('pve-qm-hostpci', \&verify_hostpci);
1652 sub verify_hostpci {
1653     my ($value, $noerr) = @_;
1654
1655     return $value if parse_hostpci($value);
1656
1657     return undef if $noerr;
1658
1659     die "unable to parse pci id\n";
1660 }
1661
1662 PVE::JSONSchema::register_format('pve-qm-watchdog', \&verify_watchdog);
1663 sub verify_watchdog {
1664     my ($value, $noerr) = @_;
1665
1666     return $value if parse_watchdog($value);
1667
1668     return undef if $noerr;
1669
1670     die "unable to parse watchdog options\n";
1671 }
1672
1673 sub parse_watchdog {
1674     my ($value) = @_;
1675
1676     return undef if !$value;
1677
1678     my $res = {};
1679
1680     foreach my $p (split(/,/, $value)) {
1681         next if $p =~ m/^\s*$/;
1682
1683         if ($p =~ m/^(model=)?(i6300esb|ib700)$/) {
1684             $res->{model} = $2;
1685         } elsif ($p =~ m/^(action=)?(reset|shutdown|poweroff|pause|debug|none)$/) {
1686             $res->{action} = $2;
1687         } else {
1688             return undef;
1689         }
1690     }
1691
1692     return $res;
1693 }
1694
1695 PVE::JSONSchema::register_format('pve-qm-startup', \&verify_startup);
1696 sub verify_startup {
1697     my ($value, $noerr) = @_;
1698
1699     return $value if parse_startup($value);
1700
1701     return undef if $noerr;
1702
1703     die "unable to parse startup options\n";
1704 }
1705
1706 sub parse_startup {
1707     my ($value) = @_;
1708
1709     return undef if !$value;
1710
1711     my $res = {};
1712
1713     foreach my $p (split(/,/, $value)) {
1714         next if $p =~ m/^\s*$/;
1715
1716         if ($p =~ m/^(order=)?(\d+)$/) {
1717             $res->{order} = $2;
1718         } elsif ($p =~ m/^up=(\d+)$/) {
1719             $res->{up} = $1;
1720         } elsif ($p =~ m/^down=(\d+)$/) {
1721             $res->{down} = $1;
1722         } else {
1723             return undef;
1724         }
1725     }
1726
1727     return $res;
1728 }
1729
1730 sub parse_usb_device {
1731     my ($value) = @_;
1732
1733     return undef if !$value;
1734
1735     my @dl = split(/,/, $value);
1736     my $found;
1737
1738     my $res = {};
1739     foreach my $v (@dl) {
1740         if ($v =~ m/^host=(0x)?([0-9A-Fa-f]{4}):(0x)?([0-9A-Fa-f]{4})$/) {
1741             $found = 1;
1742             $res->{vendorid} = $2;
1743             $res->{productid} = $4;
1744         } elsif ($v =~ m/^host=(\d+)\-(\d+(\.\d+)*)$/) {
1745             $found = 1;
1746             $res->{hostbus} = $1;
1747             $res->{hostport} = $2;
1748         } elsif ($v =~ m/^spice$/) {
1749             $found = 1;
1750             $res->{spice} = 1;
1751         } else {
1752             return undef;
1753         }
1754     }
1755     return undef if !$found;
1756
1757     return $res;
1758 }
1759
1760 PVE::JSONSchema::register_format('pve-qm-usb-device', \&verify_usb_device);
1761 sub verify_usb_device {
1762     my ($value, $noerr) = @_;
1763
1764     return $value if parse_usb_device($value);
1765
1766     return undef if $noerr;
1767
1768     die "unable to parse usb device\n";
1769 }
1770
1771 # add JSON properties for create and set function
1772 sub json_config_properties {
1773     my $prop = shift;
1774
1775     foreach my $opt (keys %$confdesc) {
1776         next if $opt eq 'parent' || $opt eq 'snaptime' || $opt eq 'vmstate';
1777         $prop->{$opt} = $confdesc->{$opt};
1778     }
1779
1780     return $prop;
1781 }
1782
1783 sub check_type {
1784     my ($key, $value) = @_;
1785
1786     die "unknown setting '$key'\n" if !$confdesc->{$key};
1787
1788     my $type = $confdesc->{$key}->{type};
1789
1790     if (!defined($value)) {
1791         die "got undefined value\n";
1792     }
1793
1794     if ($value =~ m/[\n\r]/) {
1795         die "property contains a line feed\n";
1796     }
1797
1798     if ($type eq 'boolean') {
1799         return 1 if ($value eq '1') || ($value =~ m/^(on|yes|true)$/i);
1800         return 0 if ($value eq '0') || ($value =~ m/^(off|no|false)$/i);
1801         die "type check ('boolean') failed - got '$value'\n";
1802     } elsif ($type eq 'integer') {
1803         return int($1) if $value =~ m/^(\d+)$/;
1804         die "type check ('integer') failed - got '$value'\n";
1805     } elsif ($type eq 'number') {
1806         return $value if $value =~ m/^(\d+)(\.\d+)?$/;
1807         die "type check ('number') failed - got '$value'\n";
1808     } elsif ($type eq 'string') {
1809         if (my $fmt = $confdesc->{$key}->{format}) {
1810             if ($fmt eq 'pve-qm-drive') {
1811                 # special case - we need to pass $key to parse_drive()
1812                 my $drive = parse_drive($key, $value);
1813                 return $value if $drive;
1814                 die "unable to parse drive options\n";
1815             }
1816             PVE::JSONSchema::check_format($fmt, $value);
1817             return $value;
1818         }
1819         $value =~ s/^\"(.*)\"$/$1/;
1820         return $value;
1821     } else {
1822         die "internal error"
1823     }
1824 }
1825
1826 sub lock_config_full {
1827     my ($vmid, $timeout, $code, @param) = @_;
1828
1829     my $filename = config_file_lock($vmid);
1830
1831     my $res = lock_file($filename, $timeout, $code, @param);
1832
1833     die $@ if $@;
1834
1835     return $res;
1836 }
1837
1838 sub lock_config_mode {
1839     my ($vmid, $timeout, $shared, $code, @param) = @_;
1840
1841     my $filename = config_file_lock($vmid);
1842
1843     my $res = lock_file_full($filename, $timeout, $shared, $code, @param);
1844
1845     die $@ if $@;
1846
1847     return $res;
1848 }
1849
1850 sub lock_config {
1851     my ($vmid, $code, @param) = @_;
1852
1853     return lock_config_full($vmid, 10, $code, @param);
1854 }
1855
1856 sub cfs_config_path {
1857     my ($vmid, $node) = @_;
1858
1859     $node = $nodename if !$node;
1860     return "nodes/$node/qemu-server/$vmid.conf";
1861 }
1862
1863 sub check_iommu_support{
1864     #fixme : need to check IOMMU support
1865     #http://www.linux-kvm.org/page/How_to_assign_devices_with_VT-d_in_KVM
1866
1867     my $iommu=1;
1868     return $iommu;
1869
1870 }
1871
1872 sub config_file {
1873     my ($vmid, $node) = @_;
1874
1875     my $cfspath = cfs_config_path($vmid, $node);
1876     return "/etc/pve/$cfspath";
1877 }
1878
1879 sub config_file_lock {
1880     my ($vmid) = @_;
1881
1882     return "$lock_dir/lock-$vmid.conf";
1883 }
1884
1885 sub touch_config {
1886     my ($vmid) = @_;
1887
1888     my $conf = config_file($vmid);
1889     utime undef, undef, $conf;
1890 }
1891
1892 sub destroy_vm {
1893     my ($storecfg, $vmid, $keep_empty_config) = @_;
1894
1895     my $conffile = config_file($vmid);
1896
1897     my $conf = load_config($vmid);
1898
1899     check_lock($conf);
1900
1901     # only remove disks owned by this VM
1902     foreach_drive($conf, sub {
1903         my ($ds, $drive) = @_;
1904
1905         return if drive_is_cdrom($drive);
1906
1907         my $volid = $drive->{file};
1908
1909         return if !$volid || $volid =~ m|^/|;
1910
1911         my ($path, $owner) = PVE::Storage::path($storecfg, $volid);
1912         return if !$path || !$owner || ($owner != $vmid);
1913
1914         PVE::Storage::vdisk_free($storecfg, $volid);
1915     });
1916
1917     if ($keep_empty_config) {
1918         PVE::Tools::file_set_contents($conffile, "memory: 128\n");
1919     } else {
1920         unlink $conffile;
1921     }
1922
1923     # also remove unused disk
1924     eval {
1925         my $dl = PVE::Storage::vdisk_list($storecfg, undef, $vmid);
1926
1927         eval {
1928             PVE::Storage::foreach_volid($dl, sub {
1929                 my ($volid, $sid, $volname, $d) = @_;
1930                 PVE::Storage::vdisk_free($storecfg, $volid);
1931             });
1932         };
1933         warn $@ if $@;
1934
1935     };
1936     warn $@ if $@;
1937 }
1938
1939 sub load_config {
1940     my ($vmid, $node) = @_;
1941
1942     my $cfspath = cfs_config_path($vmid, $node);
1943
1944     my $conf = PVE::Cluster::cfs_read_file($cfspath);
1945
1946     die "no such VM ('$vmid')\n" if !defined($conf);
1947
1948     return $conf;
1949 }
1950
1951 sub parse_vm_config {
1952     my ($filename, $raw) = @_;
1953
1954     return undef if !defined($raw);
1955
1956     my $res = {
1957         digest => Digest::SHA::sha1_hex($raw),
1958         snapshots => {},
1959         pending => {},
1960     };
1961
1962     $filename =~ m|/qemu-server/(\d+)\.conf$|
1963         || die "got strange filename '$filename'";
1964
1965     my $vmid = $1;
1966
1967     my $conf = $res;
1968     my $descr = '';
1969     my $section = '';
1970
1971     my @lines = split(/\n/, $raw);
1972     foreach my $line (@lines) {
1973         next if $line =~ m/^\s*$/;
1974
1975         if ($line =~ m/^\[PENDING\]\s*$/i) {
1976             $section = 'pending';
1977             $conf->{description} = $descr if $descr;
1978             $descr = '';
1979             $conf = $res->{$section} = {};
1980             next;
1981
1982         } elsif ($line =~ m/^\[([a-z][a-z0-9_\-]+)\]\s*$/i) {
1983             $section = $1;
1984             $conf->{description} = $descr if $descr;
1985             $descr = '';
1986             $conf = $res->{snapshots}->{$section} = {};
1987             next;
1988         }
1989
1990         if ($line =~ m/^\#(.*)\s*$/) {
1991             $descr .= PVE::Tools::decode_text($1) . "\n";
1992             next;
1993         }
1994
1995         if ($line =~ m/^(description):\s*(.*\S)\s*$/) {
1996             $descr .= PVE::Tools::decode_text($2);
1997         } elsif ($line =~ m/snapstate:\s*(prepare|delete)\s*$/) {
1998             $conf->{snapstate} = $1;
1999         } elsif ($line =~ m/^(args):\s*(.*\S)\s*$/) {
2000             my $key = $1;
2001             my $value = $2;
2002             $conf->{$key} = $value;
2003         } elsif ($line =~ m/^delete:\s*(.*\S)\s*$/) {
2004             my $value = $1;
2005             if ($section eq 'pending') {
2006                 $conf->{delete} = $value; # we parse this later
2007             } else {
2008                 warn "vm $vmid - propertry 'delete' is only allowed in [PENDING]\n";
2009             }
2010         } elsif ($line =~ m/^([a-z][a-z_]*\d*):\s*(\S+)\s*$/) {
2011             my $key = $1;
2012             my $value = $2;
2013             eval { $value = check_type($key, $value); };
2014             if ($@) {
2015                 warn "vm $vmid - unable to parse value of '$key' - $@";
2016             } else {
2017                 my $fmt = $confdesc->{$key}->{format};
2018                 if ($fmt && $fmt eq 'pve-qm-drive') {
2019                     my $v = parse_drive($key, $value);
2020                     if (my $volid = filename_to_volume_id($vmid, $v->{file}, $v->{media})) {
2021                         $v->{file} = $volid;
2022                         $value = print_drive($vmid, $v);
2023                     } else {
2024                         warn "vm $vmid - unable to parse value of '$key'\n";
2025                         next;
2026                     }
2027                 }
2028
2029                 if ($key eq 'cdrom') {
2030                     $conf->{ide2} = $value;
2031                 } else {
2032                     $conf->{$key} = $value;
2033                 }
2034             }
2035         }
2036     }
2037
2038     $conf->{description} = $descr if $descr;
2039
2040     delete $res->{snapstate}; # just to be sure
2041
2042     return $res;
2043 }
2044
2045 sub write_vm_config {
2046     my ($filename, $conf) = @_;
2047
2048     delete $conf->{snapstate}; # just to be sure
2049
2050     if ($conf->{cdrom}) {
2051         die "option ide2 conflicts with cdrom\n" if $conf->{ide2};
2052         $conf->{ide2} = $conf->{cdrom};
2053         delete $conf->{cdrom};
2054     }
2055
2056     # we do not use 'smp' any longer
2057     if ($conf->{sockets}) {
2058         delete $conf->{smp};
2059     } elsif ($conf->{smp}) {
2060         $conf->{sockets} = $conf->{smp};
2061         delete $conf->{cores};
2062         delete $conf->{smp};
2063     }
2064
2065     my $used_volids = {};
2066
2067     my $cleanup_config = sub {
2068         my ($cref, $pending, $snapname) = @_;
2069
2070         foreach my $key (keys %$cref) {
2071             next if $key eq 'digest' || $key eq 'description' || $key eq 'snapshots' ||
2072                 $key eq 'snapstate' || $key eq 'pending';
2073             my $value = $cref->{$key};
2074             if ($key eq 'delete') {
2075                 die "propertry 'delete' is only allowed in [PENDING]\n"
2076                     if !$pending;
2077                 # fixme: check syntax?
2078                 next;
2079             }
2080             eval { $value = check_type($key, $value); };
2081             die "unable to parse value of '$key' - $@" if $@;
2082
2083             $cref->{$key} = $value;
2084
2085             if (!$snapname && valid_drivename($key)) {
2086                 my $drive = parse_drive($key, $value);
2087                 $used_volids->{$drive->{file}} = 1 if $drive && $drive->{file};
2088             }
2089         }
2090     };
2091
2092     &$cleanup_config($conf);
2093
2094     &$cleanup_config($conf->{pending}, 1);
2095
2096     foreach my $snapname (keys %{$conf->{snapshots}}) {
2097         die "internal error" if $snapname eq 'pending';
2098         &$cleanup_config($conf->{snapshots}->{$snapname}, undef, $snapname);
2099     }
2100
2101     # remove 'unusedX' settings if we re-add a volume
2102     foreach my $key (keys %$conf) {
2103         my $value = $conf->{$key};
2104         if ($key =~ m/^unused/ && $used_volids->{$value}) {
2105             delete $conf->{$key};
2106         }
2107     }
2108
2109     my $generate_raw_config = sub {
2110         my ($conf) = @_;
2111
2112         my $raw = '';
2113
2114         # add description as comment to top of file
2115         my $descr = $conf->{description} || '';
2116         foreach my $cl (split(/\n/, $descr)) {
2117             $raw .= '#' .  PVE::Tools::encode_text($cl) . "\n";
2118         }
2119
2120         foreach my $key (sort keys %$conf) {
2121             next if $key eq 'digest' || $key eq 'description' || $key eq 'pending' || $key eq 'snapshots';
2122             $raw .= "$key: $conf->{$key}\n";
2123         }
2124         return $raw;
2125     };
2126
2127     my $raw = &$generate_raw_config($conf);
2128
2129     if (scalar(keys %{$conf->{pending}})){
2130         $raw .= "\n[PENDING]\n";
2131         $raw .= &$generate_raw_config($conf->{pending});
2132     }
2133
2134     foreach my $snapname (sort keys %{$conf->{snapshots}}) {
2135         $raw .= "\n[$snapname]\n";
2136         $raw .= &$generate_raw_config($conf->{snapshots}->{$snapname});
2137     }
2138
2139     return $raw;
2140 }
2141
2142 sub update_config_nolock {
2143     my ($vmid, $conf, $skiplock) = @_;
2144
2145     check_lock($conf) if !$skiplock;
2146
2147     my $cfspath = cfs_config_path($vmid);
2148
2149     PVE::Cluster::cfs_write_file($cfspath, $conf);
2150 }
2151
2152 sub update_config {
2153     my ($vmid, $conf, $skiplock) = @_;
2154
2155     lock_config($vmid, &update_config_nolock, $conf, $skiplock);
2156 }
2157
2158 sub load_defaults {
2159
2160     my $res = {};
2161
2162     # we use static defaults from our JSON schema configuration
2163     foreach my $key (keys %$confdesc) {
2164         if (defined(my $default = $confdesc->{$key}->{default})) {
2165             $res->{$key} = $default;
2166         }
2167     }
2168
2169     my $conf = PVE::Cluster::cfs_read_file('datacenter.cfg');
2170     $res->{keyboard} = $conf->{keyboard} if $conf->{keyboard};
2171
2172     return $res;
2173 }
2174
2175 sub config_list {
2176     my $vmlist = PVE::Cluster::get_vmlist();
2177     my $res = {};
2178     return $res if !$vmlist || !$vmlist->{ids};
2179     my $ids = $vmlist->{ids};
2180
2181     foreach my $vmid (keys %$ids) {
2182         my $d = $ids->{$vmid};
2183         next if !$d->{node} || $d->{node} ne $nodename;
2184         next if !$d->{type} || $d->{type} ne 'qemu';
2185         $res->{$vmid}->{exists} = 1;
2186     }
2187     return $res;
2188 }
2189
2190 # test if VM uses local resources (to prevent migration)
2191 sub check_local_resources {
2192     my ($conf, $noerr) = @_;
2193
2194     my $loc_res = 0;
2195
2196     $loc_res = 1 if $conf->{hostusb}; # old syntax
2197     $loc_res = 1 if $conf->{hostpci}; # old syntax
2198
2199     foreach my $k (keys %$conf) {
2200         next if $k =~ m/^usb/ && ($conf->{$k} eq 'spice');
2201         $loc_res = 1 if $k =~ m/^(usb|hostpci|serial|parallel)\d+$/;
2202     }
2203
2204     die "VM uses local resources\n" if $loc_res && !$noerr;
2205
2206     return $loc_res;
2207 }
2208
2209 # check if used storages are available on all nodes (use by migrate)
2210 sub check_storage_availability {
2211     my ($storecfg, $conf, $node) = @_;
2212
2213     foreach_drive($conf, sub {
2214         my ($ds, $drive) = @_;
2215
2216         my $volid = $drive->{file};
2217         return if !$volid;
2218
2219         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
2220         return if !$sid;
2221
2222         # check if storage is available on both nodes
2223         my $scfg = PVE::Storage::storage_check_node($storecfg, $sid);
2224         PVE::Storage::storage_check_node($storecfg, $sid, $node);
2225    });
2226 }
2227
2228 # list nodes where all VM images are available (used by has_feature API)
2229 sub shared_nodes {
2230     my ($conf, $storecfg) = @_;
2231
2232     my $nodelist = PVE::Cluster::get_nodelist();
2233     my $nodehash = { map { $_ => 1 } @$nodelist };
2234     my $nodename = PVE::INotify::nodename();
2235
2236     foreach_drive($conf, sub {
2237         my ($ds, $drive) = @_;
2238
2239         my $volid = $drive->{file};
2240         return if !$volid;
2241
2242         my ($storeid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
2243         if ($storeid) {
2244             my $scfg = PVE::Storage::storage_config($storecfg, $storeid);
2245             if ($scfg->{disable}) {
2246                 $nodehash = {};
2247             } elsif (my $avail = $scfg->{nodes}) {
2248                 foreach my $node (keys %$nodehash) {
2249                     delete $nodehash->{$node} if !$avail->{$node};
2250                 }
2251             } elsif (!$scfg->{shared}) {
2252                 foreach my $node (keys %$nodehash) {
2253                     delete $nodehash->{$node} if $node ne $nodename
2254                 }
2255             }
2256         }
2257     });
2258
2259     return $nodehash
2260 }
2261
2262 sub check_lock {
2263     my ($conf) = @_;
2264
2265     die "VM is locked ($conf->{lock})\n" if $conf->{lock};
2266 }
2267
2268 sub check_cmdline {
2269     my ($pidfile, $pid) = @_;
2270
2271     my $fh = IO::File->new("/proc/$pid/cmdline", "r");
2272     if (defined($fh)) {
2273         my $line = <$fh>;
2274         $fh->close;
2275         return undef if !$line;
2276         my @param = split(/\0/, $line);
2277
2278         my $cmd = $param[0];
2279         return if !$cmd || ($cmd !~ m|kvm$| && $cmd !~ m|qemu-system-x86_64$|);
2280
2281         for (my $i = 0; $i < scalar (@param); $i++) {
2282             my $p = $param[$i];
2283             next if !$p;
2284             if (($p eq '-pidfile') || ($p eq '--pidfile')) {
2285                 my $p = $param[$i+1];
2286                 return 1 if $p && ($p eq $pidfile);
2287                 return undef;
2288             }
2289         }
2290     }
2291     return undef;
2292 }
2293
2294 sub check_running {
2295     my ($vmid, $nocheck, $node) = @_;
2296
2297     my $filename = config_file($vmid, $node);
2298
2299     die "unable to find configuration file for VM $vmid - no such machine\n"
2300         if !$nocheck && ! -f $filename;
2301
2302     my $pidfile = pidfile_name($vmid);
2303
2304     if (my $fd = IO::File->new("<$pidfile")) {
2305         my $st = stat($fd);
2306         my $line = <$fd>;
2307         close($fd);
2308
2309         my $mtime = $st->mtime;
2310         if ($mtime > time()) {
2311             warn "file '$filename' modified in future\n";
2312         }
2313
2314         if ($line =~ m/^(\d+)$/) {
2315             my $pid = $1;
2316             if (check_cmdline($pidfile, $pid)) {
2317                 if (my $pinfo = PVE::ProcFSTools::check_process_running($pid)) {
2318                     return $pid;
2319                 }
2320             }
2321         }
2322     }
2323
2324     return undef;
2325 }
2326
2327 sub vzlist {
2328
2329     my $vzlist = config_list();
2330
2331     my $fd = IO::Dir->new($var_run_tmpdir) || return $vzlist;
2332
2333     while (defined(my $de = $fd->read)) {
2334         next if $de !~ m/^(\d+)\.pid$/;
2335         my $vmid = $1;
2336         next if !defined($vzlist->{$vmid});
2337         if (my $pid = check_running($vmid)) {
2338             $vzlist->{$vmid}->{pid} = $pid;
2339         }
2340     }
2341
2342     return $vzlist;
2343 }
2344
2345 sub disksize {
2346     my ($storecfg, $conf) = @_;
2347
2348     my $bootdisk = $conf->{bootdisk};
2349     return undef if !$bootdisk;
2350     return undef if !valid_drivename($bootdisk);
2351
2352     return undef if !$conf->{$bootdisk};
2353
2354     my $drive = parse_drive($bootdisk, $conf->{$bootdisk});
2355     return undef if !defined($drive);
2356
2357     return undef if drive_is_cdrom($drive);
2358
2359     my $volid = $drive->{file};
2360     return undef if !$volid;
2361
2362     return $drive->{size};
2363 }
2364
2365 my $last_proc_pid_stat;
2366
2367 # get VM status information
2368 # This must be fast and should not block ($full == false)
2369 # We only query KVM using QMP if $full == true (this can be slow)
2370 sub vmstatus {
2371     my ($opt_vmid, $full) = @_;
2372
2373     my $res = {};
2374
2375     my $storecfg = PVE::Storage::config();
2376
2377     my $list = vzlist();
2378     my ($uptime) = PVE::ProcFSTools::read_proc_uptime(1);
2379
2380     my $cpucount = $cpuinfo->{cpus} || 1;
2381
2382     foreach my $vmid (keys %$list) {
2383         next if $opt_vmid && ($vmid ne $opt_vmid);
2384
2385         my $cfspath = cfs_config_path($vmid);
2386         my $conf = PVE::Cluster::cfs_read_file($cfspath) || {};
2387
2388         my $d = {};
2389         $d->{pid} = $list->{$vmid}->{pid};
2390
2391         # fixme: better status?
2392         $d->{status} = $list->{$vmid}->{pid} ? 'running' : 'stopped';
2393
2394         my $size = disksize($storecfg, $conf);
2395         if (defined($size)) {
2396             $d->{disk} = 0; # no info available
2397             $d->{maxdisk} = $size;
2398         } else {
2399             $d->{disk} = 0;
2400             $d->{maxdisk} = 0;
2401         }
2402
2403         $d->{cpus} = ($conf->{sockets} || 1) * ($conf->{cores} || 1);
2404         $d->{cpus} = $cpucount if $d->{cpus} > $cpucount;
2405         $d->{cpus} = $conf->{vcpus} if $conf->{vcpus};
2406
2407         $d->{name} = $conf->{name} || "VM $vmid";
2408         $d->{maxmem} = $conf->{memory} ? $conf->{memory}*(1024*1024) : 0;
2409
2410         if ($conf->{balloon}) {
2411             $d->{balloon_min} = $conf->{balloon}*(1024*1024);
2412             $d->{shares} = defined($conf->{shares}) ? $conf->{shares} : 1000;
2413         }
2414
2415         $d->{uptime} = 0;
2416         $d->{cpu} = 0;
2417         $d->{mem} = 0;
2418
2419         $d->{netout} = 0;
2420         $d->{netin} = 0;
2421
2422         $d->{diskread} = 0;
2423         $d->{diskwrite} = 0;
2424
2425         $d->{template} = is_template($conf);
2426
2427         $res->{$vmid} = $d;
2428     }
2429
2430     my $netdev = PVE::ProcFSTools::read_proc_net_dev();
2431     foreach my $dev (keys %$netdev) {
2432         next if $dev !~ m/^tap([1-9]\d*)i/;
2433         my $vmid = $1;
2434         my $d = $res->{$vmid};
2435         next if !$d;
2436
2437         $d->{netout} += $netdev->{$dev}->{receive};
2438         $d->{netin} += $netdev->{$dev}->{transmit};
2439     }
2440
2441     my $ctime = gettimeofday;
2442
2443     foreach my $vmid (keys %$list) {
2444
2445         my $d = $res->{$vmid};
2446         my $pid = $d->{pid};
2447         next if !$pid;
2448
2449         my $pstat = PVE::ProcFSTools::read_proc_pid_stat($pid);
2450         next if !$pstat; # not running
2451
2452         my $used = $pstat->{utime} + $pstat->{stime};
2453
2454         $d->{uptime} = int(($uptime - $pstat->{starttime})/$cpuinfo->{user_hz});
2455
2456         if ($pstat->{vsize}) {
2457             $d->{mem} = int(($pstat->{rss}/$pstat->{vsize})*$d->{maxmem});
2458         }
2459
2460         my $old = $last_proc_pid_stat->{$pid};
2461         if (!$old) {
2462             $last_proc_pid_stat->{$pid} = {
2463                 time => $ctime,
2464                 used => $used,
2465                 cpu => 0,
2466             };
2467             next;
2468         }
2469
2470         my $dtime = ($ctime -  $old->{time}) * $cpucount * $cpuinfo->{user_hz};
2471
2472         if ($dtime > 1000) {
2473             my $dutime = $used -  $old->{used};
2474
2475             $d->{cpu} = (($dutime/$dtime)* $cpucount) / $d->{cpus};
2476             $last_proc_pid_stat->{$pid} = {
2477                 time => $ctime,
2478                 used => $used,
2479                 cpu => $d->{cpu},
2480             };
2481         } else {
2482             $d->{cpu} = $old->{cpu};
2483         }
2484     }
2485
2486     return $res if !$full;
2487
2488     my $qmpclient = PVE::QMPClient->new();
2489
2490     my $ballooncb = sub {
2491         my ($vmid, $resp) = @_;
2492
2493         my $info = $resp->{'return'};
2494         return if !$info->{max_mem};
2495
2496         my $d = $res->{$vmid};
2497
2498         # use memory assigned to VM
2499         $d->{maxmem} = $info->{max_mem};
2500         $d->{balloon} = $info->{actual};
2501
2502         if (defined($info->{total_mem}) && defined($info->{free_mem})) {
2503             $d->{mem} = $info->{total_mem} - $info->{free_mem};
2504             $d->{freemem} = $info->{free_mem};
2505         }
2506
2507     };
2508
2509     my $blockstatscb = sub {
2510         my ($vmid, $resp) = @_;
2511         my $data = $resp->{'return'} || [];
2512         my $totalrdbytes = 0;
2513         my $totalwrbytes = 0;
2514         for my $blockstat (@$data) {
2515             $totalrdbytes = $totalrdbytes + $blockstat->{stats}->{rd_bytes};
2516             $totalwrbytes = $totalwrbytes + $blockstat->{stats}->{wr_bytes};
2517         }
2518         $res->{$vmid}->{diskread} = $totalrdbytes;
2519         $res->{$vmid}->{diskwrite} = $totalwrbytes;
2520     };
2521
2522     my $statuscb = sub {
2523         my ($vmid, $resp) = @_;
2524
2525         $qmpclient->queue_cmd($vmid, $blockstatscb, 'query-blockstats');
2526         # this fails if ballon driver is not loaded, so this must be
2527         # the last commnand (following command are aborted if this fails).
2528         $qmpclient->queue_cmd($vmid, $ballooncb, 'query-balloon');
2529
2530         my $status = 'unknown';
2531         if (!defined($status = $resp->{'return'}->{status})) {
2532             warn "unable to get VM status\n";
2533             return;
2534         }
2535
2536         $res->{$vmid}->{qmpstatus} = $resp->{'return'}->{status};
2537     };
2538
2539     foreach my $vmid (keys %$list) {
2540         next if $opt_vmid && ($vmid ne $opt_vmid);
2541         next if !$res->{$vmid}->{pid}; # not running
2542         $qmpclient->queue_cmd($vmid, $statuscb, 'query-status');
2543     }
2544
2545     $qmpclient->queue_execute(undef, 1);
2546
2547     foreach my $vmid (keys %$list) {
2548         next if $opt_vmid && ($vmid ne $opt_vmid);
2549         $res->{$vmid}->{qmpstatus} = $res->{$vmid}->{status} if !$res->{$vmid}->{qmpstatus};
2550     }
2551
2552     return $res;
2553 }
2554
2555 sub foreach_dimm {
2556     my ($conf, $vmid, $memory, $sockets, $func) = @_;
2557
2558     my $dimm_id = 0;
2559     my $current_size = 1024;
2560     my $dimm_size = 512;
2561     return if $current_size == $memory;
2562
2563     for (my $j = 0; $j < 8; $j++) {
2564         for (my $i = 0; $i < 32; $i++) {
2565             my $name = "dimm${dimm_id}";
2566             $dimm_id++;
2567             my $numanode = $i % $sockets;
2568             $current_size += $dimm_size;
2569             &$func($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory);
2570             return  $current_size if $current_size >= $memory;
2571         }
2572         $dimm_size *= 2;
2573     }
2574 }
2575
2576 sub foreach_drive {
2577     my ($conf, $func) = @_;
2578
2579     foreach my $ds (keys %$conf) {
2580         next if !valid_drivename($ds);
2581
2582         my $drive = parse_drive($ds, $conf->{$ds});
2583         next if !$drive;
2584
2585         &$func($ds, $drive);
2586     }
2587 }
2588
2589 sub foreach_volid {
2590     my ($conf, $func) = @_;
2591
2592     my $volhash = {};
2593
2594     my $test_volid = sub {
2595         my ($volid, $is_cdrom) = @_;
2596
2597         return if !$volid;
2598
2599         $volhash->{$volid} = $is_cdrom || 0;
2600     };
2601
2602     foreach_drive($conf, sub {
2603         my ($ds, $drive) = @_;
2604         &$test_volid($drive->{file}, drive_is_cdrom($drive));
2605     });
2606
2607     foreach my $snapname (keys %{$conf->{snapshots}}) {
2608         my $snap = $conf->{snapshots}->{$snapname};
2609         &$test_volid($snap->{vmstate}, 0);
2610         foreach_drive($snap, sub {
2611             my ($ds, $drive) = @_;
2612             &$test_volid($drive->{file}, drive_is_cdrom($drive));
2613         });
2614     }
2615
2616     foreach my $volid (keys %$volhash) {
2617         &$func($volid, $volhash->{$volid});
2618     }
2619 }
2620
2621 sub vga_conf_has_spice {
2622     my ($vga) = @_;
2623
2624     return 0 if !$vga || $vga !~ m/^qxl([234])?$/;
2625
2626     return $1 || 1;
2627 }
2628
2629 sub config_to_command {
2630     my ($storecfg, $vmid, $conf, $defaults, $forcemachine) = @_;
2631
2632     my $cmd = [];
2633     my $globalFlags = [];
2634     my $machineFlags = [];
2635     my $rtcFlags = [];
2636     my $cpuFlags = [];
2637     my $devices = [];
2638     my $pciaddr = '';
2639     my $bridges = {};
2640     my $kvmver = kvm_user_version();
2641     my $vernum = 0; # unknown
2642     if ($kvmver =~ m/^(\d+)\.(\d+)$/) {
2643         $vernum = $1*1000000+$2*1000;
2644     } elsif ($kvmver =~ m/^(\d+)\.(\d+)\.(\d+)$/) {
2645         $vernum = $1*1000000+$2*1000+$3;
2646     }
2647
2648     die "detected old qemu-kvm binary ($kvmver)\n" if $vernum < 15000;
2649
2650     my $have_ovz = -f '/proc/vz/vestat';
2651
2652     my $q35 = machine_type_is_q35($conf);
2653     my $hotplug_features = parse_hotplug_features(defined($conf->{hotplug}) ? $conf->{hotplug} : '1');
2654     my $machine_type = $forcemachine || $conf->{machine};
2655
2656     push @$cmd, '/usr/bin/kvm';
2657
2658     push @$cmd, '-id', $vmid;
2659
2660     my $use_virtio = 0;
2661
2662     my $qmpsocket = qmp_socket($vmid);
2663     push @$cmd, '-chardev', "socket,id=qmp,path=$qmpsocket,server,nowait";
2664     push @$cmd, '-mon', "chardev=qmp,mode=control";
2665
2666     my $socket = vnc_socket($vmid);
2667     push @$cmd,  '-vnc', "unix:$socket,x509,password";
2668
2669     push @$cmd, '-pidfile' , pidfile_name($vmid);
2670
2671     push @$cmd, '-daemonize';
2672
2673     if ($conf->{smbios1}) {
2674         push @$cmd, '-smbios', "type=1,$conf->{smbios1}";
2675     }
2676
2677     push @$cmd, '-object', "iothread,id=iothread0" if $conf->{iothread};
2678
2679     if ($q35) {
2680         # the q35 chipset support native usb2, so we enable usb controller
2681         # by default for this machine type
2682         push @$devices, '-readconfig', '/usr/share/qemu-server/pve-q35.cfg';
2683     } else {
2684         $pciaddr = print_pci_addr("piix3", $bridges);
2685         push @$devices, '-device', "piix3-usb-uhci,id=uhci$pciaddr.0x2";
2686
2687         my $use_usb2 = 0;
2688         for (my $i = 0; $i < $MAX_USB_DEVICES; $i++)  {
2689             next if !$conf->{"usb$i"};
2690             $use_usb2 = 1;
2691         }
2692         # include usb device config
2693         push @$devices, '-readconfig', '/usr/share/qemu-server/pve-usb.cfg' if $use_usb2;
2694     }
2695
2696     my $vga = $conf->{vga};
2697
2698     my $qxlnum = vga_conf_has_spice($vga);
2699     $vga = 'qxl' if $qxlnum;
2700
2701     if (!$vga) {
2702         if ($conf->{ostype} && ($conf->{ostype} eq 'win8' ||
2703                                 $conf->{ostype} eq 'win7' ||
2704                                 $conf->{ostype} eq 'w2k8')) {
2705             $vga = 'std';
2706         } else {
2707             $vga = 'cirrus';
2708         }
2709     }
2710
2711     # enable absolute mouse coordinates (needed by vnc)
2712     my $tablet;
2713     if (defined($conf->{tablet})) {
2714         $tablet = $conf->{tablet};
2715     } else {
2716         $tablet = $defaults->{tablet};
2717         $tablet = 0 if $qxlnum; # disable for spice because it is not needed
2718         $tablet = 0 if $vga =~ m/^serial\d+$/; # disable if we use serial terminal (no vga card)
2719     }
2720
2721     push @$devices, '-device', print_tabletdevice_full($conf) if $tablet;
2722
2723     # host pci devices
2724     for (my $i = 0; $i < $MAX_HOSTPCI_DEVICES; $i++)  {
2725         my $d = parse_hostpci($conf->{"hostpci$i"});
2726         next if !$d;
2727
2728         my $pcie = $d->{pcie};
2729         if($pcie){
2730             die "q35 machine model is not enabled" if !$q35;
2731             $pciaddr = print_pcie_addr("hostpci$i");
2732         }else{
2733             $pciaddr = print_pci_addr("hostpci$i", $bridges);
2734         }
2735
2736         my $rombar = $d->{rombar} && $d->{rombar} eq 'off' ? ",rombar=0" : "";
2737         my $driver = $d->{driver} && $d->{driver} eq 'vfio' ? "vfio-pci" : "pci-assign";
2738         my $xvga = $d->{'x-vga'} && $d->{'x-vga'} eq 'on' ? ",x-vga=on" : "";
2739         if ($xvga && $xvga ne '') {
2740             push @$cpuFlags, 'kvm=off';
2741             $vga = 'none';
2742         }
2743         $driver = "vfio-pci" if $xvga ne '';
2744         my $pcidevices = $d->{pciid};
2745         my $multifunction = 1 if @$pcidevices > 1;
2746
2747         my $j=0;
2748         foreach my $pcidevice (@$pcidevices) {
2749
2750             my $id = "hostpci$i";
2751             $id .= ".$j" if $multifunction;
2752             my $addr = $pciaddr;
2753             $addr .= ".$j" if $multifunction;
2754             my $devicestr = "$driver,host=$pcidevice->{id}.$pcidevice->{function},id=$id$addr";
2755
2756             if($j == 0){
2757                 $devicestr .= "$rombar$xvga";
2758                 $devicestr .= ",multifunction=on" if $multifunction;
2759             }
2760
2761             push @$devices, '-device', $devicestr;
2762             $j++;
2763         }
2764     }
2765
2766     # usb devices
2767     for (my $i = 0; $i < $MAX_USB_DEVICES; $i++)  {
2768         my $d = parse_usb_device($conf->{"usb$i"});
2769         next if !$d;
2770         if ($d->{vendorid} && $d->{productid}) {
2771             push @$devices, '-device', "usb-host,vendorid=0x$d->{vendorid},productid=0x$d->{productid}";
2772         } elsif (defined($d->{hostbus}) && defined($d->{hostport})) {
2773             push @$devices, '-device', "usb-host,hostbus=$d->{hostbus},hostport=$d->{hostport}";
2774         } elsif ($d->{spice}) {
2775             # usb redir support for spice
2776             push @$devices, '-chardev', "spicevmc,id=usbredirchardev$i,name=usbredir";
2777             push @$devices, '-device', "usb-redir,chardev=usbredirchardev$i,id=usbredirdev$i,bus=ehci.0";
2778         }
2779     }
2780
2781     # serial devices
2782     for (my $i = 0; $i < $MAX_SERIAL_PORTS; $i++)  {
2783         if (my $path = $conf->{"serial$i"}) {
2784             if ($path eq 'socket') {
2785                 my $socket = "/var/run/qemu-server/${vmid}.serial$i";
2786                 push @$devices, '-chardev', "socket,id=serial$i,path=$socket,server,nowait";
2787                 push @$devices, '-device', "isa-serial,chardev=serial$i";
2788             } else {
2789                 die "no such serial device\n" if ! -c $path;
2790                 push @$devices, '-chardev', "tty,id=serial$i,path=$path";
2791                 push @$devices, '-device', "isa-serial,chardev=serial$i";
2792             }
2793         }
2794     }
2795
2796     # parallel devices
2797     for (my $i = 0; $i < $MAX_PARALLEL_PORTS; $i++)  {
2798         if (my $path = $conf->{"parallel$i"}) {
2799             die "no such parallel device\n" if ! -c $path;
2800             my $devtype = $path =~ m!^/dev/usb/lp! ? 'tty' : 'parport';
2801             push @$devices, '-chardev', "$devtype,id=parallel$i,path=$path";
2802             push @$devices, '-device', "isa-parallel,chardev=parallel$i";
2803         }
2804     }
2805
2806     my $vmname = $conf->{name} || "vm$vmid";
2807
2808     push @$cmd, '-name', $vmname;
2809
2810     my $sockets = 1;
2811     $sockets = $conf->{smp} if $conf->{smp}; # old style - no longer iused
2812     $sockets = $conf->{sockets} if  $conf->{sockets};
2813
2814     my $cores = $conf->{cores} || 1;
2815
2816     my $maxcpus = $sockets * $cores;
2817
2818     my $vcpus = $conf->{vcpus} ? $conf->{vcpus} : $maxcpus;
2819
2820     my $allowed_vcpus = $cpuinfo->{cpus};
2821
2822     die "MAX $maxcpus vcpus allowed per VM on this node\n"
2823         if ($allowed_vcpus < $maxcpus);
2824
2825     push @$cmd, '-smp', "$vcpus,sockets=$sockets,cores=$cores,maxcpus=$maxcpus";
2826
2827     push @$cmd, '-nodefaults';
2828
2829     my $bootorder = $conf->{boot} || $confdesc->{boot}->{default};
2830
2831     my $bootindex_hash = {};
2832     my $i = 1;
2833     foreach my $o (split(//, $bootorder)) {
2834         $bootindex_hash->{$o} = $i*100;
2835         $i++;
2836     }
2837
2838     push @$cmd, '-boot', "menu=on,strict=on,reboot-timeout=1000";
2839
2840     push @$cmd, '-no-acpi' if defined($conf->{acpi}) && $conf->{acpi} == 0;
2841
2842     push @$cmd, '-no-reboot' if  defined($conf->{reboot}) && $conf->{reboot} == 0;
2843
2844     push @$cmd, '-vga', $vga if $vga && $vga !~ m/^serial\d+$/; # for kvm 77 and later
2845
2846     # time drift fix
2847     my $tdf = defined($conf->{tdf}) ? $conf->{tdf} : $defaults->{tdf};
2848
2849     my $nokvm = defined($conf->{kvm}) && $conf->{kvm} == 0 ? 1 : 0;
2850     my $useLocaltime = $conf->{localtime};
2851
2852     if (my $ost = $conf->{ostype}) {
2853         # other, wxp, w2k, w2k3, w2k8, wvista, win7, win8, l24, l26, solaris
2854
2855         if ($ost =~ m/^w/) { # windows
2856             $useLocaltime = 1 if !defined($conf->{localtime});
2857
2858             # use time drift fix when acpi is enabled
2859             if (!(defined($conf->{acpi}) && $conf->{acpi} == 0)) {
2860                 $tdf = 1 if !defined($conf->{tdf});
2861             }
2862         }
2863
2864         if ($ost eq 'win7' || $ost eq 'win8' || $ost eq 'w2k8' ||
2865             $ost eq 'wvista') {
2866             push @$globalFlags, 'kvm-pit.lost_tick_policy=discard';
2867             push @$cmd, '-no-hpet';
2868             if (qemu_machine_feature_enabled ($machine_type, $kvmver, 2, 3)) {
2869                 push @$cpuFlags , 'hv_spinlocks=0x1fff' if !$nokvm;
2870                 push @$cpuFlags , 'hv_vapic' if !$nokvm;
2871                 push @$cpuFlags , 'hv_time' if !$nokvm;
2872
2873             } else {
2874                 push @$cpuFlags , 'hv_spinlocks=0xffff' if !$nokvm;
2875             }
2876         }
2877
2878         if ($ost eq 'win7' || $ost eq 'win8') {
2879             push @$cpuFlags , 'hv_relaxed' if !$nokvm;
2880         }
2881     }
2882
2883     push @$rtcFlags, 'driftfix=slew' if $tdf;
2884
2885     if ($nokvm) {
2886         push @$machineFlags, 'accel=tcg';
2887     } else {
2888         die "No accelerator found!\n" if !$cpuinfo->{hvm};
2889     }
2890
2891     if ($machine_type) {
2892         push @$machineFlags, "type=${machine_type}";
2893     }
2894
2895     if ($conf->{startdate}) {
2896         push @$rtcFlags, "base=$conf->{startdate}";
2897     } elsif ($useLocaltime) {
2898         push @$rtcFlags, 'base=localtime';
2899     }
2900
2901     my $cpu = $nokvm ? "qemu64" : "kvm64";
2902     $cpu = $conf->{cpu} if $conf->{cpu};
2903
2904     push @$cpuFlags , '+lahf_lm' if $cpu eq 'kvm64';
2905
2906     push @$cpuFlags , '+x2apic' if !$nokvm && $conf->{ostype} ne 'solaris';
2907
2908     push @$cpuFlags , '-x2apic' if $conf->{ostype} eq 'solaris';
2909
2910     push @$cpuFlags, '+sep' if $cpu eq 'kvm64' || $cpu eq 'kvm32';
2911
2912     if (qemu_machine_feature_enabled ($machine_type, $kvmver, 2, 3)) {
2913
2914         push @$cpuFlags , '+kvm_pv_unhalt' if !$nokvm;
2915     }
2916
2917     $cpu .= "," . join(',', @$cpuFlags) if scalar(@$cpuFlags);
2918
2919     # Note: enforce needs kernel 3.10, so we do not use it for now
2920     # push @$cmd, '-cpu', "$cpu,enforce";
2921     push @$cmd, '-cpu', $cpu;
2922
2923     my $memory = $conf->{memory} || $defaults->{memory};
2924     my $static_memory = 0;
2925     my $dimm_memory = 0;
2926
2927     if ($hotplug_features->{memory}) {
2928         die "Numa need to be enabled for memory hotplug\n" if !$conf->{numa};
2929         die "Total memory is bigger than ${MAX_MEM}MB\n" if $memory > $MAX_MEM;
2930         $static_memory = $STATICMEM;
2931         die "minimum memory must be ${static_memory}MB\n" if($memory < $static_memory);
2932         $dimm_memory = $memory - $static_memory;
2933         push @$cmd, '-m', "size=${static_memory},slots=255,maxmem=${MAX_MEM}M";
2934
2935     } else {
2936
2937         $static_memory = $memory;
2938         push @$cmd, '-m', $static_memory;
2939     }
2940
2941     if ($conf->{numa}) {
2942
2943         my $numa_totalmemory = undef;
2944         for (my $i = 0; $i < $MAX_NUMA; $i++) {
2945             next if !$conf->{"numa$i"};
2946             my $numa = parse_numa($conf->{"numa$i"});
2947             next if !$numa;
2948             # memory
2949             die "missing numa node$i memory value\n" if !$numa->{memory};
2950             my $numa_memory = $numa->{memory};
2951             $numa_totalmemory += $numa_memory;
2952             my $numa_object = "memory-backend-ram,id=ram-node$i,size=${numa_memory}M";
2953
2954             # cpus
2955             my $cpus_start = $numa->{cpus}->{start};
2956             die "missing numa node$i cpus\n" if !defined($cpus_start);
2957             my $cpus_end = $numa->{cpus}->{end} if defined($numa->{cpus}->{end});
2958             my $cpus = $cpus_start;
2959             if (defined($cpus_end)) {
2960                 $cpus .= "-$cpus_end";
2961                 die "numa node$i :  cpu range $cpus is incorrect\n" if $cpus_end <= $cpus_start;
2962             }
2963
2964             # hostnodes
2965             my $hostnodes_start = $numa->{hostnodes}->{start};
2966             if (defined($hostnodes_start)) {
2967                 my $hostnodes_end = $numa->{hostnodes}->{end} if defined($numa->{hostnodes}->{end});
2968                 my $hostnodes = $hostnodes_start;
2969                 if (defined($hostnodes_end)) {
2970                     $hostnodes .= "-$hostnodes_end";
2971                     die "host node $hostnodes range is incorrect\n" if $hostnodes_end <= $hostnodes_start;
2972                 }
2973
2974                 my $hostnodes_end_range = defined($hostnodes_end) ? $hostnodes_end : $hostnodes_start;
2975                 for (my $i = $hostnodes_start; $i <= $hostnodes_end_range; $i++ ) {
2976                     die "host numa node$i don't exist\n" if ! -d "/sys/devices/system/node/node$i/";
2977                 }
2978
2979                 # policy
2980                 my $policy = $numa->{policy};
2981                 die "you need to define a policy for hostnode $hostnodes\n" if !$policy;
2982                 $numa_object .= ",host-nodes=$hostnodes,policy=$policy";
2983             }
2984
2985             push @$cmd, '-object', $numa_object;
2986             push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i";
2987         }
2988
2989         die "total memory for NUMA nodes must be equal to vm static memory\n"
2990             if $numa_totalmemory && $numa_totalmemory != $static_memory;
2991
2992         #if no custom tology, we split memory and cores across numa nodes
2993         if(!$numa_totalmemory) {
2994
2995             my $numa_memory = ($static_memory / $sockets) . "M";
2996
2997             for (my $i = 0; $i < $sockets; $i++)  {
2998
2999                 my $cpustart = ($cores * $i);
3000                 my $cpuend = ($cpustart + $cores - 1) if $cores && $cores > 1;
3001                 my $cpus = $cpustart;
3002                 $cpus .= "-$cpuend" if $cpuend;
3003
3004                 push @$cmd, '-object', "memory-backend-ram,size=$numa_memory,id=ram-node$i";
3005                 push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i";
3006             }
3007         }
3008     }
3009
3010     if ($hotplug_features->{memory}) {
3011         foreach_dimm($conf, $vmid, $memory, $sockets, sub {
3012             my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
3013             push @$cmd, "-object" , "memory-backend-ram,id=mem-$name,size=${dimm_size}M";
3014             push @$cmd, "-device", "pc-dimm,id=$name,memdev=mem-$name,node=$numanode";
3015
3016             #if dimm_memory is not aligned to dimm map
3017             if($current_size > $memory) {
3018                  $conf->{memory} = $current_size;
3019                  update_config_nolock($vmid, $conf, 1);
3020             }
3021         });
3022     }
3023
3024     push @$cmd, '-S' if $conf->{freeze};
3025
3026     # set keyboard layout
3027     my $kb = $conf->{keyboard} || $defaults->{keyboard};
3028     push @$cmd, '-k', $kb if $kb;
3029
3030     # enable sound
3031     #my $soundhw = $conf->{soundhw} || $defaults->{soundhw};
3032     #push @$cmd, '-soundhw', 'es1370';
3033     #push @$cmd, '-soundhw', $soundhw if $soundhw;
3034
3035     if($conf->{agent}) {
3036         my $qgasocket = qmp_socket($vmid, 1);
3037         my $pciaddr = print_pci_addr("qga0", $bridges);
3038         push @$devices, '-chardev', "socket,path=$qgasocket,server,nowait,id=qga0";
3039         push @$devices, '-device', "virtio-serial,id=qga0$pciaddr";
3040         push @$devices, '-device', 'virtserialport,chardev=qga0,name=org.qemu.guest_agent.0';
3041     }
3042
3043     my $spice_port;
3044
3045     if ($qxlnum) {
3046         if ($qxlnum > 1) {
3047             if ($conf->{ostype} && $conf->{ostype} =~ m/^w/){
3048                 for(my $i = 1; $i < $qxlnum; $i++){
3049                     my $pciaddr = print_pci_addr("vga$i", $bridges);
3050                     push @$cmd, '-device', "qxl,id=vga$i,ram_size=67108864,vram_size=33554432$pciaddr";
3051                 }
3052             } else {
3053                 # assume other OS works like Linux
3054                 push @$cmd, '-global', 'qxl-vga.ram_size=134217728';
3055                 push @$cmd, '-global', 'qxl-vga.vram_size=67108864';
3056             }
3057         }
3058
3059         my $pciaddr = print_pci_addr("spice", $bridges);
3060
3061         $spice_port = PVE::Tools::next_spice_port();
3062
3063         push @$devices, '-spice', "tls-port=${spice_port},addr=127.0.0.1,tls-ciphers=DES-CBC3-SHA,seamless-migration=on";
3064
3065         push @$devices, '-device', "virtio-serial,id=spice$pciaddr";
3066         push @$devices, '-chardev', "spicevmc,id=vdagent,name=vdagent";
3067         push @$devices, '-device', "virtserialport,chardev=vdagent,name=com.redhat.spice.0";
3068     }
3069
3070     # enable balloon by default, unless explicitly disabled
3071     if (!defined($conf->{balloon}) || $conf->{balloon}) {
3072         $pciaddr = print_pci_addr("balloon0", $bridges);
3073         push @$devices, '-device', "virtio-balloon-pci,id=balloon0$pciaddr";
3074     }
3075
3076     if ($conf->{watchdog}) {
3077         my $wdopts = parse_watchdog($conf->{watchdog});
3078         $pciaddr = print_pci_addr("watchdog", $bridges);
3079         my $watchdog = $wdopts->{model} || 'i6300esb';
3080         push @$devices, '-device', "$watchdog$pciaddr";
3081         push @$devices, '-watchdog-action', $wdopts->{action} if $wdopts->{action};
3082     }
3083
3084     my $vollist = [];
3085     my $scsicontroller = {};
3086     my $ahcicontroller = {};
3087     my $scsihw = defined($conf->{scsihw}) ? $conf->{scsihw} : $defaults->{scsihw};
3088
3089     # Add iscsi initiator name if available
3090     if (my $initiator = get_initiator_name()) {
3091         push @$devices, '-iscsi', "initiator-name=$initiator";
3092     }
3093
3094     foreach_drive($conf, sub {
3095         my ($ds, $drive) = @_;
3096
3097         if (PVE::Storage::parse_volume_id($drive->{file}, 1)) {
3098             push @$vollist, $drive->{file};
3099         }
3100
3101         $use_virtio = 1 if $ds =~ m/^virtio/;
3102
3103         if (drive_is_cdrom ($drive)) {
3104             if ($bootindex_hash->{d}) {
3105                 $drive->{bootindex} = $bootindex_hash->{d};
3106                 $bootindex_hash->{d} += 1;
3107             }
3108         } else {
3109             if ($bootindex_hash->{c}) {
3110                 $drive->{bootindex} = $bootindex_hash->{c} if $conf->{bootdisk} && ($conf->{bootdisk} eq $ds);
3111                 $bootindex_hash->{c} += 1;
3112             }
3113         }
3114
3115         if ($drive->{interface} eq 'scsi') {
3116
3117             my $maxdev = ($scsihw !~ m/^lsi/) ? 256 : 7;
3118             my $controller = int($drive->{index} / $maxdev);
3119             $pciaddr = print_pci_addr("scsihw$controller", $bridges);
3120             push @$devices, '-device', "$scsihw,id=scsihw$controller$pciaddr" if !$scsicontroller->{$controller};
3121             $scsicontroller->{$controller}=1;
3122         }
3123
3124         if ($drive->{interface} eq 'sata') {
3125            my $controller = int($drive->{index} / $MAX_SATA_DISKS);
3126            $pciaddr = print_pci_addr("ahci$controller", $bridges);
3127            push @$devices, '-device', "ahci,id=ahci$controller,multifunction=on$pciaddr" if !$ahcicontroller->{$controller};
3128            $ahcicontroller->{$controller}=1;
3129         }
3130
3131         my $drive_cmd = print_drive_full($storecfg, $vmid, $drive);
3132         push @$devices, '-drive',$drive_cmd;
3133         push @$devices, '-device', print_drivedevice_full($storecfg, $conf, $vmid, $drive, $bridges);
3134     });
3135
3136     for (my $i = 0; $i < $MAX_NETS; $i++) {
3137          next if !$conf->{"net$i"};
3138          my $d = parse_net($conf->{"net$i"});
3139          next if !$d;
3140
3141          $use_virtio = 1 if $d->{model} eq 'virtio';
3142
3143          if ($bootindex_hash->{n}) {
3144             $d->{bootindex} = $bootindex_hash->{n};
3145             $bootindex_hash->{n} += 1;
3146          }
3147
3148          my $netdevfull = print_netdev_full($vmid,$conf,$d,"net$i");
3149          push @$devices, '-netdev', $netdevfull;
3150
3151          my $netdevicefull = print_netdevice_full($vmid,$conf,$d,"net$i",$bridges);
3152          push @$devices, '-device', $netdevicefull;
3153     }
3154
3155     if (!$q35) {
3156         # add pci bridges
3157         while (my ($k, $v) = each %$bridges) {
3158             $pciaddr = print_pci_addr("pci.$k");
3159             unshift @$devices, '-device', "pci-bridge,id=pci.$k,chassis_nr=$k$pciaddr" if $k > 0;
3160         }
3161     }
3162
3163     # hack: virtio with fairsched is unreliable, so we do not use fairsched
3164     # when the VM uses virtio devices.
3165     if (!$use_virtio && $have_ovz) {
3166
3167         my $cpuunits = defined($conf->{cpuunits}) ?
3168             $conf->{cpuunits} : $defaults->{cpuunits};
3169
3170         push @$cmd, '-cpuunits', $cpuunits if $cpuunits;
3171
3172         # fixme: cpulimit is currently ignored
3173         #push @$cmd, '-cpulimit', $conf->{cpulimit} if $conf->{cpulimit};
3174     }
3175
3176     # add custom args
3177     if ($conf->{args}) {
3178         my $aa = PVE::Tools::split_args($conf->{args});
3179         push @$cmd, @$aa;
3180     }
3181
3182     push @$cmd, @$devices;
3183     push @$cmd, '-rtc', join(',', @$rtcFlags)
3184         if scalar(@$rtcFlags);
3185     push @$cmd, '-machine', join(',', @$machineFlags)
3186         if scalar(@$machineFlags);
3187     push @$cmd, '-global', join(',', @$globalFlags)
3188         if scalar(@$globalFlags);
3189
3190     return wantarray ? ($cmd, $vollist, $spice_port) : $cmd;
3191 }
3192
3193 sub vnc_socket {
3194     my ($vmid) = @_;
3195     return "${var_run_tmpdir}/$vmid.vnc";
3196 }
3197
3198 sub spice_port {
3199     my ($vmid) = @_;
3200
3201     my $res = vm_mon_cmd($vmid, 'query-spice');
3202
3203     return $res->{'tls-port'} || $res->{'port'} || die "no spice port\n";
3204 }
3205
3206 sub qmp_socket {
3207     my ($vmid, $qga) = @_;
3208     my $sockettype = $qga ? 'qga' : 'qmp';
3209     return "${var_run_tmpdir}/$vmid.$sockettype";
3210 }
3211
3212 sub pidfile_name {
3213     my ($vmid) = @_;
3214     return "${var_run_tmpdir}/$vmid.pid";
3215 }
3216
3217 sub vm_devices_list {
3218     my ($vmid) = @_;
3219
3220     my $res = vm_mon_cmd($vmid, 'query-pci');
3221     my $devices = {};
3222     foreach my $pcibus (@$res) {
3223         foreach my $device (@{$pcibus->{devices}}) {
3224             next if !$device->{'qdev_id'};
3225             if ($device->{'pci_bridge'}) {
3226                 $devices->{$device->{'qdev_id'}} = 1;
3227                 foreach my $bridge_device (@{$device->{'pci_bridge'}->{devices}}) {
3228                     next if !$bridge_device->{'qdev_id'};
3229                     $devices->{$bridge_device->{'qdev_id'}} = 1;
3230                     $devices->{$device->{'qdev_id'}}++;
3231                 }
3232             } else {
3233                 $devices->{$device->{'qdev_id'}} = 1;
3234             }
3235         }
3236     }
3237
3238     my $resblock = vm_mon_cmd($vmid, 'query-block');
3239     foreach my $block (@$resblock) {
3240         if($block->{device} =~ m/^drive-(\S+)/){
3241                 $devices->{$1} = 1;
3242         }
3243     }
3244
3245     my $resmice = vm_mon_cmd($vmid, 'query-mice');
3246     foreach my $mice (@$resmice) {
3247         if ($mice->{name} eq 'QEMU HID Tablet') {
3248             $devices->{tablet} = 1;
3249             last;
3250         }
3251     }
3252
3253     return $devices;
3254 }
3255
3256 sub vm_deviceplug {
3257     my ($storecfg, $conf, $vmid, $deviceid, $device) = @_;
3258
3259     my $q35 = machine_type_is_q35($conf);
3260
3261     my $devices_list = vm_devices_list($vmid);
3262     return 1 if defined($devices_list->{$deviceid});
3263
3264     qemu_add_pci_bridge($storecfg, $conf, $vmid, $deviceid); # add PCI bridge if we need it for the device
3265
3266     if ($deviceid eq 'tablet') {
3267
3268         qemu_deviceadd($vmid, print_tabletdevice_full($conf));
3269
3270     } elsif ($deviceid =~ m/^(virtio)(\d+)$/) {
3271
3272         qemu_driveadd($storecfg, $vmid, $device);
3273         my $devicefull = print_drivedevice_full($storecfg, $conf, $vmid, $device);
3274
3275         qemu_deviceadd($vmid, $devicefull);
3276         eval { qemu_deviceaddverify($vmid, $deviceid); };
3277         if (my $err = $@) {
3278             eval { qemu_drivedel($vmid, $deviceid); };
3279             warn $@ if $@;
3280             die $err;
3281         }
3282
3283     } elsif ($deviceid =~ m/^(scsihw)(\d+)$/) {
3284
3285         my $scsihw = defined($conf->{scsihw}) ? $conf->{scsihw} : "lsi";
3286         my $pciaddr = print_pci_addr($deviceid);
3287         my $devicefull = "$scsihw,id=$deviceid$pciaddr";
3288
3289         qemu_deviceadd($vmid, $devicefull);
3290         qemu_deviceaddverify($vmid, $deviceid);
3291
3292     } elsif ($deviceid =~ m/^(scsi)(\d+)$/) {
3293
3294         qemu_findorcreatescsihw($storecfg,$conf, $vmid, $device);
3295         qemu_driveadd($storecfg, $vmid, $device);
3296
3297         my $devicefull = print_drivedevice_full($storecfg, $conf, $vmid, $device);
3298         eval { qemu_deviceadd($vmid, $devicefull); };
3299         if (my $err = $@) {
3300             eval { qemu_drivedel($vmid, $deviceid); };
3301             warn $@ if $@;
3302             die $err;
3303         }
3304
3305     } elsif ($deviceid =~ m/^(net)(\d+)$/) {
3306
3307         return undef if !qemu_netdevadd($vmid, $conf, $device, $deviceid);
3308         my $netdevicefull = print_netdevice_full($vmid, $conf, $device, $deviceid);
3309         qemu_deviceadd($vmid, $netdevicefull);
3310         eval { qemu_deviceaddverify($vmid, $deviceid); };
3311         if (my $err = $@) {
3312             eval { qemu_netdevdel($vmid, $deviceid); };
3313             warn $@ if $@;
3314             die $err;
3315         }
3316
3317     } elsif (!$q35 && $deviceid =~ m/^(pci\.)(\d+)$/) {
3318
3319         my $bridgeid = $2;
3320         my $pciaddr = print_pci_addr($deviceid);
3321         my $devicefull = "pci-bridge,id=pci.$bridgeid,chassis_nr=$bridgeid$pciaddr";
3322
3323         qemu_deviceadd($vmid, $devicefull);
3324         qemu_deviceaddverify($vmid, $deviceid);
3325
3326     } else {
3327         die "can't hotplug device '$deviceid'\n";
3328     }
3329
3330     return 1;
3331 }
3332
3333 # fixme: this should raise exceptions on error!
3334 sub vm_deviceunplug {
3335     my ($vmid, $conf, $deviceid) = @_;
3336
3337     my $devices_list = vm_devices_list($vmid);
3338     return 1 if !defined($devices_list->{$deviceid});
3339
3340     die "can't unplug bootdisk" if $conf->{bootdisk} && $conf->{bootdisk} eq $deviceid;
3341
3342     if ($deviceid eq 'tablet') {
3343
3344         qemu_devicedel($vmid, $deviceid);
3345
3346     } elsif ($deviceid =~ m/^(virtio)(\d+)$/) {
3347
3348         qemu_devicedel($vmid, $deviceid);
3349         qemu_devicedelverify($vmid, $deviceid);
3350         qemu_drivedel($vmid, $deviceid);
3351
3352     } elsif ($deviceid =~ m/^(lsi)(\d+)$/) {
3353
3354         qemu_devicedel($vmid, $deviceid);
3355
3356     } elsif ($deviceid =~ m/^(scsi)(\d+)$/) {
3357
3358         qemu_devicedel($vmid, $deviceid);
3359         qemu_drivedel($vmid, $deviceid);
3360
3361     } elsif ($deviceid =~ m/^(net)(\d+)$/) {
3362
3363         qemu_devicedel($vmid, $deviceid);
3364         qemu_devicedelverify($vmid, $deviceid);
3365         qemu_netdevdel($vmid, $deviceid);
3366
3367     } else {
3368         die "can't unplug device '$deviceid'\n";
3369     }
3370
3371     return 1;
3372 }
3373
3374 sub qemu_deviceadd {
3375     my ($vmid, $devicefull) = @_;
3376
3377     $devicefull = "driver=".$devicefull;
3378     my %options =  split(/[=,]/, $devicefull);
3379
3380     vm_mon_cmd($vmid, "device_add" , %options);
3381 }
3382
3383 sub qemu_devicedel {
3384     my ($vmid, $deviceid) = @_;
3385
3386     my $ret = vm_mon_cmd($vmid, "device_del", id => $deviceid);
3387 }
3388
3389 sub qemu_objectadd {
3390     my($vmid, $objectid, $qomtype) = @_;
3391
3392     vm_mon_cmd($vmid, "object-add", id => $objectid, "qom-type" => $qomtype);
3393
3394     return 1;
3395 }
3396
3397 sub qemu_objectdel {
3398     my($vmid, $objectid) = @_;
3399
3400     vm_mon_cmd($vmid, "object-del", id => $objectid);
3401
3402     return 1;
3403 }
3404
3405 sub qemu_driveadd {
3406     my ($storecfg, $vmid, $device) = @_;
3407
3408     my $drive = print_drive_full($storecfg, $vmid, $device);
3409     $drive =~ s/\\/\\\\/g;
3410     my $ret = vm_human_monitor_command($vmid, "drive_add auto \"$drive\"");
3411
3412     # If the command succeeds qemu prints: "OK"
3413     return 1 if $ret =~ m/OK/s;
3414
3415     die "adding drive failed: $ret\n";
3416 }
3417
3418 sub qemu_drivedel {
3419     my($vmid, $deviceid) = @_;
3420
3421     my $ret = vm_human_monitor_command($vmid, "drive_del drive-$deviceid");
3422     $ret =~ s/^\s+//;
3423
3424     return 1 if $ret eq "";
3425
3426     # NB: device not found errors mean the drive was auto-deleted and we ignore the error
3427     return 1 if $ret =~ m/Device \'.*?\' not found/s;
3428
3429     die "deleting drive $deviceid failed : $ret\n";
3430 }
3431
3432 sub qemu_deviceaddverify {
3433     my ($vmid, $deviceid) = @_;
3434
3435     for (my $i = 0; $i <= 5; $i++) {
3436          my $devices_list = vm_devices_list($vmid);
3437          return 1 if defined($devices_list->{$deviceid});
3438          sleep 1;
3439     }
3440
3441     die "error on hotplug device '$deviceid'\n";
3442 }
3443
3444
3445 sub qemu_devicedelverify {
3446     my ($vmid, $deviceid) = @_;
3447
3448     # need to verify that the device is correctly removed as device_del
3449     # is async and empty return is not reliable
3450
3451     for (my $i = 0; $i <= 5; $i++) {
3452          my $devices_list = vm_devices_list($vmid);
3453          return 1 if !defined($devices_list->{$deviceid});
3454          sleep 1;
3455     }
3456
3457     die "error on hot-unplugging device '$deviceid'\n";
3458 }
3459
3460 sub qemu_findorcreatescsihw {
3461     my ($storecfg, $conf, $vmid, $device) = @_;
3462
3463     my $maxdev = ($conf->{scsihw} && ($conf->{scsihw} !~ m/^lsi/)) ? 256 : 7;
3464     my $controller = int($device->{index} / $maxdev);
3465     my $scsihwid="scsihw$controller";
3466     my $devices_list = vm_devices_list($vmid);
3467
3468     if(!defined($devices_list->{$scsihwid})) {
3469         vm_deviceplug($storecfg, $conf, $vmid, $scsihwid);
3470     }
3471
3472     return 1;
3473 }
3474
3475 sub qemu_add_pci_bridge {
3476     my ($storecfg, $conf, $vmid, $device) = @_;
3477
3478     my $bridges = {};
3479
3480     my $bridgeid;
3481
3482     print_pci_addr($device, $bridges);
3483
3484     while (my ($k, $v) = each %$bridges) {
3485         $bridgeid = $k;
3486     }
3487     return 1 if !defined($bridgeid) || $bridgeid < 1;
3488
3489     my $bridge = "pci.$bridgeid";
3490     my $devices_list = vm_devices_list($vmid);
3491
3492     if (!defined($devices_list->{$bridge})) {
3493         vm_deviceplug($storecfg, $conf, $vmid, $bridge);
3494     }
3495
3496     return 1;
3497 }
3498
3499 sub qemu_set_link_status {
3500     my ($vmid, $device, $up) = @_;
3501
3502     vm_mon_cmd($vmid, "set_link", name => $device,
3503                up => $up ? JSON::true : JSON::false);
3504 }
3505
3506 sub qemu_netdevadd {
3507     my ($vmid, $conf, $device, $deviceid) = @_;
3508
3509     my $netdev = print_netdev_full($vmid, $conf, $device, $deviceid);
3510     my %options =  split(/[=,]/, $netdev);
3511
3512     vm_mon_cmd($vmid, "netdev_add",  %options);
3513     return 1;
3514 }
3515
3516 sub qemu_netdevdel {
3517     my ($vmid, $deviceid) = @_;
3518
3519     vm_mon_cmd($vmid, "netdev_del", id => $deviceid);
3520 }
3521
3522 sub qemu_cpu_hotplug {
3523     my ($vmid, $conf, $vcpus) = @_;
3524
3525     my $sockets = 1;
3526     $sockets = $conf->{smp} if $conf->{smp}; # old style - no longer iused
3527     $sockets = $conf->{sockets} if  $conf->{sockets};
3528     my $cores = $conf->{cores} || 1;
3529     my $maxcpus = $sockets * $cores;
3530
3531     $vcpus = $maxcpus if !$vcpus;
3532
3533     die "you can't add more vcpus than maxcpus\n"
3534         if $vcpus > $maxcpus;
3535
3536     my $currentvcpus = $conf->{vcpus} || $maxcpus;
3537     die "online cpu unplug is not yet possible\n"
3538         if $vcpus < $currentvcpus;
3539
3540     my $currentrunningvcpus = vm_mon_cmd($vmid, "query-cpus");
3541     die "vcpus in running vm is different than configuration\n"
3542         if scalar(@{$currentrunningvcpus}) != $currentvcpus;
3543
3544     for (my $i = $currentvcpus; $i < $vcpus; $i++) {
3545         vm_mon_cmd($vmid, "cpu-add", id => int($i));
3546     }
3547 }
3548
3549 sub qemu_memory_hotplug {
3550     my ($vmid, $conf, $defaults, $opt, $value) = @_;
3551
3552     return $value if !check_running($vmid);
3553
3554     my $memory = $conf->{memory} || $defaults->{memory};
3555     $value = $defaults->{memory} if !$value;
3556     return $value if $value == $memory;
3557
3558     my $static_memory = $STATICMEM;
3559     my $dimm_memory = $memory - $static_memory;
3560
3561     die "memory can't be lower than $static_memory MB" if $value < $static_memory;
3562     die "memory unplug is not yet available" if $value < $memory;
3563     die "you cannot add more memory than $MAX_MEM MB!\n" if $memory > $MAX_MEM;
3564
3565
3566     my $sockets = 1;
3567     $sockets = $conf->{sockets} if $conf->{sockets};
3568
3569     foreach_dimm($conf, $vmid, $value, $sockets, sub {
3570         my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
3571
3572             return if $current_size <= $conf->{memory};
3573
3574             eval { vm_mon_cmd($vmid, "object-add", 'qom-type' => "memory-backend-ram", id => "mem-$name", props => { size => int($dimm_size*1024*1024) } ) };
3575             if (my $err = $@) {
3576                 eval { qemu_objectdel($vmid, "mem-$name"); };
3577                 die $err;
3578             }
3579
3580             eval { vm_mon_cmd($vmid, "device_add", driver => "pc-dimm", id => "$name", memdev => "mem-$name", node => $numanode) };
3581             if (my $err = $@) {
3582                 eval { qemu_objectdel($vmid, "mem-$name"); };
3583                 die $err;
3584             }
3585             #update conf after each succesful module hotplug
3586             $conf->{memory} = $current_size;
3587             update_config_nolock($vmid, $conf, 1);
3588     });
3589 }
3590
3591 sub qemu_block_set_io_throttle {
3592     my ($vmid, $deviceid, $bps, $bps_rd, $bps_wr, $iops, $iops_rd, $iops_wr) = @_;
3593
3594     return if !check_running($vmid) ;
3595
3596     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($iops_wr));
3597
3598 }
3599
3600 # old code, only used to shutdown old VM after update
3601 sub __read_avail {
3602     my ($fh, $timeout) = @_;
3603
3604     my $sel = new IO::Select;
3605     $sel->add($fh);
3606
3607     my $res = '';
3608     my $buf;
3609
3610     my @ready;
3611     while (scalar (@ready = $sel->can_read($timeout))) {
3612         my $count;
3613         if ($count = $fh->sysread($buf, 8192)) {
3614             if ($buf =~ /^(.*)\(qemu\) $/s) {
3615                 $res .= $1;
3616                 last;
3617             } else {
3618                 $res .= $buf;
3619             }
3620         } else {
3621             if (!defined($count)) {
3622                 die "$!\n";
3623             }
3624             last;
3625         }
3626     }
3627
3628     die "monitor read timeout\n" if !scalar(@ready);
3629
3630     return $res;
3631 }
3632
3633 # old code, only used to shutdown old VM after update
3634 sub vm_monitor_command {
3635     my ($vmid, $cmdstr, $nocheck) = @_;
3636
3637     my $res;
3638
3639     eval {
3640         die "VM $vmid not running\n" if !check_running($vmid, $nocheck);
3641
3642         my $sname = "${var_run_tmpdir}/$vmid.mon";
3643
3644         my $sock = IO::Socket::UNIX->new( Peer => $sname ) ||
3645             die "unable to connect to VM $vmid socket - $!\n";
3646
3647         my $timeout = 3;
3648
3649         # hack: migrate sometime blocks the monitor (when migrate_downtime
3650         # is set)
3651         if ($cmdstr =~ m/^(info\s+migrate|migrate\s)/) {
3652             $timeout = 60*60; # 1 hour
3653         }
3654
3655         # read banner;
3656         my $data = __read_avail($sock, $timeout);
3657
3658         if ($data !~ m/^QEMU\s+(\S+)\s+monitor\s/) {
3659             die "got unexpected qemu monitor banner\n";
3660         }
3661
3662         my $sel = new IO::Select;
3663         $sel->add($sock);
3664
3665         if (!scalar(my @ready = $sel->can_write($timeout))) {
3666             die "monitor write error - timeout";
3667         }
3668
3669         my $fullcmd = "$cmdstr\r";
3670
3671         # syslog('info', "VM $vmid monitor command: $cmdstr");
3672
3673         my $b;
3674         if (!($b = $sock->syswrite($fullcmd)) || ($b != length($fullcmd))) {
3675             die "monitor write error - $!";
3676         }
3677
3678         return if ($cmdstr eq 'q') || ($cmdstr eq 'quit');
3679
3680         $timeout = 20;
3681
3682         if ($cmdstr =~ m/^(info\s+migrate|migrate\s)/) {
3683             $timeout = 60*60; # 1 hour
3684         } elsif ($cmdstr =~ m/^(eject|change)/) {
3685             $timeout = 60; # note: cdrom mount command is slow
3686         }
3687         if ($res = __read_avail($sock, $timeout)) {
3688
3689             my @lines = split("\r?\n", $res);
3690
3691             shift @lines if $lines[0] !~ m/^unknown command/; # skip echo
3692
3693             $res = join("\n", @lines);
3694             $res .= "\n";
3695         }
3696     };
3697
3698     my $err = $@;
3699
3700     if ($err) {
3701         syslog("err", "VM $vmid monitor command failed - $err");
3702         die $err;
3703     }
3704
3705     return $res;
3706 }
3707
3708 sub qemu_block_resize {
3709     my ($vmid, $deviceid, $storecfg, $volid, $size) = @_;
3710
3711     my $running = check_running($vmid);
3712
3713     return if !PVE::Storage::volume_resize($storecfg, $volid, $size, $running);
3714
3715     return if !$running;
3716
3717     vm_mon_cmd($vmid, "block_resize", device => $deviceid, size => int($size));
3718
3719 }
3720
3721 sub qemu_volume_snapshot {
3722     my ($vmid, $deviceid, $storecfg, $volid, $snap) = @_;
3723
3724     my $running = check_running($vmid);
3725
3726     return if !PVE::Storage::volume_snapshot($storecfg, $volid, $snap, $running);
3727
3728     return if !$running;
3729
3730     vm_mon_cmd($vmid, "snapshot-drive", device => $deviceid, name => $snap);
3731
3732 }
3733
3734 sub qemu_volume_snapshot_delete {
3735     my ($vmid, $deviceid, $storecfg, $volid, $snap) = @_;
3736
3737     my $running = check_running($vmid);
3738
3739     return if !PVE::Storage::volume_snapshot_delete($storecfg, $volid, $snap, $running);
3740
3741     return if !$running;
3742
3743     vm_mon_cmd($vmid, "delete-drive-snapshot", device => $deviceid, name => $snap);
3744 }
3745
3746 sub set_migration_caps {
3747     my ($vmid) = @_;
3748
3749     my $cap_ref = [];
3750
3751     my $enabled_cap = {
3752         "auto-converge" => 1,
3753         "xbzrle" => 0,
3754         "x-rdma-pin-all" => 0,
3755         "zero-blocks" => 0,
3756     };
3757
3758     my $supported_capabilities = vm_mon_cmd_nocheck($vmid, "query-migrate-capabilities");
3759
3760     for my $supported_capability (@$supported_capabilities) {
3761         push @$cap_ref, {
3762             capability => $supported_capability->{capability},
3763             state => $enabled_cap->{$supported_capability->{capability}} ? JSON::true : JSON::false,
3764         };
3765     }
3766
3767     vm_mon_cmd_nocheck($vmid, "migrate-set-capabilities", capabilities => $cap_ref);
3768 }
3769
3770 my $fast_plug_option = {
3771     'lock' => 1,
3772     'name' => 1,
3773     'onboot' => 1,
3774     'shares' => 1,
3775     'startup' => 1,
3776 };
3777
3778 # hotplug changes in [PENDING]
3779 # $selection hash can be used to only apply specified options, for
3780 # example: { cores => 1 } (only apply changed 'cores')
3781 # $errors ref is used to return error messages
3782 sub vmconfig_hotplug_pending {
3783     my ($vmid, $conf, $storecfg, $selection, $errors) = @_;
3784
3785     my $defaults = load_defaults();
3786
3787     # commit values which do not have any impact on running VM first
3788     # Note: those option cannot raise errors, we we do not care about
3789     # $selection and always apply them.
3790
3791     my $add_error = sub {
3792         my ($opt, $msg) = @_;
3793         $errors->{$opt} = "hotplug problem - $msg";
3794     };
3795
3796     my $changes = 0;
3797     foreach my $opt (keys %{$conf->{pending}}) { # add/change
3798         if ($fast_plug_option->{$opt}) {
3799             $conf->{$opt} = $conf->{pending}->{$opt};
3800             delete $conf->{pending}->{$opt};
3801             $changes = 1;
3802         }
3803     }
3804
3805     if ($changes) {
3806         update_config_nolock($vmid, $conf, 1);
3807         $conf = load_config($vmid); # update/reload
3808     }
3809
3810     my $hotplug_features = parse_hotplug_features(defined($conf->{hotplug}) ? $conf->{hotplug} : '1');
3811
3812     my @delete = PVE::Tools::split_list($conf->{pending}->{delete});
3813     foreach my $opt (@delete) {
3814         next if $selection && !$selection->{$opt};
3815         eval {
3816             if ($opt eq 'hotplug') {
3817                 die "skip\n" if ($conf->{hotplug} =~ /memory/);
3818             } elsif ($opt eq 'tablet') {
3819                 die "skip\n" if !$hotplug_features->{usb};
3820                 if ($defaults->{tablet}) {
3821                     vm_deviceplug($storecfg, $conf, $vmid, $opt);
3822                 } else {
3823                     vm_deviceunplug($vmid, $conf, $opt);
3824                 }
3825             } elsif ($opt eq 'vcpus') {
3826                 die "skip\n" if !$hotplug_features->{cpu};
3827                 qemu_cpu_hotplug($vmid, $conf, undef);
3828             } elsif ($opt eq 'balloon') {
3829                 # enable balloon device is not hotpluggable
3830                 die "skip\n" if !defined($conf->{balloon}) || $conf->{balloon};
3831             } elsif ($fast_plug_option->{$opt}) {
3832                 # do nothing
3833             } elsif ($opt =~ m/^net(\d+)$/) {
3834                 die "skip\n" if !$hotplug_features->{network};
3835                 vm_deviceunplug($vmid, $conf, $opt);
3836             } elsif (valid_drivename($opt)) {
3837                 die "skip\n" if !$hotplug_features->{disk} || $opt =~ m/(ide|sata)(\d+)/;
3838                 vm_deviceunplug($vmid, $conf, $opt);
3839                 vmconfig_register_unused_drive($storecfg, $vmid, $conf, parse_drive($opt, $conf->{$opt}));
3840             } elsif ($opt =~ m/^memory$/) {
3841                 die "skip\n" if !$hotplug_features->{memory};
3842                 qemu_memory_hotplug($vmid, $conf, $defaults, $opt);
3843             } else {
3844                 die "skip\n";
3845             }
3846         };
3847         if (my $err = $@) {
3848             &$add_error($opt, $err) if $err ne "skip\n";
3849         } else {
3850             # save new config if hotplug was successful
3851             delete $conf->{$opt};
3852             vmconfig_undelete_pending_option($conf, $opt);
3853             update_config_nolock($vmid, $conf, 1);
3854             $conf = load_config($vmid); # update/reload
3855         }
3856     }
3857
3858     foreach my $opt (keys %{$conf->{pending}}) {
3859         next if $selection && !$selection->{$opt};
3860         my $value = $conf->{pending}->{$opt};
3861         eval {
3862             if ($opt eq 'hotplug') {
3863                 die "skip\n" if ($value =~ /memory/) || ($value !~ /memory/ && $conf->{hotplug} =~ /memory/);
3864             } elsif ($opt eq 'tablet') {
3865                 die "skip\n" if !$hotplug_features->{usb};
3866                 if ($value == 1) {
3867                     vm_deviceplug($storecfg, $conf, $vmid, $opt);
3868                 } elsif ($value == 0) {
3869                     vm_deviceunplug($vmid, $conf, $opt);
3870                 }
3871             } elsif ($opt eq 'vcpus') {
3872                 die "skip\n" if !$hotplug_features->{cpu};
3873                 qemu_cpu_hotplug($vmid, $conf, $value);
3874             } elsif ($opt eq 'balloon') {
3875                 # enable/disable balloning device is not hotpluggable
3876                 my $old_balloon_enabled =  !!(!defined($conf->{balloon}) || $conf->{balloon});
3877                 my $new_balloon_enabled =  !!(!defined($conf->{pending}->{balloon}) || $conf->{pending}->{balloon});
3878                 die "skip\n" if $old_balloon_enabled != $new_balloon_enabled;
3879
3880                 # allow manual ballooning if shares is set to zero
3881                 if ((defined($conf->{shares}) && ($conf->{shares} == 0))) {
3882                     my $balloon = $conf->{pending}->{balloon} || $conf->{memory} || $defaults->{memory};
3883                     vm_mon_cmd($vmid, "balloon", value => $balloon*1024*1024);
3884                 }
3885             } elsif ($opt =~ m/^net(\d+)$/) {
3886                 # some changes can be done without hotplug
3887                 vmconfig_update_net($storecfg, $conf, $hotplug_features->{network},
3888                                     $vmid, $opt, $value);
3889             } elsif (valid_drivename($opt)) {
3890                 # some changes can be done without hotplug
3891                 vmconfig_update_disk($storecfg, $conf, $hotplug_features->{disk},
3892                                      $vmid, $opt, $value, 1);
3893             } elsif ($opt =~ m/^memory$/) { #dimms
3894                 die "skip\n" if !$hotplug_features->{memory};
3895                 $value = qemu_memory_hotplug($vmid, $conf, $defaults, $opt, $value);
3896             } else {
3897                 die "skip\n";  # skip non-hot-pluggable options
3898             }
3899         };
3900         if (my $err = $@) {
3901             &$add_error($opt, $err) if $err ne "skip\n";
3902         } else {
3903             # save new config if hotplug was successful
3904             $conf->{$opt} = $value;
3905             delete $conf->{pending}->{$opt};
3906             update_config_nolock($vmid, $conf, 1);
3907             $conf = load_config($vmid); # update/reload
3908         }
3909     }
3910 }
3911
3912 sub vmconfig_apply_pending {
3913     my ($vmid, $conf, $storecfg) = @_;
3914
3915     # cold plug
3916
3917     my @delete = PVE::Tools::split_list($conf->{pending}->{delete});
3918     foreach my $opt (@delete) { # delete
3919         die "internal error" if $opt =~ m/^unused/;
3920         $conf = load_config($vmid); # update/reload
3921         if (!defined($conf->{$opt})) {
3922             vmconfig_undelete_pending_option($conf, $opt);
3923             update_config_nolock($vmid, $conf, 1);
3924         } elsif (valid_drivename($opt)) {
3925             vmconfig_register_unused_drive($storecfg, $vmid, $conf, parse_drive($opt, $conf->{$opt}));
3926             vmconfig_undelete_pending_option($conf, $opt);
3927             delete $conf->{$opt};
3928             update_config_nolock($vmid, $conf, 1);
3929         } else {
3930             vmconfig_undelete_pending_option($conf, $opt);
3931             delete $conf->{$opt};
3932             update_config_nolock($vmid, $conf, 1);
3933         }
3934     }
3935
3936     $conf = load_config($vmid); # update/reload
3937
3938     foreach my $opt (keys %{$conf->{pending}}) { # add/change
3939         $conf = load_config($vmid); # update/reload
3940
3941         if (defined($conf->{$opt}) && ($conf->{$opt} eq $conf->{pending}->{$opt})) {
3942             # skip if nothing changed
3943         } elsif (valid_drivename($opt)) {
3944             vmconfig_register_unused_drive($storecfg, $vmid, $conf, parse_drive($opt, $conf->{$opt}))
3945                 if defined($conf->{$opt});
3946             $conf->{$opt} = $conf->{pending}->{$opt};
3947         } else {
3948             $conf->{$opt} = $conf->{pending}->{$opt};
3949         }
3950
3951         delete $conf->{pending}->{$opt};
3952         update_config_nolock($vmid, $conf, 1);
3953     }
3954 }
3955
3956 my $safe_num_ne = sub {
3957     my ($a, $b) = @_;
3958
3959     return 0 if !defined($a) && !defined($b);
3960     return 1 if !defined($a);
3961     return 1 if !defined($b);
3962
3963     return $a != $b;
3964 };
3965
3966 my $safe_string_ne = sub {
3967     my ($a, $b) = @_;
3968
3969     return 0 if !defined($a) && !defined($b);
3970     return 1 if !defined($a);
3971     return 1 if !defined($b);
3972
3973     return $a ne $b;
3974 };
3975
3976 sub vmconfig_update_net {
3977     my ($storecfg, $conf, $hotplug, $vmid, $opt, $value) = @_;
3978
3979     my $newnet = parse_net($value);
3980
3981     if ($conf->{$opt}) {
3982         my $oldnet = parse_net($conf->{$opt});
3983
3984         if (&$safe_string_ne($oldnet->{model}, $newnet->{model}) ||
3985             &$safe_string_ne($oldnet->{macaddr}, $newnet->{macaddr}) ||
3986             &$safe_num_ne($oldnet->{queues}, $newnet->{queues}) ||
3987             !($newnet->{bridge} && $oldnet->{bridge})) { # bridge/nat mode change
3988
3989             # for non online change, we try to hot-unplug
3990             die "skip\n" if !$hotplug;
3991             vm_deviceunplug($vmid, $conf, $opt);
3992         } else {
3993
3994             die "internal error" if $opt !~ m/net(\d+)/;
3995             my $iface = "tap${vmid}i$1";
3996
3997             if (&$safe_num_ne($oldnet->{rate}, $newnet->{rate})) {
3998                 PVE::Network::tap_rate_limit($iface, $newnet->{rate});
3999             }
4000
4001             if (&$safe_string_ne($oldnet->{bridge}, $newnet->{bridge}) ||
4002                 &$safe_num_ne($oldnet->{tag}, $newnet->{tag}) ||
4003                 &$safe_num_ne($oldnet->{firewall}, $newnet->{firewall})) {
4004                 PVE::Network::tap_unplug($iface);
4005                 PVE::Network::tap_plug($iface, $newnet->{bridge}, $newnet->{tag}, $newnet->{firewall});
4006             }
4007
4008             if (&$safe_string_ne($oldnet->{link_down}, $newnet->{link_down})) {
4009                 qemu_set_link_status($vmid, $opt, !$newnet->{link_down});
4010             }
4011
4012             return 1;
4013         }
4014     }
4015
4016     if ($hotplug) {
4017         vm_deviceplug($storecfg, $conf, $vmid, $opt, $newnet);
4018     } else {
4019         die "skip\n";
4020     }
4021 }
4022
4023 sub vmconfig_update_disk {
4024     my ($storecfg, $conf, $hotplug, $vmid, $opt, $value, $force) = @_;
4025
4026     # fixme: do we need force?
4027
4028     my $drive = parse_drive($opt, $value);
4029
4030     if ($conf->{$opt}) {
4031
4032         if (my $old_drive = parse_drive($opt, $conf->{$opt}))  {
4033
4034             my $media = $drive->{media} || 'disk';
4035             my $oldmedia = $old_drive->{media} || 'disk';
4036             die "unable to change media type\n" if $media ne $oldmedia;
4037
4038             if (!drive_is_cdrom($old_drive)) {
4039
4040                 if ($drive->{file} ne $old_drive->{file}) {
4041
4042                     die "skip\n" if !$hotplug;
4043
4044                     # unplug and register as unused
4045                     vm_deviceunplug($vmid, $conf, $opt);
4046                     vmconfig_register_unused_drive($storecfg, $vmid, $conf, $old_drive)
4047
4048                 } else {
4049                     # update existing disk
4050
4051                     # skip non hotpluggable value
4052                     if (&$safe_num_ne($drive->{discard}, $old_drive->{discard}) ||
4053                         &$safe_string_ne($drive->{cache}, $old_drive->{cache})) {
4054                         die "skip\n";
4055                     }
4056
4057                     # apply throttle
4058                     if (&$safe_num_ne($drive->{mbps}, $old_drive->{mbps}) ||
4059                         &$safe_num_ne($drive->{mbps_rd}, $old_drive->{mbps_rd}) ||
4060                         &$safe_num_ne($drive->{mbps_wr}, $old_drive->{mbps_wr}) ||
4061                         &$safe_num_ne($drive->{iops}, $old_drive->{iops}) ||
4062                         &$safe_num_ne($drive->{iops_rd}, $old_drive->{iops_rd}) ||
4063                         &$safe_num_ne($drive->{iops_wr}, $old_drive->{iops_wr}) ||
4064                         &$safe_num_ne($drive->{mbps_max}, $old_drive->{mbps_max}) ||
4065                         &$safe_num_ne($drive->{mbps_rd_max}, $old_drive->{mbps_rd_max}) ||
4066                         &$safe_num_ne($drive->{mbps_wr_max}, $old_drive->{mbps_wr_max}) ||
4067                         &$safe_num_ne($drive->{iops_max}, $old_drive->{iops_max}) ||
4068                         &$safe_num_ne($drive->{iops_rd_max}, $old_drive->{iops_rd_max}) ||
4069                         &$safe_num_ne($drive->{iops_wr_max}, $old_drive->{iops_wr_max})) {
4070
4071                         qemu_block_set_io_throttle($vmid,"drive-$opt",
4072                                                    ($drive->{mbps} || 0)*1024*1024,
4073                                                    ($drive->{mbps_rd} || 0)*1024*1024,
4074                                                    ($drive->{mbps_wr} || 0)*1024*1024,
4075                                                    $drive->{iops} || 0,
4076                                                    $drive->{iops_rd} || 0,
4077                                                    $drive->{iops_wr} || 0,
4078                                                    ($drive->{mbps_max} || 0)*1024*1024,
4079                                                    ($drive->{mbps_rd_max} || 0)*1024*1024,
4080                                                    ($drive->{mbps_wr_max} || 0)*1024*1024,
4081                                                    $drive->{iops_max} || 0,
4082                                                    $drive->{iops_rd_max} || 0,
4083                                                    $drive->{iops_wr_max} || 0);
4084
4085                     }
4086
4087                     return 1;
4088                 }
4089
4090             } else { # cdrom
4091
4092                 if ($drive->{file} eq 'none') {
4093                     vm_mon_cmd($vmid, "eject",force => JSON::true,device => "drive-$opt");
4094                 } else {
4095                     my $path = get_iso_path($storecfg, $vmid, $drive->{file});
4096                     vm_mon_cmd($vmid, "eject", force => JSON::true,device => "drive-$opt"); # force eject if locked
4097                     vm_mon_cmd($vmid, "change", device => "drive-$opt",target => "$path") if $path;
4098                 }
4099
4100                 return 1;
4101             }
4102         }
4103     }
4104
4105     die "skip\n" if !$hotplug || $opt =~ m/(ide|sata)(\d+)/;
4106     # hotplug new disks
4107     vm_deviceplug($storecfg, $conf, $vmid, $opt, $drive);
4108 }
4109
4110 sub vm_start {
4111     my ($storecfg, $vmid, $statefile, $skiplock, $migratedfrom, $paused, $forcemachine, $spice_ticket) = @_;
4112
4113     lock_config($vmid, sub {
4114         my $conf = load_config($vmid, $migratedfrom);
4115
4116         die "you can't start a vm if it's a template\n" if is_template($conf);
4117
4118         check_lock($conf) if !$skiplock;
4119
4120         die "VM $vmid already running\n" if check_running($vmid, undef, $migratedfrom);
4121
4122         if (!$statefile && scalar(keys %{$conf->{pending}})) {
4123             vmconfig_apply_pending($vmid, $conf, $storecfg);
4124             $conf = load_config($vmid); # update/reload
4125         }
4126
4127         my $defaults = load_defaults();
4128
4129         # set environment variable useful inside network script
4130         $ENV{PVE_MIGRATED_FROM} = $migratedfrom if $migratedfrom;
4131
4132         my ($cmd, $vollist, $spice_port) = config_to_command($storecfg, $vmid, $conf, $defaults, $forcemachine);
4133
4134         my $migrate_port = 0;
4135         my $migrate_uri;
4136         if ($statefile) {
4137             if ($statefile eq 'tcp') {
4138                 my $localip = "localhost";
4139                 my $datacenterconf = PVE::Cluster::cfs_read_file('datacenter.cfg');
4140                 if ($datacenterconf->{migration_unsecure}) {
4141                         my $nodename = PVE::INotify::nodename();
4142                         $localip = PVE::Cluster::remote_node_ip($nodename, 1);
4143                 }
4144                 $migrate_port = PVE::Tools::next_migrate_port();
4145                 $migrate_uri = "tcp:${localip}:${migrate_port}";
4146                 push @$cmd, '-incoming', $migrate_uri;
4147                 push @$cmd, '-S';
4148             } else {
4149                 push @$cmd, '-loadstate', $statefile;
4150             }
4151         } elsif ($paused) {
4152             push @$cmd, '-S';
4153         }
4154
4155         # host pci devices
4156         for (my $i = 0; $i < $MAX_HOSTPCI_DEVICES; $i++)  {
4157           my $d = parse_hostpci($conf->{"hostpci$i"});
4158           next if !$d;
4159           my $pcidevices = $d->{pciid};
4160           foreach my $pcidevice (@$pcidevices) {
4161                 my $pciid = $pcidevice->{id}.".".$pcidevice->{function};
4162
4163                 my $info = pci_device_info("0000:$pciid");
4164                 die "IOMMU not present\n" if !check_iommu_support();
4165                 die "no pci device info for device '$pciid'\n" if !$info;
4166
4167                 if ($d->{driver} && $d->{driver} eq "vfio") {
4168                     die "can't unbind/bind pci group to vfio '$pciid'\n" if !pci_dev_group_bind_to_vfio($pciid);
4169                 } else {
4170                     die "can't unbind/bind to stub pci device '$pciid'\n" if !pci_dev_bind_to_stub($info);
4171                 }
4172
4173                 die "can't reset pci device '$pciid'\n" if $info->{has_fl_reset} and !pci_dev_reset($info);
4174           }
4175         }
4176
4177         PVE::Storage::activate_volumes($storecfg, $vollist);
4178
4179         eval  { run_command($cmd, timeout => $statefile ? undef : 30,
4180                     umask => 0077); };
4181         my $err = $@;
4182         die "start failed: $err" if $err;
4183
4184         print "migration listens on $migrate_uri\n" if $migrate_uri;
4185
4186         if ($statefile && $statefile ne 'tcp')  {
4187             eval { vm_mon_cmd_nocheck($vmid, "cont"); };
4188             warn $@ if $@;
4189         }
4190
4191         if ($migratedfrom) {
4192
4193             eval {
4194                 set_migration_caps($vmid);
4195             };
4196             warn $@ if $@;
4197
4198             if ($spice_port) {
4199                 print "spice listens on port $spice_port\n";
4200                 if ($spice_ticket) {
4201                     vm_mon_cmd_nocheck($vmid, "set_password", protocol => 'spice', password => $spice_ticket);
4202                     vm_mon_cmd_nocheck($vmid, "expire_password", protocol => 'spice', time => "+30");
4203                 }
4204             }
4205
4206         } else {
4207
4208             if (!$statefile && (!defined($conf->{balloon}) || $conf->{balloon})) {
4209                 vm_mon_cmd_nocheck($vmid, "balloon", value => $conf->{balloon}*1024*1024)
4210                     if $conf->{balloon};
4211             }
4212
4213             foreach my $opt (keys %$conf) {
4214                 next if $opt !~  m/^net\d+$/;
4215                 my $nicconf = parse_net($conf->{$opt});
4216                 qemu_set_link_status($vmid, $opt, 0) if $nicconf->{link_down};
4217             }
4218         }
4219
4220         vm_mon_cmd_nocheck($vmid, 'qom-set',
4221                     path => "machine/peripheral/balloon0",
4222                     property => "guest-stats-polling-interval",
4223                     value => 2) if (!defined($conf->{balloon}) || $conf->{balloon});
4224
4225     });
4226 }
4227
4228 sub vm_mon_cmd {
4229     my ($vmid, $execute, %params) = @_;
4230
4231     my $cmd = { execute => $execute, arguments => \%params };
4232     vm_qmp_command($vmid, $cmd);
4233 }
4234
4235 sub vm_mon_cmd_nocheck {
4236     my ($vmid, $execute, %params) = @_;
4237
4238     my $cmd = { execute => $execute, arguments => \%params };
4239     vm_qmp_command($vmid, $cmd, 1);
4240 }
4241
4242 sub vm_qmp_command {
4243     my ($vmid, $cmd, $nocheck) = @_;
4244
4245     my $res;
4246
4247     my $timeout;
4248     if ($cmd->{arguments} && $cmd->{arguments}->{timeout}) {
4249         $timeout = $cmd->{arguments}->{timeout};
4250         delete $cmd->{arguments}->{timeout};
4251     }
4252
4253     eval {
4254         die "VM $vmid not running\n" if !check_running($vmid, $nocheck);
4255         my $sname = qmp_socket($vmid);
4256         if (-e $sname) { # test if VM is reasonambe new and supports qmp/qga
4257             my $qmpclient = PVE::QMPClient->new();
4258
4259             $res = $qmpclient->cmd($vmid, $cmd, $timeout);
4260         } elsif (-e "${var_run_tmpdir}/$vmid.mon") {
4261             die "can't execute complex command on old monitor - stop/start your vm to fix the problem\n"
4262                 if scalar(%{$cmd->{arguments}});
4263             vm_monitor_command($vmid, $cmd->{execute}, $nocheck);
4264         } else {
4265             die "unable to open monitor socket\n";
4266         }
4267     };
4268     if (my $err = $@) {
4269         syslog("err", "VM $vmid qmp command failed - $err");
4270         die $err;
4271     }
4272
4273     return $res;
4274 }
4275
4276 sub vm_human_monitor_command {
4277     my ($vmid, $cmdline) = @_;
4278
4279     my $res;
4280
4281     my $cmd = {
4282         execute => 'human-monitor-command',
4283         arguments => { 'command-line' => $cmdline},
4284     };
4285
4286     return vm_qmp_command($vmid, $cmd);
4287 }
4288
4289 sub vm_commandline {
4290     my ($storecfg, $vmid) = @_;
4291
4292     my $conf = load_config($vmid);
4293
4294     my $defaults = load_defaults();
4295
4296     my $cmd = config_to_command($storecfg, $vmid, $conf, $defaults);
4297
4298     return join(' ', @$cmd);
4299 }
4300
4301 sub vm_reset {
4302     my ($vmid, $skiplock) = @_;
4303
4304     lock_config($vmid, sub {
4305
4306         my $conf = load_config($vmid);
4307
4308         check_lock($conf) if !$skiplock;
4309
4310         vm_mon_cmd($vmid, "system_reset");
4311     });
4312 }
4313
4314 sub get_vm_volumes {
4315     my ($conf) = @_;
4316
4317     my $vollist = [];
4318     foreach_volid($conf, sub {
4319         my ($volid, $is_cdrom) = @_;
4320
4321         return if $volid =~ m|^/|;
4322
4323         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
4324         return if !$sid;
4325
4326         push @$vollist, $volid;
4327     });
4328
4329     return $vollist;
4330 }
4331
4332 sub vm_stop_cleanup {
4333     my ($storecfg, $vmid, $conf, $keepActive, $apply_pending_changes) = @_;
4334
4335     eval {
4336         fairsched_rmnod($vmid); # try to destroy group
4337
4338         if (!$keepActive) {
4339             my $vollist = get_vm_volumes($conf);
4340             PVE::Storage::deactivate_volumes($storecfg, $vollist);
4341         }
4342
4343         foreach my $ext (qw(mon qmp pid vnc qga)) {
4344             unlink "/var/run/qemu-server/${vmid}.$ext";
4345         }
4346
4347         vmconfig_apply_pending($vmid, $conf, $storecfg) if $apply_pending_changes;
4348     };
4349     warn $@ if $@; # avoid errors - just warn
4350 }
4351
4352 # Note: use $nockeck to skip tests if VM configuration file exists.
4353 # We need that when migration VMs to other nodes (files already moved)
4354 # Note: we set $keepActive in vzdump stop mode - volumes need to stay active
4355 sub vm_stop {
4356     my ($storecfg, $vmid, $skiplock, $nocheck, $timeout, $shutdown, $force, $keepActive, $migratedfrom) = @_;
4357
4358     $force = 1 if !defined($force) && !$shutdown;
4359
4360     if ($migratedfrom){
4361         my $pid = check_running($vmid, $nocheck, $migratedfrom);
4362         kill 15, $pid if $pid;
4363         my $conf = load_config($vmid, $migratedfrom);
4364         vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive, 0);
4365         return;
4366     }
4367
4368     lock_config($vmid, sub {
4369
4370         my $pid = check_running($vmid, $nocheck);
4371         return if !$pid;
4372
4373         my $conf;
4374         if (!$nocheck) {
4375             $conf = load_config($vmid);
4376             check_lock($conf) if !$skiplock;
4377             if (!defined($timeout) && $shutdown && $conf->{startup}) {
4378                 my $opts = parse_startup($conf->{startup});
4379                 $timeout = $opts->{down} if $opts->{down};
4380             }
4381         }
4382
4383         $timeout = 60 if !defined($timeout);
4384
4385         eval {
4386             if ($shutdown) {
4387                 if (defined($conf) && $conf->{agent}) {
4388                     vm_qmp_command($vmid, { execute => "guest-shutdown" }, $nocheck);
4389                 } else {
4390                     vm_qmp_command($vmid, { execute => "system_powerdown" }, $nocheck);
4391                 }
4392             } else {
4393                 vm_qmp_command($vmid, { execute => "quit" }, $nocheck);
4394             }
4395         };
4396         my $err = $@;
4397
4398         if (!$err) {
4399             my $count = 0;
4400             while (($count < $timeout) && check_running($vmid, $nocheck)) {
4401                 $count++;
4402                 sleep 1;
4403             }
4404
4405             if ($count >= $timeout) {
4406                 if ($force) {
4407                     warn "VM still running - terminating now with SIGTERM\n";
4408                     kill 15, $pid;
4409                 } else {
4410                     die "VM quit/powerdown failed - got timeout\n";
4411                 }
4412             } else {
4413                 vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive, 1) if $conf;
4414                 return;
4415             }
4416         } else {
4417             if ($force) {
4418                 warn "VM quit/powerdown failed - terminating now with SIGTERM\n";
4419                 kill 15, $pid;
4420             } else {
4421                 die "VM quit/powerdown failed\n";
4422             }
4423         }
4424
4425         # wait again
4426         $timeout = 10;
4427
4428         my $count = 0;
4429         while (($count < $timeout) && check_running($vmid, $nocheck)) {
4430             $count++;
4431             sleep 1;
4432         }
4433
4434         if ($count >= $timeout) {
4435             warn "VM still running - terminating now with SIGKILL\n";
4436             kill 9, $pid;
4437             sleep 1;
4438         }
4439
4440         vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive, 1) if $conf;
4441    });
4442 }
4443
4444 sub vm_suspend {
4445     my ($vmid, $skiplock) = @_;
4446
4447     lock_config($vmid, sub {
4448
4449         my $conf = load_config($vmid);
4450
4451         check_lock($conf) if !($skiplock || ($conf->{lock} && $conf->{lock} eq 'backup'));
4452
4453         vm_mon_cmd($vmid, "stop");
4454     });
4455 }
4456
4457 sub vm_resume {
4458     my ($vmid, $skiplock) = @_;
4459
4460     lock_config($vmid, sub {
4461
4462         my $conf = load_config($vmid);
4463
4464         check_lock($conf) if !($skiplock || ($conf->{lock} && $conf->{lock} eq 'backup'));
4465
4466         vm_mon_cmd($vmid, "cont");
4467     });
4468 }
4469
4470 sub vm_sendkey {
4471     my ($vmid, $skiplock, $key) = @_;
4472
4473     lock_config($vmid, sub {
4474
4475         my $conf = load_config($vmid);
4476
4477         # there is no qmp command, so we use the human monitor command
4478         vm_human_monitor_command($vmid, "sendkey $key");
4479     });
4480 }
4481
4482 sub vm_destroy {
4483     my ($storecfg, $vmid, $skiplock) = @_;
4484
4485     lock_config($vmid, sub {
4486
4487         my $conf = load_config($vmid);
4488
4489         check_lock($conf) if !$skiplock;
4490
4491         if (!check_running($vmid)) {
4492             fairsched_rmnod($vmid); # try to destroy group
4493             destroy_vm($storecfg, $vmid);
4494         } else {
4495             die "VM $vmid is running - destroy failed\n";
4496         }
4497     });
4498 }
4499
4500 # pci helpers
4501
4502 sub file_write {
4503     my ($filename, $buf) = @_;
4504
4505     my $fh = IO::File->new($filename, "w");
4506     return undef if !$fh;
4507
4508     my $res = print $fh $buf;
4509
4510     $fh->close();
4511
4512     return $res;
4513 }
4514
4515 sub pci_device_info {
4516     my ($name) = @_;
4517
4518     my $res;
4519
4520     return undef if $name !~ m/^([a-f0-9]{4}):([a-f0-9]{2}):([a-f0-9]{2})\.([a-f0-9])$/;
4521     my ($domain, $bus, $slot, $func) = ($1, $2, $3, $4);
4522
4523     my $irq = file_read_firstline("$pcisysfs/devices/$name/irq");
4524     return undef if !defined($irq) || $irq !~ m/^\d+$/;
4525
4526     my $vendor = file_read_firstline("$pcisysfs/devices/$name/vendor");
4527     return undef if !defined($vendor) || $vendor !~ s/^0x//;
4528
4529     my $product = file_read_firstline("$pcisysfs/devices/$name/device");
4530     return undef if !defined($product) || $product !~ s/^0x//;
4531
4532     $res = {
4533         name => $name,
4534         vendor => $vendor,
4535         product => $product,
4536         domain => $domain,
4537         bus => $bus,
4538         slot => $slot,
4539         func => $func,
4540         irq => $irq,
4541         has_fl_reset => -f "$pcisysfs/devices/$name/reset" || 0,
4542     };
4543
4544     return $res;
4545 }
4546
4547 sub pci_dev_reset {
4548     my ($dev) = @_;
4549
4550     my $name = $dev->{name};
4551
4552     my $fn = "$pcisysfs/devices/$name/reset";
4553
4554     return file_write($fn, "1");
4555 }
4556
4557 sub pci_dev_bind_to_stub {
4558     my ($dev) = @_;
4559
4560     my $name = $dev->{name};
4561
4562     my $testdir = "$pcisysfs/drivers/pci-stub/$name";
4563     return 1 if -d $testdir;
4564
4565     my $data = "$dev->{vendor} $dev->{product}";
4566     return undef if !file_write("$pcisysfs/drivers/pci-stub/new_id", $data);
4567
4568     my $fn = "$pcisysfs/devices/$name/driver/unbind";
4569     if (!file_write($fn, $name)) {
4570         return undef if -f $fn;
4571     }
4572
4573     $fn = "$pcisysfs/drivers/pci-stub/bind";
4574     if (! -d $testdir) {
4575         return undef if !file_write($fn, $name);
4576     }
4577
4578     return -d $testdir;
4579 }
4580
4581 sub pci_dev_bind_to_vfio {
4582     my ($dev) = @_;
4583
4584     my $name = $dev->{name};
4585
4586     my $vfio_basedir = "$pcisysfs/drivers/vfio-pci";
4587
4588     if (!-d $vfio_basedir) {
4589         system("/sbin/modprobe vfio-pci >/dev/null 2>/dev/null");
4590     }
4591     die "Cannot find vfio-pci module!\n" if !-d $vfio_basedir;
4592
4593     my $testdir = "$vfio_basedir/$name";
4594     return 1 if -d $testdir;
4595
4596     my $data = "$dev->{vendor} $dev->{product}";
4597     return undef if !file_write("$vfio_basedir/new_id", $data);
4598
4599     my $fn = "$pcisysfs/devices/$name/driver/unbind";
4600     if (!file_write($fn, $name)) {
4601         return undef if -f $fn;
4602     }
4603
4604     $fn = "$vfio_basedir/bind";
4605     if (! -d $testdir) {
4606         return undef if !file_write($fn, $name);
4607     }
4608
4609     return -d $testdir;
4610 }
4611
4612 sub pci_dev_group_bind_to_vfio {
4613     my ($pciid) = @_;
4614
4615     my $vfio_basedir = "$pcisysfs/drivers/vfio-pci";
4616
4617     if (!-d $vfio_basedir) {
4618         system("/sbin/modprobe vfio-pci >/dev/null 2>/dev/null");
4619     }
4620     die "Cannot find vfio-pci module!\n" if !-d $vfio_basedir;
4621
4622     # get IOMMU group devices
4623     opendir(my $D, "$pcisysfs/devices/0000:$pciid/iommu_group/devices/") || die "Cannot open iommu_group: $!\n";
4624       my @devs = grep /^0000:/, readdir($D);
4625     closedir($D);
4626
4627     foreach my $pciid (@devs) {
4628         $pciid =~ m/^([:\.\da-f]+)$/ or die "PCI ID $pciid not valid!\n";
4629
4630         # pci bridges, switches or root ports are not supported
4631         # they have a pci_bus subdirectory so skip them
4632         next if (-e "$pcisysfs/devices/$pciid/pci_bus");
4633
4634         my $info = pci_device_info($1);
4635         pci_dev_bind_to_vfio($info) || die "Cannot bind $pciid to vfio\n";
4636     }
4637
4638     return 1;
4639 }
4640
4641 sub print_pci_addr {
4642     my ($id, $bridges) = @_;
4643
4644     my $res = '';
4645     my $devices = {
4646         piix3 => { bus => 0, addr => 1 },
4647         #addr2 : first videocard
4648         balloon0 => { bus => 0, addr => 3 },
4649         watchdog => { bus => 0, addr => 4 },
4650         scsihw0 => { bus => 0, addr => 5 },
4651         scsihw1 => { bus => 0, addr => 6 },
4652         ahci0 => { bus => 0, addr => 7 },
4653         qga0 => { bus => 0, addr => 8 },
4654         spice => { bus => 0, addr => 9 },
4655         virtio0 => { bus => 0, addr => 10 },
4656         virtio1 => { bus => 0, addr => 11 },
4657         virtio2 => { bus => 0, addr => 12 },
4658         virtio3 => { bus => 0, addr => 13 },
4659         virtio4 => { bus => 0, addr => 14 },
4660         virtio5 => { bus => 0, addr => 15 },
4661         hostpci0 => { bus => 0, addr => 16 },
4662         hostpci1 => { bus => 0, addr => 17 },
4663         net0 => { bus => 0, addr => 18 },
4664         net1 => { bus => 0, addr => 19 },
4665         net2 => { bus => 0, addr => 20 },
4666         net3 => { bus => 0, addr => 21 },
4667         net4 => { bus => 0, addr => 22 },
4668         net5 => { bus => 0, addr => 23 },
4669         vga1 => { bus => 0, addr => 24 },
4670         vga2 => { bus => 0, addr => 25 },
4671         vga3 => { bus => 0, addr => 26 },
4672         hostpci2 => { bus => 0, addr => 27 },
4673         hostpci3 => { bus => 0, addr => 28 },
4674         #addr29 : usb-host (pve-usb.cfg)
4675         'pci.1' => { bus => 0, addr => 30 },
4676         'pci.2' => { bus => 0, addr => 31 },
4677         'net6' => { bus => 1, addr => 1 },
4678         'net7' => { bus => 1, addr => 2 },
4679         'net8' => { bus => 1, addr => 3 },
4680         'net9' => { bus => 1, addr => 4 },
4681         'net10' => { bus => 1, addr => 5 },
4682         'net11' => { bus => 1, addr => 6 },
4683         'net12' => { bus => 1, addr => 7 },
4684         'net13' => { bus => 1, addr => 8 },
4685         'net14' => { bus => 1, addr => 9 },
4686         'net15' => { bus => 1, addr => 10 },
4687         'net16' => { bus => 1, addr => 11 },
4688         'net17' => { bus => 1, addr => 12 },
4689         'net18' => { bus => 1, addr => 13 },
4690         'net19' => { bus => 1, addr => 14 },
4691         'net20' => { bus => 1, addr => 15 },
4692         'net21' => { bus => 1, addr => 16 },
4693         'net22' => { bus => 1, addr => 17 },
4694         'net23' => { bus => 1, addr => 18 },
4695         'net24' => { bus => 1, addr => 19 },
4696         'net25' => { bus => 1, addr => 20 },
4697         'net26' => { bus => 1, addr => 21 },
4698         'net27' => { bus => 1, addr => 22 },
4699         'net28' => { bus => 1, addr => 23 },
4700         'net29' => { bus => 1, addr => 24 },
4701         'net30' => { bus => 1, addr => 25 },
4702         'net31' => { bus => 1, addr => 26 },
4703         'virtio6' => { bus => 2, addr => 1 },
4704         'virtio7' => { bus => 2, addr => 2 },
4705         'virtio8' => { bus => 2, addr => 3 },
4706         'virtio9' => { bus => 2, addr => 4 },
4707         'virtio10' => { bus => 2, addr => 5 },
4708         'virtio11' => { bus => 2, addr => 6 },
4709         'virtio12' => { bus => 2, addr => 7 },
4710         'virtio13' => { bus => 2, addr => 8 },
4711         'virtio14' => { bus => 2, addr => 9 },
4712         'virtio15' => { bus => 2, addr => 10 },
4713     };
4714
4715     if (defined($devices->{$id}->{bus}) && defined($devices->{$id}->{addr})) {
4716            my $addr = sprintf("0x%x", $devices->{$id}->{addr});
4717            my $bus = $devices->{$id}->{bus};
4718            $res = ",bus=pci.$bus,addr=$addr";
4719            $bridges->{$bus} = 1 if $bridges;
4720     }
4721     return $res;
4722
4723 }
4724
4725 sub print_pcie_addr {
4726     my ($id) = @_;
4727
4728     my $res = '';
4729     my $devices = {
4730         hostpci0 => { bus => "ich9-pcie-port-1", addr => 0 },
4731         hostpci1 => { bus => "ich9-pcie-port-2", addr => 0 },
4732         hostpci2 => { bus => "ich9-pcie-port-3", addr => 0 },
4733         hostpci3 => { bus => "ich9-pcie-port-4", addr => 0 },
4734     };
4735
4736     if (defined($devices->{$id}->{bus}) && defined($devices->{$id}->{addr})) {
4737            my $addr = sprintf("0x%x", $devices->{$id}->{addr});
4738            my $bus = $devices->{$id}->{bus};
4739            $res = ",bus=$bus,addr=$addr";
4740     }
4741     return $res;
4742
4743 }
4744
4745 # vzdump restore implementaion
4746
4747 sub tar_archive_read_firstfile {
4748     my $archive = shift;
4749
4750     die "ERROR: file '$archive' does not exist\n" if ! -f $archive;
4751
4752     # try to detect archive type first
4753     my $pid = open (TMP, "tar tf '$archive'|") ||
4754         die "unable to open file '$archive'\n";
4755     my $firstfile = <TMP>;
4756     kill 15, $pid;
4757     close TMP;
4758
4759     die "ERROR: archive contaions no data\n" if !$firstfile;
4760     chomp $firstfile;
4761
4762     return $firstfile;
4763 }
4764
4765 sub tar_restore_cleanup {
4766     my ($storecfg, $statfile) = @_;
4767
4768     print STDERR "starting cleanup\n";
4769
4770     if (my $fd = IO::File->new($statfile, "r")) {
4771         while (defined(my $line = <$fd>)) {
4772             if ($line =~ m/vzdump:([^\s:]*):(\S+)$/) {
4773                 my $volid = $2;
4774                 eval {
4775                     if ($volid =~ m|^/|) {
4776                         unlink $volid || die 'unlink failed\n';
4777                     } else {
4778                         PVE::Storage::vdisk_free($storecfg, $volid);
4779                     }
4780                     print STDERR "temporary volume '$volid' sucessfuly removed\n";
4781                 };
4782                 print STDERR "unable to cleanup '$volid' - $@" if $@;
4783             } else {
4784                 print STDERR "unable to parse line in statfile - $line";
4785             }
4786         }
4787         $fd->close();
4788     }
4789 }
4790
4791 sub restore_archive {
4792     my ($archive, $vmid, $user, $opts) = @_;
4793
4794     my $format = $opts->{format};
4795     my $comp;
4796
4797     if ($archive =~ m/\.tgz$/ || $archive =~ m/\.tar\.gz$/) {
4798         $format = 'tar' if !$format;
4799         $comp = 'gzip';
4800     } elsif ($archive =~ m/\.tar$/) {
4801         $format = 'tar' if !$format;
4802     } elsif ($archive =~ m/.tar.lzo$/) {
4803         $format = 'tar' if !$format;
4804         $comp = 'lzop';
4805     } elsif ($archive =~ m/\.vma$/) {
4806         $format = 'vma' if !$format;
4807     } elsif ($archive =~ m/\.vma\.gz$/) {
4808         $format = 'vma' if !$format;
4809         $comp = 'gzip';
4810     } elsif ($archive =~ m/\.vma\.lzo$/) {
4811         $format = 'vma' if !$format;
4812         $comp = 'lzop';
4813     } else {
4814         $format = 'vma' if !$format; # default
4815     }
4816
4817     # try to detect archive format
4818     if ($format eq 'tar') {
4819         return restore_tar_archive($archive, $vmid, $user, $opts);
4820     } else {
4821         return restore_vma_archive($archive, $vmid, $user, $opts, $comp);
4822     }
4823 }
4824
4825 sub restore_update_config_line {
4826     my ($outfd, $cookie, $vmid, $map, $line, $unique) = @_;
4827
4828     return if $line =~ m/^\#qmdump\#/;
4829     return if $line =~ m/^\#vzdump\#/;
4830     return if $line =~ m/^lock:/;
4831     return if $line =~ m/^unused\d+:/;
4832     return if $line =~ m/^parent:/;
4833     return if $line =~ m/^template:/; # restored VM is never a template
4834
4835     if (($line =~ m/^(vlan(\d+)):\s*(\S+)\s*$/)) {
4836         # try to convert old 1.X settings
4837         my ($id, $ind, $ethcfg) = ($1, $2, $3);
4838         foreach my $devconfig (PVE::Tools::split_list($ethcfg)) {
4839             my ($model, $macaddr) = split(/\=/, $devconfig);
4840             $macaddr = PVE::Tools::random_ether_addr() if !$macaddr || $unique;
4841             my $net = {
4842                 model => $model,
4843                 bridge => "vmbr$ind",
4844                 macaddr => $macaddr,
4845             };
4846             my $netstr = print_net($net);
4847
4848             print $outfd "net$cookie->{netcount}: $netstr\n";
4849             $cookie->{netcount}++;
4850         }
4851     } elsif (($line =~ m/^(net\d+):\s*(\S+)\s*$/) && $unique) {
4852         my ($id, $netstr) = ($1, $2);
4853         my $net = parse_net($netstr);
4854         $net->{macaddr} = PVE::Tools::random_ether_addr() if $net->{macaddr};
4855         $netstr = print_net($net);
4856         print $outfd "$id: $netstr\n";
4857     } elsif ($line =~ m/^((ide|scsi|virtio|sata)\d+):\s*(\S+)\s*$/) {
4858         my $virtdev = $1;
4859         my $value = $3;
4860         if ($line =~ m/backup=no/) {
4861             print $outfd "#$line";
4862         } elsif ($virtdev && $map->{$virtdev}) {
4863             my $di = parse_drive($virtdev, $value);
4864             delete $di->{format}; # format can change on restore
4865             $di->{file} = $map->{$virtdev};
4866             $value = print_drive($vmid, $di);
4867             print $outfd "$virtdev: $value\n";
4868         } else {
4869             print $outfd $line;
4870         }
4871     } else {
4872         print $outfd $line;
4873     }
4874 }
4875
4876 sub scan_volids {
4877     my ($cfg, $vmid) = @_;
4878
4879     my $info = PVE::Storage::vdisk_list($cfg, undef, $vmid);
4880
4881     my $volid_hash = {};
4882     foreach my $storeid (keys %$info) {
4883         foreach my $item (@{$info->{$storeid}}) {
4884             next if !($item->{volid} && $item->{size});
4885             $item->{path} = PVE::Storage::path($cfg, $item->{volid});
4886             $volid_hash->{$item->{volid}} = $item;
4887         }
4888     }
4889
4890     return $volid_hash;
4891 }
4892
4893 sub get_used_paths {
4894     my ($vmid, $storecfg, $conf, $scan_snapshots, $skip_drive) = @_;
4895
4896     my $used_path = {};
4897
4898     my $scan_config = sub {
4899         my ($cref, $snapname) = @_;
4900
4901         foreach my $key (keys %$cref) {
4902             my $value = $cref->{$key};
4903             if (valid_drivename($key)) {
4904                 next if $skip_drive && $key eq $skip_drive;
4905                 my $drive = parse_drive($key, $value);
4906                 next if !$drive || !$drive->{file} || drive_is_cdrom($drive);
4907                 if ($drive->{file} =~ m!^/!) {
4908                     $used_path->{$drive->{file}}++; # = 1;
4909                 } else {
4910                     my ($storeid, $volname) = PVE::Storage::parse_volume_id($drive->{file}, 1);
4911                     next if !$storeid;
4912                     my $scfg = PVE::Storage::storage_config($storecfg, $storeid, 1);
4913                     next if !$scfg;
4914                     my $path = PVE::Storage::path($storecfg, $drive->{file}, $snapname);
4915                     $used_path->{$path}++; # = 1;
4916                 }
4917             }
4918         }
4919     };
4920
4921     &$scan_config($conf);
4922
4923     undef $skip_drive;
4924
4925     if ($scan_snapshots) {
4926         foreach my $snapname (keys %{$conf->{snapshots}}) {
4927             &$scan_config($conf->{snapshots}->{$snapname}, $snapname);
4928         }
4929     }
4930
4931     return $used_path;
4932 }
4933
4934 sub update_disksize {
4935     my ($vmid, $conf, $volid_hash) = @_;
4936
4937     my $changes;
4938
4939     my $used = {};
4940
4941     # Note: it is allowed to define multiple storages with same path (alias), so
4942     # we need to check both 'volid' and real 'path' (two different volid can point
4943     # to the same path).
4944
4945     my $usedpath = {};
4946
4947     # update size info
4948     foreach my $opt (keys %$conf) {
4949         if (valid_drivename($opt)) {
4950             my $drive = parse_drive($opt, $conf->{$opt});
4951             my $volid = $drive->{file};
4952             next if !$volid;
4953
4954             $used->{$volid} = 1;
4955             if ($volid_hash->{$volid} &&
4956                 (my $path = $volid_hash->{$volid}->{path})) {
4957                 $usedpath->{$path} = 1;
4958             }
4959
4960             next if drive_is_cdrom($drive);
4961             next if !$volid_hash->{$volid};
4962
4963             $drive->{size} = $volid_hash->{$volid}->{size};
4964             my $new = print_drive($vmid, $drive);
4965             if ($new ne $conf->{$opt}) {
4966                 $changes = 1;
4967                 $conf->{$opt} = $new;
4968             }
4969         }
4970     }
4971
4972     # remove 'unusedX' entry if volume is used
4973     foreach my $opt (keys %$conf) {
4974         next if $opt !~ m/^unused\d+$/;
4975         my $volid = $conf->{$opt};
4976         my $path = $volid_hash->{$volid}->{path} if $volid_hash->{$volid};
4977         if ($used->{$volid} || ($path && $usedpath->{$path})) {
4978             $changes = 1;
4979             delete $conf->{$opt};
4980         }
4981     }
4982
4983     foreach my $volid (sort keys %$volid_hash) {
4984         next if $volid =~ m/vm-$vmid-state-/;
4985         next if $used->{$volid};
4986         my $path = $volid_hash->{$volid}->{path};
4987         next if !$path; # just to be sure
4988         next if $usedpath->{$path};
4989         $changes = 1;
4990         add_unused_volume($conf, $volid);
4991         $usedpath->{$path} = 1; # avoid to add more than once (aliases)
4992     }
4993
4994     return $changes;
4995 }
4996
4997 sub rescan {
4998     my ($vmid, $nolock) = @_;
4999
5000     my $cfg = PVE::Cluster::cfs_read_file("storage.cfg");
5001
5002     my $volid_hash = scan_volids($cfg, $vmid);
5003
5004     my $updatefn =  sub {
5005         my ($vmid) = @_;
5006
5007         my $conf = load_config($vmid);
5008
5009         check_lock($conf);
5010
5011         my $vm_volids = {};
5012         foreach my $volid (keys %$volid_hash) {
5013             my $info = $volid_hash->{$volid};
5014             $vm_volids->{$volid} = $info if $info->{vmid} && $info->{vmid} == $vmid;
5015         }
5016
5017         my $changes = update_disksize($vmid, $conf, $vm_volids);
5018
5019         update_config_nolock($vmid, $conf, 1) if $changes;
5020     };
5021
5022     if (defined($vmid)) {
5023         if ($nolock) {
5024             &$updatefn($vmid);
5025         } else {
5026             lock_config($vmid, $updatefn, $vmid);
5027         }
5028     } else {
5029         my $vmlist = config_list();
5030         foreach my $vmid (keys %$vmlist) {
5031             if ($nolock) {
5032                 &$updatefn($vmid);
5033             } else {
5034                 lock_config($vmid, $updatefn, $vmid);
5035             }
5036         }
5037     }
5038 }
5039
5040 sub restore_vma_archive {
5041     my ($archive, $vmid, $user, $opts, $comp) = @_;
5042
5043     my $input = $archive eq '-' ? "<&STDIN" : undef;
5044     my $readfrom = $archive;
5045
5046     my $uncomp = '';
5047     if ($comp) {
5048         $readfrom = '-';
5049         my $qarchive = PVE::Tools::shellquote($archive);
5050         if ($comp eq 'gzip') {
5051             $uncomp = "zcat $qarchive|";
5052         } elsif ($comp eq 'lzop') {
5053             $uncomp = "lzop -d -c $qarchive|";
5054         } else {
5055             die "unknown compression method '$comp'\n";
5056         }
5057
5058     }
5059
5060     my $tmpdir = "/var/tmp/vzdumptmp$$";
5061     rmtree $tmpdir;
5062
5063     # disable interrupts (always do cleanups)
5064     local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = sub {
5065         warn "got interrupt - ignored\n";
5066     };
5067
5068     my $mapfifo = "/var/tmp/vzdumptmp$$.fifo";
5069     POSIX::mkfifo($mapfifo, 0600);
5070     my $fifofh;
5071
5072     my $openfifo = sub {
5073         open($fifofh, '>', $mapfifo) || die $!;
5074     };
5075
5076     my $cmd = "${uncomp}vma extract -v -r $mapfifo $readfrom $tmpdir";
5077
5078     my $oldtimeout;
5079     my $timeout = 5;
5080
5081     my $devinfo = {};
5082
5083     my $rpcenv = PVE::RPCEnvironment::get();
5084
5085     my $conffile = config_file($vmid);
5086     my $tmpfn = "$conffile.$$.tmp";
5087
5088     # Note: $oldconf is undef if VM does not exists
5089     my $oldconf = PVE::Cluster::cfs_read_file(cfs_config_path($vmid));
5090
5091     my $print_devmap = sub {
5092         my $virtdev_hash = {};
5093
5094         my $cfgfn = "$tmpdir/qemu-server.conf";
5095
5096         # we can read the config - that is already extracted
5097         my $fh = IO::File->new($cfgfn, "r") ||
5098             "unable to read qemu-server.conf - $!\n";
5099
5100         while (defined(my $line = <$fh>)) {
5101             if ($line =~ m/^\#qmdump\#map:(\S+):(\S+):(\S*):(\S*):$/) {
5102                 my ($virtdev, $devname, $storeid, $format) = ($1, $2, $3, $4);
5103                 die "archive does not contain data for drive '$virtdev'\n"
5104                     if !$devinfo->{$devname};
5105                 if (defined($opts->{storage})) {
5106                     $storeid = $opts->{storage} || 'local';
5107                 } elsif (!$storeid) {
5108                     $storeid = 'local';
5109                 }
5110                 $format = 'raw' if !$format;
5111                 $devinfo->{$devname}->{devname} = $devname;
5112                 $devinfo->{$devname}->{virtdev} = $virtdev;
5113                 $devinfo->{$devname}->{format} = $format;
5114                 $devinfo->{$devname}->{storeid} = $storeid;
5115
5116                 # check permission on storage
5117                 my $pool = $opts->{pool}; # todo: do we need that?
5118                 if ($user ne 'root@pam') {
5119                     $rpcenv->check($user, "/storage/$storeid", ['Datastore.AllocateSpace']);
5120                 }
5121
5122                 $virtdev_hash->{$virtdev} = $devinfo->{$devname};
5123             }
5124         }
5125
5126         foreach my $devname (keys %$devinfo) {
5127             die "found no device mapping information for device '$devname'\n"
5128                 if !$devinfo->{$devname}->{virtdev};
5129         }
5130
5131         my $cfg = cfs_read_file('storage.cfg');
5132
5133         # create empty/temp config
5134         if ($oldconf) {
5135             PVE::Tools::file_set_contents($conffile, "memory: 128\n");
5136             foreach_drive($oldconf, sub {
5137                 my ($ds, $drive) = @_;
5138
5139                 return if drive_is_cdrom($drive);
5140
5141                 my $volid = $drive->{file};
5142
5143                 return if !$volid || $volid =~ m|^/|;
5144
5145                 my ($path, $owner) = PVE::Storage::path($cfg, $volid);
5146                 return if !$path || !$owner || ($owner != $vmid);
5147
5148                 # Note: only delete disk we want to restore
5149                 # other volumes will become unused
5150                 if ($virtdev_hash->{$ds}) {
5151                     PVE::Storage::vdisk_free($cfg, $volid);
5152                 }
5153             });
5154         }
5155
5156         my $map = {};
5157         foreach my $virtdev (sort keys %$virtdev_hash) {
5158             my $d = $virtdev_hash->{$virtdev};
5159             my $alloc_size = int(($d->{size} + 1024 - 1)/1024);
5160             my $scfg = PVE::Storage::storage_config($cfg, $d->{storeid});
5161
5162             # test if requested format is supported
5163             my ($defFormat, $validFormats) = PVE::Storage::storage_default_format($cfg, $d->{storeid});
5164             my $supported = grep { $_ eq $d->{format} } @$validFormats;
5165             $d->{format} = $defFormat if !$supported;
5166
5167             my $volid = PVE::Storage::vdisk_alloc($cfg, $d->{storeid}, $vmid,
5168                                                   $d->{format}, undef, $alloc_size);
5169             print STDERR "new volume ID is '$volid'\n";
5170             $d->{volid} = $volid;
5171             my $path = PVE::Storage::path($cfg, $volid);
5172
5173             my $write_zeros = 1;
5174             # fixme: what other storages types initialize volumes with zero?
5175             if ($scfg->{type} eq 'dir' || $scfg->{type} eq 'nfs' || $scfg->{type} eq 'glusterfs' ||
5176                 $scfg->{type} eq 'sheepdog' || $scfg->{type} eq 'rbd') {
5177                 $write_zeros = 0;
5178             }
5179
5180             print $fifofh "${write_zeros}:$d->{devname}=$path\n";
5181
5182             print "map '$d->{devname}' to '$path' (write zeros = ${write_zeros})\n";
5183             $map->{$virtdev} = $volid;
5184         }
5185
5186         $fh->seek(0, 0) || die "seek failed - $!\n";
5187
5188         my $outfd = new IO::File ($tmpfn, "w") ||
5189             die "unable to write config for VM $vmid\n";
5190
5191         my $cookie = { netcount => 0 };
5192         while (defined(my $line = <$fh>)) {
5193             restore_update_config_line($outfd, $cookie, $vmid, $map, $line, $opts->{unique});
5194         }
5195
5196         $fh->close();
5197         $outfd->close();
5198     };
5199
5200     eval {
5201         # enable interrupts
5202         local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = $SIG{PIPE} = sub {
5203             die "interrupted by signal\n";
5204         };
5205         local $SIG{ALRM} = sub { die "got timeout\n"; };
5206
5207         $oldtimeout = alarm($timeout);
5208
5209         my $parser = sub {
5210             my $line = shift;
5211
5212             print "$line\n";
5213
5214             if ($line =~ m/^DEV:\sdev_id=(\d+)\ssize:\s(\d+)\sdevname:\s(\S+)$/) {
5215                 my ($dev_id, $size, $devname) = ($1, $2, $3);
5216                 $devinfo->{$devname} = { size => $size, dev_id => $dev_id };
5217             } elsif ($line =~ m/^CTIME: /) {
5218                 # we correctly received the vma config, so we can disable
5219                 # the timeout now for disk allocation (set to 10 minutes, so
5220                 # that we always timeout if something goes wrong)
5221                 alarm(600);
5222                 &$print_devmap();
5223                 print $fifofh "done\n";
5224                 my $tmp = $oldtimeout || 0;
5225                 $oldtimeout = undef;
5226                 alarm($tmp);
5227                 close($fifofh);
5228             }
5229         };
5230
5231         print "restore vma archive: $cmd\n";
5232         run_command($cmd, input => $input, outfunc => $parser, afterfork => $openfifo);
5233     };
5234     my $err = $@;
5235
5236     alarm($oldtimeout) if $oldtimeout;
5237
5238     unlink $mapfifo;
5239
5240     if ($err) {
5241         rmtree $tmpdir;
5242         unlink $tmpfn;
5243
5244         my $cfg = cfs_read_file('storage.cfg');
5245         foreach my $devname (keys %$devinfo) {
5246             my $volid = $devinfo->{$devname}->{volid};
5247             next if !$volid;
5248             eval {
5249                 if ($volid =~ m|^/|) {
5250                     unlink $volid || die 'unlink failed\n';
5251                 } else {
5252                     PVE::Storage::vdisk_free($cfg, $volid);
5253                 }
5254                 print STDERR "temporary volume '$volid' sucessfuly removed\n";
5255             };
5256             print STDERR "unable to cleanup '$volid' - $@" if $@;
5257         }
5258         die $err;
5259     }
5260
5261     rmtree $tmpdir;
5262
5263     rename($tmpfn, $conffile) ||
5264         die "unable to commit configuration file '$conffile'\n";
5265
5266     PVE::Cluster::cfs_update(); # make sure we read new file
5267
5268     eval { rescan($vmid, 1); };
5269     warn $@ if $@;
5270 }
5271
5272 sub restore_tar_archive {
5273     my ($archive, $vmid, $user, $opts) = @_;
5274
5275     if ($archive ne '-') {
5276         my $firstfile = tar_archive_read_firstfile($archive);
5277         die "ERROR: file '$archive' dos not lock like a QemuServer vzdump backup\n"
5278             if $firstfile ne 'qemu-server.conf';
5279     }
5280
5281     my $storecfg = cfs_read_file('storage.cfg');
5282
5283     # destroy existing data - keep empty config
5284     my $vmcfgfn = config_file($vmid);
5285     destroy_vm($storecfg, $vmid, 1) if -f $vmcfgfn;
5286
5287     my $tocmd = "/usr/lib/qemu-server/qmextract";
5288
5289     $tocmd .= " --storage " . PVE::Tools::shellquote($opts->{storage}) if $opts->{storage};
5290     $tocmd .= " --pool " . PVE::Tools::shellquote($opts->{pool}) if $opts->{pool};
5291     $tocmd .= ' --prealloc' if $opts->{prealloc};
5292     $tocmd .= ' --info' if $opts->{info};
5293
5294     # tar option "xf" does not autodetect compression when read from STDIN,
5295     # so we pipe to zcat
5296     my $cmd = "zcat -f|tar xf " . PVE::Tools::shellquote($archive) . " " .
5297         PVE::Tools::shellquote("--to-command=$tocmd");
5298
5299     my $tmpdir = "/var/tmp/vzdumptmp$$";
5300     mkpath $tmpdir;
5301
5302     local $ENV{VZDUMP_TMPDIR} = $tmpdir;
5303     local $ENV{VZDUMP_VMID} = $vmid;
5304     local $ENV{VZDUMP_USER} = $user;
5305
5306     my $conffile = config_file($vmid);
5307     my $tmpfn = "$conffile.$$.tmp";
5308
5309     # disable interrupts (always do cleanups)
5310     local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = sub {
5311         print STDERR "got interrupt - ignored\n";
5312     };
5313
5314     eval {
5315         # enable interrupts
5316         local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = $SIG{PIPE} = sub {
5317             die "interrupted by signal\n";
5318         };
5319
5320         if ($archive eq '-') {
5321             print "extracting archive from STDIN\n";
5322             run_command($cmd, input => "<&STDIN");
5323         } else {
5324             print "extracting archive '$archive'\n";
5325             run_command($cmd);
5326         }
5327
5328         return if $opts->{info};
5329
5330         # read new mapping
5331         my $map = {};
5332         my $statfile = "$tmpdir/qmrestore.stat";
5333         if (my $fd = IO::File->new($statfile, "r")) {
5334             while (defined (my $line = <$fd>)) {
5335                 if ($line =~ m/vzdump:([^\s:]*):(\S+)$/) {
5336                     $map->{$1} = $2 if $1;
5337                 } else {
5338                     print STDERR "unable to parse line in statfile - $line\n";
5339                 }
5340             }
5341             $fd->close();
5342         }
5343
5344         my $confsrc = "$tmpdir/qemu-server.conf";
5345
5346         my $srcfd = new IO::File($confsrc, "r") ||
5347             die "unable to open file '$confsrc'\n";
5348
5349         my $outfd = new IO::File ($tmpfn, "w") ||
5350             die "unable to write config for VM $vmid\n";
5351
5352         my $cookie = { netcount => 0 };
5353         while (defined (my $line = <$srcfd>)) {
5354             restore_update_config_line($outfd, $cookie, $vmid, $map, $line, $opts->{unique});
5355         }
5356
5357         $srcfd->close();
5358         $outfd->close();
5359     };
5360     my $err = $@;
5361
5362     if ($err) {
5363
5364         unlink $tmpfn;
5365
5366         tar_restore_cleanup($storecfg, "$tmpdir/qmrestore.stat") if !$opts->{info};
5367
5368         die $err;
5369     }
5370
5371     rmtree $tmpdir;
5372
5373     rename $tmpfn, $conffile ||
5374         die "unable to commit configuration file '$conffile'\n";
5375
5376     PVE::Cluster::cfs_update(); # make sure we read new file
5377
5378     eval { rescan($vmid, 1); };
5379     warn $@ if $@;
5380 };
5381
5382
5383 # Internal snapshots
5384
5385 # NOTE: Snapshot create/delete involves several non-atomic
5386 # action, and can take a long time.
5387 # So we try to avoid locking the file and use 'lock' variable
5388 # inside the config file instead.
5389
5390 my $snapshot_copy_config = sub {
5391     my ($source, $dest) = @_;
5392
5393     foreach my $k (keys %$source) {
5394         next if $k eq 'snapshots';
5395         next if $k eq 'snapstate';
5396         next if $k eq 'snaptime';
5397         next if $k eq 'vmstate';
5398         next if $k eq 'lock';
5399         next if $k eq 'digest';
5400         next if $k eq 'description';
5401         next if $k =~ m/^unused\d+$/;
5402
5403         $dest->{$k} = $source->{$k};
5404     }
5405 };
5406
5407 my $snapshot_apply_config = sub {
5408     my ($conf, $snap) = @_;
5409
5410     # copy snapshot list
5411     my $newconf = {
5412         snapshots => $conf->{snapshots},
5413     };
5414
5415     # keep description and list of unused disks
5416     foreach my $k (keys %$conf) {
5417         next if !($k =~ m/^unused\d+$/ || $k eq 'description');
5418         $newconf->{$k} = $conf->{$k};
5419     }
5420
5421     &$snapshot_copy_config($snap, $newconf);
5422
5423     return $newconf;
5424 };
5425
5426 sub foreach_writable_storage {
5427     my ($conf, $func) = @_;
5428
5429     my $sidhash = {};
5430
5431     foreach my $ds (keys %$conf) {
5432         next if !valid_drivename($ds);
5433
5434         my $drive = parse_drive($ds, $conf->{$ds});
5435         next if !$drive;
5436         next if drive_is_cdrom($drive);
5437
5438         my $volid = $drive->{file};
5439
5440         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
5441         $sidhash->{$sid} = $sid if $sid;
5442     }
5443
5444     foreach my $sid (sort keys %$sidhash) {
5445         &$func($sid);
5446     }
5447 }
5448
5449 my $alloc_vmstate_volid = sub {
5450     my ($storecfg, $vmid, $conf, $snapname) = @_;
5451
5452     # Note: we try to be smart when selecting a $target storage
5453
5454     my $target;
5455
5456     # search shared storage first
5457     foreach_writable_storage($conf, sub {
5458         my ($sid) = @_;
5459         my $scfg = PVE::Storage::storage_config($storecfg, $sid);
5460         return if !$scfg->{shared};
5461
5462         $target = $sid if !$target || $scfg->{path}; # prefer file based storage
5463     });
5464
5465     if (!$target) {
5466         # now search local storage
5467         foreach_writable_storage($conf, sub {
5468             my ($sid) = @_;
5469             my $scfg = PVE::Storage::storage_config($storecfg, $sid);
5470             return if $scfg->{shared};
5471
5472             $target = $sid if !$target || $scfg->{path}; # prefer file based storage;
5473         });
5474     }
5475
5476     $target = 'local' if !$target;
5477
5478     my $driver_state_size = 500; # assume 32MB is enough to safe all driver state;
5479     # we abort live save after $conf->{memory}, so we need at max twice that space
5480     my $size = $conf->{memory}*2 + $driver_state_size;
5481
5482     my $name = "vm-$vmid-state-$snapname";
5483     my $scfg = PVE::Storage::storage_config($storecfg, $target);
5484     $name .= ".raw" if $scfg->{path}; # add filename extension for file base storage
5485     my $volid = PVE::Storage::vdisk_alloc($storecfg, $target, $vmid, 'raw', $name, $size*1024);
5486
5487     return $volid;
5488 };
5489
5490 my $snapshot_prepare = sub {
5491     my ($vmid, $snapname, $save_vmstate, $comment) = @_;
5492
5493     my $snap;
5494
5495     my $updatefn =  sub {
5496
5497         my $conf = load_config($vmid);
5498
5499         die "you can't take a snapshot if it's a template\n"
5500             if is_template($conf);
5501
5502         check_lock($conf);
5503
5504         $conf->{lock} = 'snapshot';
5505
5506         die "snapshot name '$snapname' already used\n"
5507             if defined($conf->{snapshots}->{$snapname});
5508
5509         my $storecfg = PVE::Storage::config();
5510         die "snapshot feature is not available" if !has_feature('snapshot', $conf, $storecfg);
5511
5512         $snap = $conf->{snapshots}->{$snapname} = {};
5513
5514         if ($save_vmstate && check_running($vmid)) {
5515             $snap->{vmstate} = &$alloc_vmstate_volid($storecfg, $vmid, $conf, $snapname);
5516         }
5517
5518         &$snapshot_copy_config($conf, $snap);
5519
5520         $snap->{snapstate} = "prepare";
5521         $snap->{snaptime} = time();
5522         $snap->{description} = $comment if $comment;
5523
5524         # always overwrite machine if we save vmstate. This makes sure we
5525         # can restore it later using correct machine type
5526         $snap->{machine} = get_current_qemu_machine($vmid) if $snap->{vmstate};
5527
5528         update_config_nolock($vmid, $conf, 1);
5529     };
5530
5531     lock_config($vmid, $updatefn);
5532
5533     return $snap;
5534 };
5535
5536 my $snapshot_commit = sub {
5537     my ($vmid, $snapname) = @_;
5538
5539     my $updatefn = sub {
5540
5541         my $conf = load_config($vmid);
5542
5543         die "missing snapshot lock\n"
5544             if !($conf->{lock} && $conf->{lock} eq 'snapshot');
5545
5546         my $has_machine_config = defined($conf->{machine});
5547
5548         my $snap = $conf->{snapshots}->{$snapname};
5549
5550         die "snapshot '$snapname' does not exist\n" if !defined($snap);
5551
5552         die "wrong snapshot state\n"
5553             if !($snap->{snapstate} && $snap->{snapstate} eq "prepare");
5554
5555         delete $snap->{snapstate};
5556         delete $conf->{lock};
5557
5558         my $newconf = &$snapshot_apply_config($conf, $snap);
5559
5560         delete $newconf->{machine} if !$has_machine_config;
5561
5562         $newconf->{parent} = $snapname;
5563
5564         update_config_nolock($vmid, $newconf, 1);
5565     };
5566
5567     lock_config($vmid, $updatefn);
5568 };
5569
5570 sub snapshot_rollback {
5571     my ($vmid, $snapname) = @_;
5572
5573     my $prepare = 1;
5574
5575     my $storecfg = PVE::Storage::config();
5576
5577     my $conf = load_config($vmid);
5578
5579     my $get_snapshot_config = sub {
5580
5581         die "you can't rollback if vm is a template\n" if is_template($conf);
5582
5583         my $res = $conf->{snapshots}->{$snapname};
5584
5585         die "snapshot '$snapname' does not exist\n" if !defined($res);
5586
5587         return $res;
5588     };
5589
5590     my $snap = &$get_snapshot_config();
5591
5592     foreach_drive($snap, sub {
5593         my ($ds, $drive) = @_;
5594
5595         return if drive_is_cdrom($drive);
5596
5597         my $volid = $drive->{file};
5598
5599         PVE::Storage::volume_rollback_is_possible($storecfg, $volid, $snapname);
5600     });
5601
5602     my $updatefn = sub {
5603
5604         $conf = load_config($vmid);
5605
5606         $snap = &$get_snapshot_config();
5607
5608         die "unable to rollback to incomplete snapshot (snapstate = $snap->{snapstate})\n"
5609             if $snap->{snapstate};
5610
5611         if ($prepare) {
5612             check_lock($conf);
5613             vm_stop($storecfg, $vmid, undef, undef, 5, undef, undef);
5614         }
5615
5616         die "unable to rollback vm $vmid: vm is running\n"
5617             if check_running($vmid);
5618
5619         if ($prepare) {
5620             $conf->{lock} = 'rollback';
5621         } else {
5622             die "got wrong lock\n" if !($conf->{lock} && $conf->{lock} eq 'rollback');
5623             delete $conf->{lock};
5624         }
5625
5626         my $forcemachine;
5627
5628         if (!$prepare) {
5629             my $has_machine_config = defined($conf->{machine});
5630
5631             # copy snapshot config to current config
5632             $conf = &$snapshot_apply_config($conf, $snap);
5633             $conf->{parent} = $snapname;
5634
5635             # Note: old code did not store 'machine', so we try to be smart
5636             # and guess the snapshot was generated with kvm 1.4 (pc-i440fx-1.4).
5637             $forcemachine = $conf->{machine} || 'pc-i440fx-1.4';
5638             # we remove the 'machine' configuration if not explicitly specified
5639             # in the original config.
5640             delete $conf->{machine} if $snap->{vmstate} && !$has_machine_config;
5641         }
5642
5643         update_config_nolock($vmid, $conf, 1);
5644
5645         if (!$prepare && $snap->{vmstate}) {
5646             my $statefile = PVE::Storage::path($storecfg, $snap->{vmstate});
5647             vm_start($storecfg, $vmid, $statefile, undef, undef, undef, $forcemachine);
5648         }
5649     };
5650
5651     lock_config($vmid, $updatefn);
5652
5653     foreach_drive($snap, sub {
5654         my ($ds, $drive) = @_;
5655
5656         return if drive_is_cdrom($drive);
5657
5658         my $volid = $drive->{file};
5659         my $device = "drive-$ds";
5660
5661         PVE::Storage::volume_snapshot_rollback($storecfg, $volid, $snapname);
5662     });
5663
5664     $prepare = 0;
5665     lock_config($vmid, $updatefn);
5666 }
5667
5668 my $savevm_wait = sub {
5669     my ($vmid) = @_;
5670
5671     for(;;) {
5672         my $stat = vm_mon_cmd_nocheck($vmid, "query-savevm");
5673         if (!$stat->{status}) {
5674             die "savevm not active\n";
5675         } elsif ($stat->{status} eq 'active') {
5676             sleep(1);
5677             next;
5678         } elsif ($stat->{status} eq 'completed') {
5679             last;
5680         } else {
5681             die "query-savevm returned status '$stat->{status}'\n";
5682         }
5683     }
5684 };
5685
5686 sub snapshot_create {
5687     my ($vmid, $snapname, $save_vmstate, $comment) = @_;
5688
5689     my $snap = &$snapshot_prepare($vmid, $snapname, $save_vmstate, $comment);
5690
5691     $save_vmstate = 0 if !$snap->{vmstate}; # vm is not running
5692
5693     my $config = load_config($vmid);
5694
5695     my $running = check_running($vmid);
5696
5697     my $freezefs = $running && $config->{agent};
5698     $freezefs = 0 if $snap->{vmstate}; # not needed if we save RAM
5699
5700     my $drivehash = {};
5701
5702     if ($freezefs) {
5703         eval { vm_mon_cmd($vmid, "guest-fsfreeze-freeze"); };
5704         warn "guest-fsfreeze-freeze problems - $@" if $@;
5705     }
5706
5707     eval {
5708         # create internal snapshots of all drives
5709
5710         my $storecfg = PVE::Storage::config();
5711
5712         if ($running) {
5713             if ($snap->{vmstate}) {
5714                 my $path = PVE::Storage::path($storecfg, $snap->{vmstate});
5715                 vm_mon_cmd($vmid, "savevm-start", statefile => $path);
5716                 &$savevm_wait($vmid);
5717             } else {
5718                 vm_mon_cmd($vmid, "savevm-start");
5719             }
5720         };
5721
5722         foreach_drive($snap, sub {
5723             my ($ds, $drive) = @_;
5724
5725             return if drive_is_cdrom($drive);
5726
5727             my $volid = $drive->{file};
5728             my $device = "drive-$ds";
5729
5730             qemu_volume_snapshot($vmid, $device, $storecfg, $volid, $snapname);
5731             $drivehash->{$ds} = 1;
5732        });
5733     };
5734     my $err = $@;
5735
5736     if ($running) {
5737         eval { vm_mon_cmd($vmid, "savevm-end")  };
5738         warn $@ if $@;
5739
5740         if ($freezefs) {
5741             eval { vm_mon_cmd($vmid, "guest-fsfreeze-thaw"); };
5742             warn "guest-fsfreeze-thaw problems - $@" if $@;
5743         }
5744
5745         # savevm-end is async, we need to wait
5746         for (;;) {
5747             my $stat = vm_mon_cmd_nocheck($vmid, "query-savevm");
5748             if (!$stat->{bytes}) {
5749                 last;
5750             } else {
5751                 print "savevm not yet finished\n";
5752                 sleep(1);
5753                 next;
5754             }
5755         }
5756     }
5757
5758     if ($err) {
5759         warn "snapshot create failed: starting cleanup\n";
5760         eval { snapshot_delete($vmid, $snapname, 0, $drivehash); };
5761         warn $@ if $@;
5762         die $err;
5763     }
5764
5765     &$snapshot_commit($vmid, $snapname);
5766 }
5767
5768 # Note: $drivehash is only set when called from snapshot_create.
5769 sub snapshot_delete {
5770     my ($vmid, $snapname, $force, $drivehash) = @_;
5771
5772     my $prepare = 1;
5773
5774     my $snap;
5775     my $unused = [];
5776
5777     my $unlink_parent = sub {
5778         my ($confref, $new_parent) = @_;
5779
5780         if ($confref->{parent} && $confref->{parent} eq $snapname) {
5781             if ($new_parent) {
5782                 $confref->{parent} = $new_parent;
5783             } else {
5784                 delete $confref->{parent};
5785             }
5786         }
5787     };
5788
5789     my $updatefn =  sub {
5790         my ($remove_drive) = @_;
5791
5792         my $conf = load_config($vmid);
5793
5794         if (!$drivehash) {
5795             check_lock($conf);
5796             die "you can't delete a snapshot if vm is a template\n"
5797                 if is_template($conf);
5798         }
5799
5800         $snap = $conf->{snapshots}->{$snapname};
5801
5802         die "snapshot '$snapname' does not exist\n" if !defined($snap);
5803
5804         # remove parent refs
5805         if (!$prepare) {
5806             &$unlink_parent($conf, $snap->{parent});
5807             foreach my $sn (keys %{$conf->{snapshots}}) {
5808                 next if $sn eq $snapname;
5809                 &$unlink_parent($conf->{snapshots}->{$sn}, $snap->{parent});
5810             }
5811         }
5812
5813         if ($remove_drive) {
5814             if ($remove_drive eq 'vmstate') {
5815                 delete $snap->{$remove_drive};
5816             } else {
5817                 my $drive = parse_drive($remove_drive, $snap->{$remove_drive});
5818                 my $volid = $drive->{file};
5819                 delete $snap->{$remove_drive};
5820                 add_unused_volume($conf, $volid);
5821             }
5822         }
5823
5824         if ($prepare) {
5825             $snap->{snapstate} = 'delete';
5826         } else {
5827             delete $conf->{snapshots}->{$snapname};
5828             delete $conf->{lock} if $drivehash;
5829             foreach my $volid (@$unused) {
5830                 add_unused_volume($conf, $volid);
5831             }
5832         }
5833
5834         update_config_nolock($vmid, $conf, 1);
5835     };
5836
5837     lock_config($vmid, $updatefn);
5838
5839     # now remove vmstate file
5840
5841     my $storecfg = PVE::Storage::config();
5842
5843     if ($snap->{vmstate}) {
5844         eval {  PVE::Storage::vdisk_free($storecfg, $snap->{vmstate}); };
5845         if (my $err = $@) {
5846             die $err if !$force;
5847             warn $err;
5848         }
5849         # save changes (remove vmstate from snapshot)
5850         lock_config($vmid, $updatefn, 'vmstate') if !$force;
5851     };
5852
5853     # now remove all internal snapshots
5854     foreach_drive($snap, sub {
5855         my ($ds, $drive) = @_;
5856
5857         return if drive_is_cdrom($drive);
5858
5859         my $volid = $drive->{file};
5860         my $device = "drive-$ds";
5861
5862         if (!$drivehash || $drivehash->{$ds}) {
5863             eval { qemu_volume_snapshot_delete($vmid, $device, $storecfg, $volid, $snapname); };
5864             if (my $err = $@) {
5865                 die $err if !$force;
5866                 warn $err;
5867             }
5868         }
5869
5870         # save changes (remove drive fron snapshot)
5871         lock_config($vmid, $updatefn, $ds) if !$force;
5872         push @$unused, $volid;
5873     });
5874
5875     # now cleanup config
5876     $prepare = 0;
5877     lock_config($vmid, $updatefn);
5878 }
5879
5880 sub has_feature {
5881     my ($feature, $conf, $storecfg, $snapname, $running) = @_;
5882
5883     my $err;
5884     foreach_drive($conf, sub {
5885         my ($ds, $drive) = @_;
5886
5887         return if drive_is_cdrom($drive);
5888         my $volid = $drive->{file};
5889         $err = 1 if !PVE::Storage::volume_has_feature($storecfg, $feature, $volid, $snapname, $running);
5890     });
5891
5892     return $err ? 0 : 1;
5893 }
5894
5895 sub template_create {
5896     my ($vmid, $conf, $disk) = @_;
5897
5898     my $storecfg = PVE::Storage::config();
5899
5900     foreach_drive($conf, sub {
5901         my ($ds, $drive) = @_;
5902
5903         return if drive_is_cdrom($drive);
5904         return if $disk && $ds ne $disk;
5905
5906         my $volid = $drive->{file};
5907         return if !PVE::Storage::volume_has_feature($storecfg, 'template', $volid);
5908
5909         my $voliddst = PVE::Storage::vdisk_create_base($storecfg, $volid);
5910         $drive->{file} = $voliddst;
5911         $conf->{$ds} = print_drive($vmid, $drive);
5912         update_config_nolock($vmid, $conf, 1);
5913     });
5914 }
5915
5916 sub is_template {
5917     my ($conf) = @_;
5918
5919     return 1 if defined $conf->{template} && $conf->{template} == 1;
5920 }
5921
5922 sub qemu_img_convert {
5923     my ($src_volid, $dst_volid, $size, $snapname) = @_;
5924
5925     my $storecfg = PVE::Storage::config();
5926     my ($src_storeid, $src_volname) = PVE::Storage::parse_volume_id($src_volid, 1);
5927     my ($dst_storeid, $dst_volname) = PVE::Storage::parse_volume_id($dst_volid, 1);
5928
5929     if ($src_storeid && $dst_storeid) {
5930         my $src_scfg = PVE::Storage::storage_config($storecfg, $src_storeid);
5931         my $dst_scfg = PVE::Storage::storage_config($storecfg, $dst_storeid);
5932
5933         my $src_format = qemu_img_format($src_scfg, $src_volname);
5934         my $dst_format = qemu_img_format($dst_scfg, $dst_volname);
5935
5936         my $src_path = PVE::Storage::path($storecfg, $src_volid, $snapname);
5937         my $dst_path = PVE::Storage::path($storecfg, $dst_volid);
5938
5939         my $cmd = [];
5940         push @$cmd, '/usr/bin/qemu-img', 'convert', '-t', 'writeback', '-p', '-n';
5941         push @$cmd, '-s', $snapname if($snapname && $src_format eq "qcow2");
5942         push @$cmd, '-f', $src_format, '-O', $dst_format, $src_path, $dst_path;
5943
5944         my $parser = sub {
5945             my $line = shift;
5946             if($line =~ m/\((\S+)\/100\%\)/){
5947                 my $percent = $1;
5948                 my $transferred = int($size * $percent / 100);
5949                 my $remaining = $size - $transferred;
5950
5951                 print "transferred: $transferred bytes remaining: $remaining bytes total: $size bytes progression: $percent %\n";
5952             }
5953
5954         };
5955
5956         eval  { run_command($cmd, timeout => undef, outfunc => $parser); };
5957         my $err = $@;
5958         die "copy failed: $err" if $err;
5959     }
5960 }
5961
5962 sub qemu_img_format {
5963     my ($scfg, $volname) = @_;
5964
5965     if ($scfg->{path} && $volname =~ m/\.(raw|qcow2|qed|vmdk)$/) {
5966         return $1;
5967     } elsif ($scfg->{type} eq 'iscsi') {
5968         return "host_device";
5969     } else {
5970         return "raw";
5971     }
5972 }
5973
5974 sub qemu_drive_mirror {
5975     my ($vmid, $drive, $dst_volid, $vmiddst) = @_;
5976
5977     my $count = 0;
5978     my $old_len = 0;
5979     my $frozen = undef;
5980     my $maxwait = 120;
5981
5982     my $storecfg = PVE::Storage::config();
5983     my ($dst_storeid, $dst_volname) = PVE::Storage::parse_volume_id($dst_volid);
5984
5985     my $dst_scfg = PVE::Storage::storage_config($storecfg, $dst_storeid);
5986
5987     my $format;
5988     if ($dst_volname =~ m/\.(raw|qcow2)$/){
5989         $format = $1;
5990     }
5991
5992     my $dst_path = PVE::Storage::path($storecfg, $dst_volid);
5993
5994     my $opts = { timeout => 10, device => "drive-$drive", mode => "existing", sync => "full", target => $dst_path };
5995     $opts->{format} = $format if $format;
5996
5997     #fixme : sometime drive-mirror timeout, but works fine after.
5998     # (I have see the problem with big volume > 200GB), so we need to eval
5999     eval { vm_mon_cmd($vmid, "drive-mirror", %$opts); };
6000     # ignore errors here
6001
6002     eval {
6003         while (1) {
6004             my $stats = vm_mon_cmd($vmid, "query-block-jobs");
6005             my $stat = @$stats[0];
6006             die "mirroring job seem to have die. Maybe do you have bad sectors?" if !$stat;
6007             die "error job is not mirroring" if $stat->{type} ne "mirror";
6008
6009             my $busy = $stat->{busy};
6010
6011             if (my $total = $stat->{len}) {
6012                 my $transferred = $stat->{offset} || 0;
6013                 my $remaining = $total - $transferred;
6014                 my $percent = sprintf "%.2f", ($transferred * 100 / $total);
6015
6016                 print "transferred: $transferred bytes remaining: $remaining bytes total: $total bytes progression: $percent % busy: $busy\n";
6017             }
6018
6019             if ($stat->{len} == $stat->{offset}) {
6020                 if ($busy eq 'false') {
6021
6022                     last if $vmiddst != $vmid;
6023
6024                     # try to switch the disk if source and destination are on the same guest
6025                     eval { vm_mon_cmd($vmid, "block-job-complete", device => "drive-$drive") };
6026                     last if !$@;
6027                     die $@ if $@ !~ m/cannot be completed/;
6028                 }
6029
6030                 if ($count > $maxwait) {
6031                     # if too much writes to disk occurs at the end of migration
6032                     #the disk needs to be freezed to be able to complete the migration
6033                     vm_suspend($vmid,1);
6034                     $frozen = 1;
6035                 }
6036                 $count ++
6037             }
6038             $old_len = $stat->{offset};
6039             sleep 1;
6040         }
6041
6042         vm_resume($vmid, 1) if $frozen;
6043
6044     };
6045     my $err = $@;
6046
6047     my $cancel_job = sub {
6048         vm_mon_cmd($vmid, "block-job-cancel", device => "drive-$drive");
6049         while (1) {
6050             my $stats = vm_mon_cmd($vmid, "query-block-jobs");
6051             my $stat = @$stats[0];
6052             last if !$stat;
6053             sleep 1;
6054         }
6055     };
6056
6057     if ($err) {
6058         eval { &$cancel_job(); };
6059         die "mirroring error: $err";
6060     }
6061
6062     if ($vmiddst != $vmid) {
6063         # if we clone a disk for a new target vm, we don't switch the disk
6064         &$cancel_job(); # so we call block-job-cancel
6065     }
6066 }
6067
6068 sub clone_disk {
6069     my ($storecfg, $vmid, $running, $drivename, $drive, $snapname,
6070         $newvmid, $storage, $format, $full, $newvollist) = @_;
6071
6072     my $newvolid;
6073
6074     if (!$full) {
6075         print "create linked clone of drive $drivename ($drive->{file})\n";
6076         $newvolid = PVE::Storage::vdisk_clone($storecfg,  $drive->{file}, $newvmid, $snapname);
6077         push @$newvollist, $newvolid;
6078     } else {
6079         my ($storeid, $volname) = PVE::Storage::parse_volume_id($drive->{file});
6080         $storeid = $storage if $storage;
6081
6082         my ($defFormat, $validFormats) = PVE::Storage::storage_default_format($storecfg, $storeid);
6083         if (!$format) {
6084             $format = $drive->{format} || $defFormat;
6085         }
6086
6087         # test if requested format is supported - else use default
6088         my $supported = grep { $_ eq $format } @$validFormats;
6089         $format = $defFormat if !$supported;
6090
6091         my ($size) = PVE::Storage::volume_size_info($storecfg, $drive->{file}, 3);
6092
6093         print "create full clone of drive $drivename ($drive->{file})\n";
6094         $newvolid = PVE::Storage::vdisk_alloc($storecfg, $storeid, $newvmid, $format, undef, ($size/1024));
6095         push @$newvollist, $newvolid;
6096
6097         if (!$running || $snapname) {
6098             qemu_img_convert($drive->{file}, $newvolid, $size, $snapname);
6099         } else {
6100             qemu_drive_mirror($vmid, $drivename, $newvolid, $newvmid);
6101         }
6102     }
6103
6104     my ($size) = PVE::Storage::volume_size_info($storecfg, $newvolid, 3);
6105
6106     my $disk = $drive;
6107     $disk->{format} = undef;
6108     $disk->{file} = $newvolid;
6109     $disk->{size} = $size;
6110
6111     return $disk;
6112 }
6113
6114 # this only works if VM is running
6115 sub get_current_qemu_machine {
6116     my ($vmid) = @_;
6117
6118     my $cmd = { execute => 'query-machines', arguments => {} };
6119     my $res = vm_qmp_command($vmid, $cmd);
6120
6121     my ($current, $default);
6122     foreach my $e (@$res) {
6123         $default = $e->{name} if $e->{'is-default'};
6124         $current = $e->{name} if $e->{'is-current'};
6125     }
6126
6127     # fallback to the default machine if current is not supported by qemu
6128     return $current || $default || 'pc';
6129 }
6130
6131 sub qemu_machine_feature_enabled {
6132     my ($machine, $kvmver, $version_major, $version_minor) = @_;
6133
6134     my $current_major;
6135     my $current_minor;
6136
6137     if ($machine && $machine =~ m/^(pc(-i440fx|-q35)?-(\d+)\.(\d+))/) {
6138
6139         $current_major = $3;
6140         $current_minor = $4;
6141
6142     } elsif ($kvmver =~ m/^(\d+)\.(\d+)/) {
6143
6144         $current_major = $1;
6145         $current_minor = $2;
6146     }
6147
6148     return 1 if $current_major >= $version_major && $current_minor >= $version_minor;
6149
6150
6151 }
6152
6153 sub lspci {
6154
6155     my $devices = {};
6156
6157     dir_glob_foreach("$pcisysfs/devices", '[a-f0-9]{4}:([a-f0-9]{2}:[a-f0-9]{2})\.([0-9])', sub {
6158             my (undef, $id, $function) = @_;
6159             my $res = { id => $id, function => $function};
6160             push @{$devices->{$id}}, $res;
6161     });
6162
6163     return $devices;
6164 }
6165
6166 1;