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 !$nokvm;
2915         push @$cpuFlags , '+kvm_pv_eoi' if !$nokvm;
2916     }
2917
2918     $cpu .= "," . join(',', @$cpuFlags) if scalar(@$cpuFlags);
2919
2920     # Note: enforce needs kernel 3.10, so we do not use it for now
2921     # push @$cmd, '-cpu', "$cpu,enforce";
2922     push @$cmd, '-cpu', $cpu;
2923
2924     my $memory = $conf->{memory} || $defaults->{memory};
2925     my $static_memory = 0;
2926     my $dimm_memory = 0;
2927
2928     if ($hotplug_features->{memory}) {
2929         die "Numa need to be enabled for memory hotplug\n" if !$conf->{numa};
2930         die "Total memory is bigger than ${MAX_MEM}MB\n" if $memory > $MAX_MEM;
2931         $static_memory = $STATICMEM;
2932         die "minimum memory must be ${static_memory}MB\n" if($memory < $static_memory);
2933         $dimm_memory = $memory - $static_memory;
2934         push @$cmd, '-m', "size=${static_memory},slots=255,maxmem=${MAX_MEM}M";
2935
2936     } else {
2937
2938         $static_memory = $memory;
2939         push @$cmd, '-m', $static_memory;
2940     }
2941
2942     if ($conf->{numa}) {
2943
2944         my $numa_totalmemory = undef;
2945         for (my $i = 0; $i < $MAX_NUMA; $i++) {
2946             next if !$conf->{"numa$i"};
2947             my $numa = parse_numa($conf->{"numa$i"});
2948             next if !$numa;
2949             # memory
2950             die "missing numa node$i memory value\n" if !$numa->{memory};
2951             my $numa_memory = $numa->{memory};
2952             $numa_totalmemory += $numa_memory;
2953             my $numa_object = "memory-backend-ram,id=ram-node$i,size=${numa_memory}M";
2954
2955             # cpus
2956             my $cpus_start = $numa->{cpus}->{start};
2957             die "missing numa node$i cpus\n" if !defined($cpus_start);
2958             my $cpus_end = $numa->{cpus}->{end} if defined($numa->{cpus}->{end});
2959             my $cpus = $cpus_start;
2960             if (defined($cpus_end)) {
2961                 $cpus .= "-$cpus_end";
2962                 die "numa node$i :  cpu range $cpus is incorrect\n" if $cpus_end <= $cpus_start;
2963             }
2964
2965             # hostnodes
2966             my $hostnodes_start = $numa->{hostnodes}->{start};
2967             if (defined($hostnodes_start)) {
2968                 my $hostnodes_end = $numa->{hostnodes}->{end} if defined($numa->{hostnodes}->{end});
2969                 my $hostnodes = $hostnodes_start;
2970                 if (defined($hostnodes_end)) {
2971                     $hostnodes .= "-$hostnodes_end";
2972                     die "host node $hostnodes range is incorrect\n" if $hostnodes_end <= $hostnodes_start;
2973                 }
2974
2975                 my $hostnodes_end_range = defined($hostnodes_end) ? $hostnodes_end : $hostnodes_start;
2976                 for (my $i = $hostnodes_start; $i <= $hostnodes_end_range; $i++ ) {
2977                     die "host numa node$i don't exist\n" if ! -d "/sys/devices/system/node/node$i/";
2978                 }
2979
2980                 # policy
2981                 my $policy = $numa->{policy};
2982                 die "you need to define a policy for hostnode $hostnodes\n" if !$policy;
2983                 $numa_object .= ",host-nodes=$hostnodes,policy=$policy";
2984             }
2985
2986             push @$cmd, '-object', $numa_object;
2987             push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i";
2988         }
2989
2990         die "total memory for NUMA nodes must be equal to vm static memory\n"
2991             if $numa_totalmemory && $numa_totalmemory != $static_memory;
2992
2993         #if no custom tology, we split memory and cores across numa nodes
2994         if(!$numa_totalmemory) {
2995
2996             my $numa_memory = ($static_memory / $sockets) . "M";
2997
2998             for (my $i = 0; $i < $sockets; $i++)  {
2999
3000                 my $cpustart = ($cores * $i);
3001                 my $cpuend = ($cpustart + $cores - 1) if $cores && $cores > 1;
3002                 my $cpus = $cpustart;
3003                 $cpus .= "-$cpuend" if $cpuend;
3004
3005                 push @$cmd, '-object', "memory-backend-ram,size=$numa_memory,id=ram-node$i";
3006                 push @$cmd, '-numa', "node,nodeid=$i,cpus=$cpus,memdev=ram-node$i";
3007             }
3008         }
3009     }
3010
3011     if ($hotplug_features->{memory}) {
3012         foreach_dimm($conf, $vmid, $memory, $sockets, sub {
3013             my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
3014             push @$cmd, "-object" , "memory-backend-ram,id=mem-$name,size=${dimm_size}M";
3015             push @$cmd, "-device", "pc-dimm,id=$name,memdev=mem-$name,node=$numanode";
3016
3017             #if dimm_memory is not aligned to dimm map
3018             if($current_size > $memory) {
3019                  $conf->{memory} = $current_size;
3020                  update_config_nolock($vmid, $conf, 1);
3021             }
3022         });
3023     }
3024
3025     push @$cmd, '-S' if $conf->{freeze};
3026
3027     # set keyboard layout
3028     my $kb = $conf->{keyboard} || $defaults->{keyboard};
3029     push @$cmd, '-k', $kb if $kb;
3030
3031     # enable sound
3032     #my $soundhw = $conf->{soundhw} || $defaults->{soundhw};
3033     #push @$cmd, '-soundhw', 'es1370';
3034     #push @$cmd, '-soundhw', $soundhw if $soundhw;
3035
3036     if($conf->{agent}) {
3037         my $qgasocket = qmp_socket($vmid, 1);
3038         my $pciaddr = print_pci_addr("qga0", $bridges);
3039         push @$devices, '-chardev', "socket,path=$qgasocket,server,nowait,id=qga0";
3040         push @$devices, '-device', "virtio-serial,id=qga0$pciaddr";
3041         push @$devices, '-device', 'virtserialport,chardev=qga0,name=org.qemu.guest_agent.0';
3042     }
3043
3044     my $spice_port;
3045
3046     if ($qxlnum) {
3047         if ($qxlnum > 1) {
3048             if ($conf->{ostype} && $conf->{ostype} =~ m/^w/){
3049                 for(my $i = 1; $i < $qxlnum; $i++){
3050                     my $pciaddr = print_pci_addr("vga$i", $bridges);
3051                     push @$cmd, '-device', "qxl,id=vga$i,ram_size=67108864,vram_size=33554432$pciaddr";
3052                 }
3053             } else {
3054                 # assume other OS works like Linux
3055                 push @$cmd, '-global', 'qxl-vga.ram_size=134217728';
3056                 push @$cmd, '-global', 'qxl-vga.vram_size=67108864';
3057             }
3058         }
3059
3060         my $pciaddr = print_pci_addr("spice", $bridges);
3061
3062         $spice_port = PVE::Tools::next_spice_port();
3063
3064         push @$devices, '-spice', "tls-port=${spice_port},addr=127.0.0.1,tls-ciphers=DES-CBC3-SHA,seamless-migration=on";
3065
3066         push @$devices, '-device', "virtio-serial,id=spice$pciaddr";
3067         push @$devices, '-chardev', "spicevmc,id=vdagent,name=vdagent";
3068         push @$devices, '-device', "virtserialport,chardev=vdagent,name=com.redhat.spice.0";
3069     }
3070
3071     # enable balloon by default, unless explicitly disabled
3072     if (!defined($conf->{balloon}) || $conf->{balloon}) {
3073         $pciaddr = print_pci_addr("balloon0", $bridges);
3074         push @$devices, '-device', "virtio-balloon-pci,id=balloon0$pciaddr";
3075     }
3076
3077     if ($conf->{watchdog}) {
3078         my $wdopts = parse_watchdog($conf->{watchdog});
3079         $pciaddr = print_pci_addr("watchdog", $bridges);
3080         my $watchdog = $wdopts->{model} || 'i6300esb';
3081         push @$devices, '-device', "$watchdog$pciaddr";
3082         push @$devices, '-watchdog-action', $wdopts->{action} if $wdopts->{action};
3083     }
3084
3085     my $vollist = [];
3086     my $scsicontroller = {};
3087     my $ahcicontroller = {};
3088     my $scsihw = defined($conf->{scsihw}) ? $conf->{scsihw} : $defaults->{scsihw};
3089
3090     # Add iscsi initiator name if available
3091     if (my $initiator = get_initiator_name()) {
3092         push @$devices, '-iscsi', "initiator-name=$initiator";
3093     }
3094
3095     foreach_drive($conf, sub {
3096         my ($ds, $drive) = @_;
3097
3098         if (PVE::Storage::parse_volume_id($drive->{file}, 1)) {
3099             push @$vollist, $drive->{file};
3100         }
3101
3102         $use_virtio = 1 if $ds =~ m/^virtio/;
3103
3104         if (drive_is_cdrom ($drive)) {
3105             if ($bootindex_hash->{d}) {
3106                 $drive->{bootindex} = $bootindex_hash->{d};
3107                 $bootindex_hash->{d} += 1;
3108             }
3109         } else {
3110             if ($bootindex_hash->{c}) {
3111                 $drive->{bootindex} = $bootindex_hash->{c} if $conf->{bootdisk} && ($conf->{bootdisk} eq $ds);
3112                 $bootindex_hash->{c} += 1;
3113             }
3114         }
3115
3116         if ($drive->{interface} eq 'scsi') {
3117
3118             my $maxdev = ($scsihw !~ m/^lsi/) ? 256 : 7;
3119             my $controller = int($drive->{index} / $maxdev);
3120             $pciaddr = print_pci_addr("scsihw$controller", $bridges);
3121             push @$devices, '-device', "$scsihw,id=scsihw$controller$pciaddr" if !$scsicontroller->{$controller};
3122             $scsicontroller->{$controller}=1;
3123         }
3124
3125         if ($drive->{interface} eq 'sata') {
3126            my $controller = int($drive->{index} / $MAX_SATA_DISKS);
3127            $pciaddr = print_pci_addr("ahci$controller", $bridges);
3128            push @$devices, '-device', "ahci,id=ahci$controller,multifunction=on$pciaddr" if !$ahcicontroller->{$controller};
3129            $ahcicontroller->{$controller}=1;
3130         }
3131
3132         my $drive_cmd = print_drive_full($storecfg, $vmid, $drive);
3133         push @$devices, '-drive',$drive_cmd;
3134         push @$devices, '-device', print_drivedevice_full($storecfg, $conf, $vmid, $drive, $bridges);
3135     });
3136
3137     for (my $i = 0; $i < $MAX_NETS; $i++) {
3138          next if !$conf->{"net$i"};
3139          my $d = parse_net($conf->{"net$i"});
3140          next if !$d;
3141
3142          $use_virtio = 1 if $d->{model} eq 'virtio';
3143
3144          if ($bootindex_hash->{n}) {
3145             $d->{bootindex} = $bootindex_hash->{n};
3146             $bootindex_hash->{n} += 1;
3147          }
3148
3149          my $netdevfull = print_netdev_full($vmid,$conf,$d,"net$i");
3150          push @$devices, '-netdev', $netdevfull;
3151
3152          my $netdevicefull = print_netdevice_full($vmid,$conf,$d,"net$i",$bridges);
3153          push @$devices, '-device', $netdevicefull;
3154     }
3155
3156     if (!$q35) {
3157         # add pci bridges
3158         while (my ($k, $v) = each %$bridges) {
3159             $pciaddr = print_pci_addr("pci.$k");
3160             unshift @$devices, '-device', "pci-bridge,id=pci.$k,chassis_nr=$k$pciaddr" if $k > 0;
3161         }
3162     }
3163
3164     # hack: virtio with fairsched is unreliable, so we do not use fairsched
3165     # when the VM uses virtio devices.
3166     if (!$use_virtio && $have_ovz) {
3167
3168         my $cpuunits = defined($conf->{cpuunits}) ?
3169             $conf->{cpuunits} : $defaults->{cpuunits};
3170
3171         push @$cmd, '-cpuunits', $cpuunits if $cpuunits;
3172
3173         # fixme: cpulimit is currently ignored
3174         #push @$cmd, '-cpulimit', $conf->{cpulimit} if $conf->{cpulimit};
3175     }
3176
3177     # add custom args
3178     if ($conf->{args}) {
3179         my $aa = PVE::Tools::split_args($conf->{args});
3180         push @$cmd, @$aa;
3181     }
3182
3183     push @$cmd, @$devices;
3184     push @$cmd, '-rtc', join(',', @$rtcFlags)
3185         if scalar(@$rtcFlags);
3186     push @$cmd, '-machine', join(',', @$machineFlags)
3187         if scalar(@$machineFlags);
3188     push @$cmd, '-global', join(',', @$globalFlags)
3189         if scalar(@$globalFlags);
3190
3191     return wantarray ? ($cmd, $vollist, $spice_port) : $cmd;
3192 }
3193
3194 sub vnc_socket {
3195     my ($vmid) = @_;
3196     return "${var_run_tmpdir}/$vmid.vnc";
3197 }
3198
3199 sub spice_port {
3200     my ($vmid) = @_;
3201
3202     my $res = vm_mon_cmd($vmid, 'query-spice');
3203
3204     return $res->{'tls-port'} || $res->{'port'} || die "no spice port\n";
3205 }
3206
3207 sub qmp_socket {
3208     my ($vmid, $qga) = @_;
3209     my $sockettype = $qga ? 'qga' : 'qmp';
3210     return "${var_run_tmpdir}/$vmid.$sockettype";
3211 }
3212
3213 sub pidfile_name {
3214     my ($vmid) = @_;
3215     return "${var_run_tmpdir}/$vmid.pid";
3216 }
3217
3218 sub vm_devices_list {
3219     my ($vmid) = @_;
3220
3221     my $res = vm_mon_cmd($vmid, 'query-pci');
3222     my $devices = {};
3223     foreach my $pcibus (@$res) {
3224         foreach my $device (@{$pcibus->{devices}}) {
3225             next if !$device->{'qdev_id'};
3226             if ($device->{'pci_bridge'}) {
3227                 $devices->{$device->{'qdev_id'}} = 1;
3228                 foreach my $bridge_device (@{$device->{'pci_bridge'}->{devices}}) {
3229                     next if !$bridge_device->{'qdev_id'};
3230                     $devices->{$bridge_device->{'qdev_id'}} = 1;
3231                     $devices->{$device->{'qdev_id'}}++;
3232                 }
3233             } else {
3234                 $devices->{$device->{'qdev_id'}} = 1;
3235             }
3236         }
3237     }
3238
3239     my $resblock = vm_mon_cmd($vmid, 'query-block');
3240     foreach my $block (@$resblock) {
3241         if($block->{device} =~ m/^drive-(\S+)/){
3242                 $devices->{$1} = 1;
3243         }
3244     }
3245
3246     my $resmice = vm_mon_cmd($vmid, 'query-mice');
3247     foreach my $mice (@$resmice) {
3248         if ($mice->{name} eq 'QEMU HID Tablet') {
3249             $devices->{tablet} = 1;
3250             last;
3251         }
3252     }
3253
3254     return $devices;
3255 }
3256
3257 sub vm_deviceplug {
3258     my ($storecfg, $conf, $vmid, $deviceid, $device) = @_;
3259
3260     my $q35 = machine_type_is_q35($conf);
3261
3262     my $devices_list = vm_devices_list($vmid);
3263     return 1 if defined($devices_list->{$deviceid});
3264
3265     qemu_add_pci_bridge($storecfg, $conf, $vmid, $deviceid); # add PCI bridge if we need it for the device
3266
3267     if ($deviceid eq 'tablet') {
3268
3269         qemu_deviceadd($vmid, print_tabletdevice_full($conf));
3270
3271     } elsif ($deviceid =~ m/^(virtio)(\d+)$/) {
3272
3273         qemu_driveadd($storecfg, $vmid, $device);
3274         my $devicefull = print_drivedevice_full($storecfg, $conf, $vmid, $device);
3275
3276         qemu_deviceadd($vmid, $devicefull);
3277         eval { qemu_deviceaddverify($vmid, $deviceid); };
3278         if (my $err = $@) {
3279             eval { qemu_drivedel($vmid, $deviceid); };
3280             warn $@ if $@;
3281             die $err;
3282         }
3283
3284     } elsif ($deviceid =~ m/^(scsihw)(\d+)$/) {
3285
3286         my $scsihw = defined($conf->{scsihw}) ? $conf->{scsihw} : "lsi";
3287         my $pciaddr = print_pci_addr($deviceid);
3288         my $devicefull = "$scsihw,id=$deviceid$pciaddr";
3289
3290         qemu_deviceadd($vmid, $devicefull);
3291         qemu_deviceaddverify($vmid, $deviceid);
3292
3293     } elsif ($deviceid =~ m/^(scsi)(\d+)$/) {
3294
3295         qemu_findorcreatescsihw($storecfg,$conf, $vmid, $device);
3296         qemu_driveadd($storecfg, $vmid, $device);
3297
3298         my $devicefull = print_drivedevice_full($storecfg, $conf, $vmid, $device);
3299         eval { qemu_deviceadd($vmid, $devicefull); };
3300         if (my $err = $@) {
3301             eval { qemu_drivedel($vmid, $deviceid); };
3302             warn $@ if $@;
3303             die $err;
3304         }
3305
3306     } elsif ($deviceid =~ m/^(net)(\d+)$/) {
3307
3308         return undef if !qemu_netdevadd($vmid, $conf, $device, $deviceid);
3309         my $netdevicefull = print_netdevice_full($vmid, $conf, $device, $deviceid);
3310         qemu_deviceadd($vmid, $netdevicefull);
3311         eval { qemu_deviceaddverify($vmid, $deviceid); };
3312         if (my $err = $@) {
3313             eval { qemu_netdevdel($vmid, $deviceid); };
3314             warn $@ if $@;
3315             die $err;
3316         }
3317
3318     } elsif (!$q35 && $deviceid =~ m/^(pci\.)(\d+)$/) {
3319
3320         my $bridgeid = $2;
3321         my $pciaddr = print_pci_addr($deviceid);
3322         my $devicefull = "pci-bridge,id=pci.$bridgeid,chassis_nr=$bridgeid$pciaddr";
3323
3324         qemu_deviceadd($vmid, $devicefull);
3325         qemu_deviceaddverify($vmid, $deviceid);
3326
3327     } else {
3328         die "can't hotplug device '$deviceid'\n";
3329     }
3330
3331     return 1;
3332 }
3333
3334 # fixme: this should raise exceptions on error!
3335 sub vm_deviceunplug {
3336     my ($vmid, $conf, $deviceid) = @_;
3337
3338     my $devices_list = vm_devices_list($vmid);
3339     return 1 if !defined($devices_list->{$deviceid});
3340
3341     die "can't unplug bootdisk" if $conf->{bootdisk} && $conf->{bootdisk} eq $deviceid;
3342
3343     if ($deviceid eq 'tablet') {
3344
3345         qemu_devicedel($vmid, $deviceid);
3346
3347     } elsif ($deviceid =~ m/^(virtio)(\d+)$/) {
3348
3349         qemu_devicedel($vmid, $deviceid);
3350         qemu_devicedelverify($vmid, $deviceid);
3351         qemu_drivedel($vmid, $deviceid);
3352
3353     } elsif ($deviceid =~ m/^(lsi)(\d+)$/) {
3354
3355         qemu_devicedel($vmid, $deviceid);
3356
3357     } elsif ($deviceid =~ m/^(scsi)(\d+)$/) {
3358
3359         qemu_devicedel($vmid, $deviceid);
3360         qemu_drivedel($vmid, $deviceid);
3361
3362     } elsif ($deviceid =~ m/^(net)(\d+)$/) {
3363
3364         qemu_devicedel($vmid, $deviceid);
3365         qemu_devicedelverify($vmid, $deviceid);
3366         qemu_netdevdel($vmid, $deviceid);
3367
3368     } else {
3369         die "can't unplug device '$deviceid'\n";
3370     }
3371
3372     return 1;
3373 }
3374
3375 sub qemu_deviceadd {
3376     my ($vmid, $devicefull) = @_;
3377
3378     $devicefull = "driver=".$devicefull;
3379     my %options =  split(/[=,]/, $devicefull);
3380
3381     vm_mon_cmd($vmid, "device_add" , %options);
3382 }
3383
3384 sub qemu_devicedel {
3385     my ($vmid, $deviceid) = @_;
3386
3387     my $ret = vm_mon_cmd($vmid, "device_del", id => $deviceid);
3388 }
3389
3390 sub qemu_objectadd {
3391     my($vmid, $objectid, $qomtype) = @_;
3392
3393     vm_mon_cmd($vmid, "object-add", id => $objectid, "qom-type" => $qomtype);
3394
3395     return 1;
3396 }
3397
3398 sub qemu_objectdel {
3399     my($vmid, $objectid) = @_;
3400
3401     vm_mon_cmd($vmid, "object-del", id => $objectid);
3402
3403     return 1;
3404 }
3405
3406 sub qemu_driveadd {
3407     my ($storecfg, $vmid, $device) = @_;
3408
3409     my $drive = print_drive_full($storecfg, $vmid, $device);
3410     $drive =~ s/\\/\\\\/g;
3411     my $ret = vm_human_monitor_command($vmid, "drive_add auto \"$drive\"");
3412
3413     # If the command succeeds qemu prints: "OK"
3414     return 1 if $ret =~ m/OK/s;
3415
3416     die "adding drive failed: $ret\n";
3417 }
3418
3419 sub qemu_drivedel {
3420     my($vmid, $deviceid) = @_;
3421
3422     my $ret = vm_human_monitor_command($vmid, "drive_del drive-$deviceid");
3423     $ret =~ s/^\s+//;
3424
3425     return 1 if $ret eq "";
3426
3427     # NB: device not found errors mean the drive was auto-deleted and we ignore the error
3428     return 1 if $ret =~ m/Device \'.*?\' not found/s;
3429
3430     die "deleting drive $deviceid failed : $ret\n";
3431 }
3432
3433 sub qemu_deviceaddverify {
3434     my ($vmid, $deviceid) = @_;
3435
3436     for (my $i = 0; $i <= 5; $i++) {
3437          my $devices_list = vm_devices_list($vmid);
3438          return 1 if defined($devices_list->{$deviceid});
3439          sleep 1;
3440     }
3441
3442     die "error on hotplug device '$deviceid'\n";
3443 }
3444
3445
3446 sub qemu_devicedelverify {
3447     my ($vmid, $deviceid) = @_;
3448
3449     # need to verify that the device is correctly removed as device_del
3450     # is async and empty return is not reliable
3451
3452     for (my $i = 0; $i <= 5; $i++) {
3453          my $devices_list = vm_devices_list($vmid);
3454          return 1 if !defined($devices_list->{$deviceid});
3455          sleep 1;
3456     }
3457
3458     die "error on hot-unplugging device '$deviceid'\n";
3459 }
3460
3461 sub qemu_findorcreatescsihw {
3462     my ($storecfg, $conf, $vmid, $device) = @_;
3463
3464     my $maxdev = ($conf->{scsihw} && ($conf->{scsihw} !~ m/^lsi/)) ? 256 : 7;
3465     my $controller = int($device->{index} / $maxdev);
3466     my $scsihwid="scsihw$controller";
3467     my $devices_list = vm_devices_list($vmid);
3468
3469     if(!defined($devices_list->{$scsihwid})) {
3470         vm_deviceplug($storecfg, $conf, $vmid, $scsihwid);
3471     }
3472
3473     return 1;
3474 }
3475
3476 sub qemu_add_pci_bridge {
3477     my ($storecfg, $conf, $vmid, $device) = @_;
3478
3479     my $bridges = {};
3480
3481     my $bridgeid;
3482
3483     print_pci_addr($device, $bridges);
3484
3485     while (my ($k, $v) = each %$bridges) {
3486         $bridgeid = $k;
3487     }
3488     return 1 if !defined($bridgeid) || $bridgeid < 1;
3489
3490     my $bridge = "pci.$bridgeid";
3491     my $devices_list = vm_devices_list($vmid);
3492
3493     if (!defined($devices_list->{$bridge})) {
3494         vm_deviceplug($storecfg, $conf, $vmid, $bridge);
3495     }
3496
3497     return 1;
3498 }
3499
3500 sub qemu_set_link_status {
3501     my ($vmid, $device, $up) = @_;
3502
3503     vm_mon_cmd($vmid, "set_link", name => $device,
3504                up => $up ? JSON::true : JSON::false);
3505 }
3506
3507 sub qemu_netdevadd {
3508     my ($vmid, $conf, $device, $deviceid) = @_;
3509
3510     my $netdev = print_netdev_full($vmid, $conf, $device, $deviceid);
3511     my %options =  split(/[=,]/, $netdev);
3512
3513     vm_mon_cmd($vmid, "netdev_add",  %options);
3514     return 1;
3515 }
3516
3517 sub qemu_netdevdel {
3518     my ($vmid, $deviceid) = @_;
3519
3520     vm_mon_cmd($vmid, "netdev_del", id => $deviceid);
3521 }
3522
3523 sub qemu_cpu_hotplug {
3524     my ($vmid, $conf, $vcpus) = @_;
3525
3526     my $sockets = 1;
3527     $sockets = $conf->{smp} if $conf->{smp}; # old style - no longer iused
3528     $sockets = $conf->{sockets} if  $conf->{sockets};
3529     my $cores = $conf->{cores} || 1;
3530     my $maxcpus = $sockets * $cores;
3531
3532     $vcpus = $maxcpus if !$vcpus;
3533
3534     die "you can't add more vcpus than maxcpus\n"
3535         if $vcpus > $maxcpus;
3536
3537     my $currentvcpus = $conf->{vcpus} || $maxcpus;
3538     die "online cpu unplug is not yet possible\n"
3539         if $vcpus < $currentvcpus;
3540
3541     my $currentrunningvcpus = vm_mon_cmd($vmid, "query-cpus");
3542     die "vcpus in running vm is different than configuration\n"
3543         if scalar(@{$currentrunningvcpus}) != $currentvcpus;
3544
3545     for (my $i = $currentvcpus; $i < $vcpus; $i++) {
3546         vm_mon_cmd($vmid, "cpu-add", id => int($i));
3547     }
3548 }
3549
3550 sub qemu_memory_hotplug {
3551     my ($vmid, $conf, $defaults, $opt, $value) = @_;
3552
3553     return $value if !check_running($vmid);
3554
3555     my $memory = $conf->{memory} || $defaults->{memory};
3556     $value = $defaults->{memory} if !$value;
3557     return $value if $value == $memory;
3558
3559     my $static_memory = $STATICMEM;
3560     my $dimm_memory = $memory - $static_memory;
3561
3562     die "memory can't be lower than $static_memory MB" if $value < $static_memory;
3563     die "memory unplug is not yet available" if $value < $memory;
3564     die "you cannot add more memory than $MAX_MEM MB!\n" if $memory > $MAX_MEM;
3565
3566
3567     my $sockets = 1;
3568     $sockets = $conf->{sockets} if $conf->{sockets};
3569
3570     foreach_dimm($conf, $vmid, $value, $sockets, sub {
3571         my ($conf, $vmid, $name, $dimm_size, $numanode, $current_size, $memory) = @_;
3572
3573             return if $current_size <= $conf->{memory};
3574
3575             eval { vm_mon_cmd($vmid, "object-add", 'qom-type' => "memory-backend-ram", id => "mem-$name", props => { size => int($dimm_size*1024*1024) } ) };
3576             if (my $err = $@) {
3577                 eval { qemu_objectdel($vmid, "mem-$name"); };
3578                 die $err;
3579             }
3580
3581             eval { vm_mon_cmd($vmid, "device_add", driver => "pc-dimm", id => "$name", memdev => "mem-$name", node => $numanode) };
3582             if (my $err = $@) {
3583                 eval { qemu_objectdel($vmid, "mem-$name"); };
3584                 die $err;
3585             }
3586             #update conf after each succesful module hotplug
3587             $conf->{memory} = $current_size;
3588             update_config_nolock($vmid, $conf, 1);
3589     });
3590 }
3591
3592 sub qemu_block_set_io_throttle {
3593     my ($vmid, $deviceid, $bps, $bps_rd, $bps_wr, $iops, $iops_rd, $iops_wr) = @_;
3594
3595     return if !check_running($vmid) ;
3596
3597     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($iops_wr));
3598
3599 }
3600
3601 # old code, only used to shutdown old VM after update
3602 sub __read_avail {
3603     my ($fh, $timeout) = @_;
3604
3605     my $sel = new IO::Select;
3606     $sel->add($fh);
3607
3608     my $res = '';
3609     my $buf;
3610
3611     my @ready;
3612     while (scalar (@ready = $sel->can_read($timeout))) {
3613         my $count;
3614         if ($count = $fh->sysread($buf, 8192)) {
3615             if ($buf =~ /^(.*)\(qemu\) $/s) {
3616                 $res .= $1;
3617                 last;
3618             } else {
3619                 $res .= $buf;
3620             }
3621         } else {
3622             if (!defined($count)) {
3623                 die "$!\n";
3624             }
3625             last;
3626         }
3627     }
3628
3629     die "monitor read timeout\n" if !scalar(@ready);
3630
3631     return $res;
3632 }
3633
3634 # old code, only used to shutdown old VM after update
3635 sub vm_monitor_command {
3636     my ($vmid, $cmdstr, $nocheck) = @_;
3637
3638     my $res;
3639
3640     eval {
3641         die "VM $vmid not running\n" if !check_running($vmid, $nocheck);
3642
3643         my $sname = "${var_run_tmpdir}/$vmid.mon";
3644
3645         my $sock = IO::Socket::UNIX->new( Peer => $sname ) ||
3646             die "unable to connect to VM $vmid socket - $!\n";
3647
3648         my $timeout = 3;
3649
3650         # hack: migrate sometime blocks the monitor (when migrate_downtime
3651         # is set)
3652         if ($cmdstr =~ m/^(info\s+migrate|migrate\s)/) {
3653             $timeout = 60*60; # 1 hour
3654         }
3655
3656         # read banner;
3657         my $data = __read_avail($sock, $timeout);
3658
3659         if ($data !~ m/^QEMU\s+(\S+)\s+monitor\s/) {
3660             die "got unexpected qemu monitor banner\n";
3661         }
3662
3663         my $sel = new IO::Select;
3664         $sel->add($sock);
3665
3666         if (!scalar(my @ready = $sel->can_write($timeout))) {
3667             die "monitor write error - timeout";
3668         }
3669
3670         my $fullcmd = "$cmdstr\r";
3671
3672         # syslog('info', "VM $vmid monitor command: $cmdstr");
3673
3674         my $b;
3675         if (!($b = $sock->syswrite($fullcmd)) || ($b != length($fullcmd))) {
3676             die "monitor write error - $!";
3677         }
3678
3679         return if ($cmdstr eq 'q') || ($cmdstr eq 'quit');
3680
3681         $timeout = 20;
3682
3683         if ($cmdstr =~ m/^(info\s+migrate|migrate\s)/) {
3684             $timeout = 60*60; # 1 hour
3685         } elsif ($cmdstr =~ m/^(eject|change)/) {
3686             $timeout = 60; # note: cdrom mount command is slow
3687         }
3688         if ($res = __read_avail($sock, $timeout)) {
3689
3690             my @lines = split("\r?\n", $res);
3691
3692             shift @lines if $lines[0] !~ m/^unknown command/; # skip echo
3693
3694             $res = join("\n", @lines);
3695             $res .= "\n";
3696         }
3697     };
3698
3699     my $err = $@;
3700
3701     if ($err) {
3702         syslog("err", "VM $vmid monitor command failed - $err");
3703         die $err;
3704     }
3705
3706     return $res;
3707 }
3708
3709 sub qemu_block_resize {
3710     my ($vmid, $deviceid, $storecfg, $volid, $size) = @_;
3711
3712     my $running = check_running($vmid);
3713
3714     return if !PVE::Storage::volume_resize($storecfg, $volid, $size, $running);
3715
3716     return if !$running;
3717
3718     vm_mon_cmd($vmid, "block_resize", device => $deviceid, size => int($size));
3719
3720 }
3721
3722 sub qemu_volume_snapshot {
3723     my ($vmid, $deviceid, $storecfg, $volid, $snap) = @_;
3724
3725     my $running = check_running($vmid);
3726
3727     return if !PVE::Storage::volume_snapshot($storecfg, $volid, $snap, $running);
3728
3729     return if !$running;
3730
3731     vm_mon_cmd($vmid, "snapshot-drive", device => $deviceid, name => $snap);
3732
3733 }
3734
3735 sub qemu_volume_snapshot_delete {
3736     my ($vmid, $deviceid, $storecfg, $volid, $snap) = @_;
3737
3738     my $running = check_running($vmid);
3739
3740     return if !PVE::Storage::volume_snapshot_delete($storecfg, $volid, $snap, $running);
3741
3742     return if !$running;
3743
3744     vm_mon_cmd($vmid, "delete-drive-snapshot", device => $deviceid, name => $snap);
3745 }
3746
3747 sub set_migration_caps {
3748     my ($vmid) = @_;
3749
3750     my $cap_ref = [];
3751
3752     my $enabled_cap = {
3753         "auto-converge" => 1,
3754         "xbzrle" => 0,
3755         "x-rdma-pin-all" => 0,
3756         "zero-blocks" => 0,
3757     };
3758
3759     my $supported_capabilities = vm_mon_cmd_nocheck($vmid, "query-migrate-capabilities");
3760
3761     for my $supported_capability (@$supported_capabilities) {
3762         push @$cap_ref, {
3763             capability => $supported_capability->{capability},
3764             state => $enabled_cap->{$supported_capability->{capability}} ? JSON::true : JSON::false,
3765         };
3766     }
3767
3768     vm_mon_cmd_nocheck($vmid, "migrate-set-capabilities", capabilities => $cap_ref);
3769 }
3770
3771 my $fast_plug_option = {
3772     'lock' => 1,
3773     'name' => 1,
3774     'onboot' => 1,
3775     'shares' => 1,
3776     'startup' => 1,
3777 };
3778
3779 # hotplug changes in [PENDING]
3780 # $selection hash can be used to only apply specified options, for
3781 # example: { cores => 1 } (only apply changed 'cores')
3782 # $errors ref is used to return error messages
3783 sub vmconfig_hotplug_pending {
3784     my ($vmid, $conf, $storecfg, $selection, $errors) = @_;
3785
3786     my $defaults = load_defaults();
3787
3788     # commit values which do not have any impact on running VM first
3789     # Note: those option cannot raise errors, we we do not care about
3790     # $selection and always apply them.
3791
3792     my $add_error = sub {
3793         my ($opt, $msg) = @_;
3794         $errors->{$opt} = "hotplug problem - $msg";
3795     };
3796
3797     my $changes = 0;
3798     foreach my $opt (keys %{$conf->{pending}}) { # add/change
3799         if ($fast_plug_option->{$opt}) {
3800             $conf->{$opt} = $conf->{pending}->{$opt};
3801             delete $conf->{pending}->{$opt};
3802             $changes = 1;
3803         }
3804     }
3805
3806     if ($changes) {
3807         update_config_nolock($vmid, $conf, 1);
3808         $conf = load_config($vmid); # update/reload
3809     }
3810
3811     my $hotplug_features = parse_hotplug_features(defined($conf->{hotplug}) ? $conf->{hotplug} : '1');
3812
3813     my @delete = PVE::Tools::split_list($conf->{pending}->{delete});
3814     foreach my $opt (@delete) {
3815         next if $selection && !$selection->{$opt};
3816         eval {
3817             if ($opt eq 'hotplug') {
3818                 die "skip\n" if ($conf->{hotplug} =~ /memory/);
3819             } elsif ($opt eq 'tablet') {
3820                 die "skip\n" if !$hotplug_features->{usb};
3821                 if ($defaults->{tablet}) {
3822                     vm_deviceplug($storecfg, $conf, $vmid, $opt);
3823                 } else {
3824                     vm_deviceunplug($vmid, $conf, $opt);
3825                 }
3826             } elsif ($opt eq 'vcpus') {
3827                 die "skip\n" if !$hotplug_features->{cpu};
3828                 qemu_cpu_hotplug($vmid, $conf, undef);
3829             } elsif ($opt eq 'balloon') {
3830                 # enable balloon device is not hotpluggable
3831                 die "skip\n" if !defined($conf->{balloon}) || $conf->{balloon};
3832             } elsif ($fast_plug_option->{$opt}) {
3833                 # do nothing
3834             } elsif ($opt =~ m/^net(\d+)$/) {
3835                 die "skip\n" if !$hotplug_features->{network};
3836                 vm_deviceunplug($vmid, $conf, $opt);
3837             } elsif (valid_drivename($opt)) {
3838                 die "skip\n" if !$hotplug_features->{disk} || $opt =~ m/(ide|sata)(\d+)/;
3839                 vm_deviceunplug($vmid, $conf, $opt);
3840                 vmconfig_register_unused_drive($storecfg, $vmid, $conf, parse_drive($opt, $conf->{$opt}));
3841             } elsif ($opt =~ m/^memory$/) {
3842                 die "skip\n" if !$hotplug_features->{memory};
3843                 qemu_memory_hotplug($vmid, $conf, $defaults, $opt);
3844             } else {
3845                 die "skip\n";
3846             }
3847         };
3848         if (my $err = $@) {
3849             &$add_error($opt, $err) if $err ne "skip\n";
3850         } else {
3851             # save new config if hotplug was successful
3852             delete $conf->{$opt};
3853             vmconfig_undelete_pending_option($conf, $opt);
3854             update_config_nolock($vmid, $conf, 1);
3855             $conf = load_config($vmid); # update/reload
3856         }
3857     }
3858
3859     foreach my $opt (keys %{$conf->{pending}}) {
3860         next if $selection && !$selection->{$opt};
3861         my $value = $conf->{pending}->{$opt};
3862         eval {
3863             if ($opt eq 'hotplug') {
3864                 die "skip\n" if ($value =~ /memory/) || ($value !~ /memory/ && $conf->{hotplug} =~ /memory/);
3865             } elsif ($opt eq 'tablet') {
3866                 die "skip\n" if !$hotplug_features->{usb};
3867                 if ($value == 1) {
3868                     vm_deviceplug($storecfg, $conf, $vmid, $opt);
3869                 } elsif ($value == 0) {
3870                     vm_deviceunplug($vmid, $conf, $opt);
3871                 }
3872             } elsif ($opt eq 'vcpus') {
3873                 die "skip\n" if !$hotplug_features->{cpu};
3874                 qemu_cpu_hotplug($vmid, $conf, $value);
3875             } elsif ($opt eq 'balloon') {
3876                 # enable/disable balloning device is not hotpluggable
3877                 my $old_balloon_enabled =  !!(!defined($conf->{balloon}) || $conf->{balloon});
3878                 my $new_balloon_enabled =  !!(!defined($conf->{pending}->{balloon}) || $conf->{pending}->{balloon});
3879                 die "skip\n" if $old_balloon_enabled != $new_balloon_enabled;
3880
3881                 # allow manual ballooning if shares is set to zero
3882                 if ((defined($conf->{shares}) && ($conf->{shares} == 0))) {
3883                     my $balloon = $conf->{pending}->{balloon} || $conf->{memory} || $defaults->{memory};
3884                     vm_mon_cmd($vmid, "balloon", value => $balloon*1024*1024);
3885                 }
3886             } elsif ($opt =~ m/^net(\d+)$/) {
3887                 # some changes can be done without hotplug
3888                 vmconfig_update_net($storecfg, $conf, $hotplug_features->{network},
3889                                     $vmid, $opt, $value);
3890             } elsif (valid_drivename($opt)) {
3891                 # some changes can be done without hotplug
3892                 vmconfig_update_disk($storecfg, $conf, $hotplug_features->{disk},
3893                                      $vmid, $opt, $value, 1);
3894             } elsif ($opt =~ m/^memory$/) { #dimms
3895                 die "skip\n" if !$hotplug_features->{memory};
3896                 $value = qemu_memory_hotplug($vmid, $conf, $defaults, $opt, $value);
3897             } else {
3898                 die "skip\n";  # skip non-hot-pluggable options
3899             }
3900         };
3901         if (my $err = $@) {
3902             &$add_error($opt, $err) if $err ne "skip\n";
3903         } else {
3904             # save new config if hotplug was successful
3905             $conf->{$opt} = $value;
3906             delete $conf->{pending}->{$opt};
3907             update_config_nolock($vmid, $conf, 1);
3908             $conf = load_config($vmid); # update/reload
3909         }
3910     }
3911 }
3912
3913 sub vmconfig_apply_pending {
3914     my ($vmid, $conf, $storecfg) = @_;
3915
3916     # cold plug
3917
3918     my @delete = PVE::Tools::split_list($conf->{pending}->{delete});
3919     foreach my $opt (@delete) { # delete
3920         die "internal error" if $opt =~ m/^unused/;
3921         $conf = load_config($vmid); # update/reload
3922         if (!defined($conf->{$opt})) {
3923             vmconfig_undelete_pending_option($conf, $opt);
3924             update_config_nolock($vmid, $conf, 1);
3925         } elsif (valid_drivename($opt)) {
3926             vmconfig_register_unused_drive($storecfg, $vmid, $conf, parse_drive($opt, $conf->{$opt}));
3927             vmconfig_undelete_pending_option($conf, $opt);
3928             delete $conf->{$opt};
3929             update_config_nolock($vmid, $conf, 1);
3930         } else {
3931             vmconfig_undelete_pending_option($conf, $opt);
3932             delete $conf->{$opt};
3933             update_config_nolock($vmid, $conf, 1);
3934         }
3935     }
3936
3937     $conf = load_config($vmid); # update/reload
3938
3939     foreach my $opt (keys %{$conf->{pending}}) { # add/change
3940         $conf = load_config($vmid); # update/reload
3941
3942         if (defined($conf->{$opt}) && ($conf->{$opt} eq $conf->{pending}->{$opt})) {
3943             # skip if nothing changed
3944         } elsif (valid_drivename($opt)) {
3945             vmconfig_register_unused_drive($storecfg, $vmid, $conf, parse_drive($opt, $conf->{$opt}))
3946                 if defined($conf->{$opt});
3947             $conf->{$opt} = $conf->{pending}->{$opt};
3948         } else {
3949             $conf->{$opt} = $conf->{pending}->{$opt};
3950         }
3951
3952         delete $conf->{pending}->{$opt};
3953         update_config_nolock($vmid, $conf, 1);
3954     }
3955 }
3956
3957 my $safe_num_ne = sub {
3958     my ($a, $b) = @_;
3959
3960     return 0 if !defined($a) && !defined($b);
3961     return 1 if !defined($a);
3962     return 1 if !defined($b);
3963
3964     return $a != $b;
3965 };
3966
3967 my $safe_string_ne = sub {
3968     my ($a, $b) = @_;
3969
3970     return 0 if !defined($a) && !defined($b);
3971     return 1 if !defined($a);
3972     return 1 if !defined($b);
3973
3974     return $a ne $b;
3975 };
3976
3977 sub vmconfig_update_net {
3978     my ($storecfg, $conf, $hotplug, $vmid, $opt, $value) = @_;
3979
3980     my $newnet = parse_net($value);
3981
3982     if ($conf->{$opt}) {
3983         my $oldnet = parse_net($conf->{$opt});
3984
3985         if (&$safe_string_ne($oldnet->{model}, $newnet->{model}) ||
3986             &$safe_string_ne($oldnet->{macaddr}, $newnet->{macaddr}) ||
3987             &$safe_num_ne($oldnet->{queues}, $newnet->{queues}) ||
3988             !($newnet->{bridge} && $oldnet->{bridge})) { # bridge/nat mode change
3989
3990             # for non online change, we try to hot-unplug
3991             die "skip\n" if !$hotplug;
3992             vm_deviceunplug($vmid, $conf, $opt);
3993         } else {
3994
3995             die "internal error" if $opt !~ m/net(\d+)/;
3996             my $iface = "tap${vmid}i$1";
3997
3998             if (&$safe_num_ne($oldnet->{rate}, $newnet->{rate})) {
3999                 PVE::Network::tap_rate_limit($iface, $newnet->{rate});
4000             }
4001
4002             if (&$safe_string_ne($oldnet->{bridge}, $newnet->{bridge}) ||
4003                 &$safe_num_ne($oldnet->{tag}, $newnet->{tag}) ||
4004                 &$safe_num_ne($oldnet->{firewall}, $newnet->{firewall})) {
4005                 PVE::Network::tap_unplug($iface);
4006                 PVE::Network::tap_plug($iface, $newnet->{bridge}, $newnet->{tag}, $newnet->{firewall});
4007             }
4008
4009             if (&$safe_string_ne($oldnet->{link_down}, $newnet->{link_down})) {
4010                 qemu_set_link_status($vmid, $opt, !$newnet->{link_down});
4011             }
4012
4013             return 1;
4014         }
4015     }
4016
4017     if ($hotplug) {
4018         vm_deviceplug($storecfg, $conf, $vmid, $opt, $newnet);
4019     } else {
4020         die "skip\n";
4021     }
4022 }
4023
4024 sub vmconfig_update_disk {
4025     my ($storecfg, $conf, $hotplug, $vmid, $opt, $value, $force) = @_;
4026
4027     # fixme: do we need force?
4028
4029     my $drive = parse_drive($opt, $value);
4030
4031     if ($conf->{$opt}) {
4032
4033         if (my $old_drive = parse_drive($opt, $conf->{$opt}))  {
4034
4035             my $media = $drive->{media} || 'disk';
4036             my $oldmedia = $old_drive->{media} || 'disk';
4037             die "unable to change media type\n" if $media ne $oldmedia;
4038
4039             if (!drive_is_cdrom($old_drive)) {
4040
4041                 if ($drive->{file} ne $old_drive->{file}) {
4042
4043                     die "skip\n" if !$hotplug;
4044
4045                     # unplug and register as unused
4046                     vm_deviceunplug($vmid, $conf, $opt);
4047                     vmconfig_register_unused_drive($storecfg, $vmid, $conf, $old_drive)
4048
4049                 } else {
4050                     # update existing disk
4051
4052                     # skip non hotpluggable value
4053                     if (&$safe_num_ne($drive->{discard}, $old_drive->{discard}) ||
4054                         &$safe_string_ne($drive->{cache}, $old_drive->{cache})) {
4055                         die "skip\n";
4056                     }
4057
4058                     # apply throttle
4059                     if (&$safe_num_ne($drive->{mbps}, $old_drive->{mbps}) ||
4060                         &$safe_num_ne($drive->{mbps_rd}, $old_drive->{mbps_rd}) ||
4061                         &$safe_num_ne($drive->{mbps_wr}, $old_drive->{mbps_wr}) ||
4062                         &$safe_num_ne($drive->{iops}, $old_drive->{iops}) ||
4063                         &$safe_num_ne($drive->{iops_rd}, $old_drive->{iops_rd}) ||
4064                         &$safe_num_ne($drive->{iops_wr}, $old_drive->{iops_wr}) ||
4065                         &$safe_num_ne($drive->{mbps_max}, $old_drive->{mbps_max}) ||
4066                         &$safe_num_ne($drive->{mbps_rd_max}, $old_drive->{mbps_rd_max}) ||
4067                         &$safe_num_ne($drive->{mbps_wr_max}, $old_drive->{mbps_wr_max}) ||
4068                         &$safe_num_ne($drive->{iops_max}, $old_drive->{iops_max}) ||
4069                         &$safe_num_ne($drive->{iops_rd_max}, $old_drive->{iops_rd_max}) ||
4070                         &$safe_num_ne($drive->{iops_wr_max}, $old_drive->{iops_wr_max})) {
4071
4072                         qemu_block_set_io_throttle($vmid,"drive-$opt",
4073                                                    ($drive->{mbps} || 0)*1024*1024,
4074                                                    ($drive->{mbps_rd} || 0)*1024*1024,
4075                                                    ($drive->{mbps_wr} || 0)*1024*1024,
4076                                                    $drive->{iops} || 0,
4077                                                    $drive->{iops_rd} || 0,
4078                                                    $drive->{iops_wr} || 0,
4079                                                    ($drive->{mbps_max} || 0)*1024*1024,
4080                                                    ($drive->{mbps_rd_max} || 0)*1024*1024,
4081                                                    ($drive->{mbps_wr_max} || 0)*1024*1024,
4082                                                    $drive->{iops_max} || 0,
4083                                                    $drive->{iops_rd_max} || 0,
4084                                                    $drive->{iops_wr_max} || 0);
4085
4086                     }
4087
4088                     return 1;
4089                 }
4090
4091             } else { # cdrom
4092
4093                 if ($drive->{file} eq 'none') {
4094                     vm_mon_cmd($vmid, "eject",force => JSON::true,device => "drive-$opt");
4095                 } else {
4096                     my $path = get_iso_path($storecfg, $vmid, $drive->{file});
4097                     vm_mon_cmd($vmid, "eject", force => JSON::true,device => "drive-$opt"); # force eject if locked
4098                     vm_mon_cmd($vmid, "change", device => "drive-$opt",target => "$path") if $path;
4099                 }
4100
4101                 return 1;
4102             }
4103         }
4104     }
4105
4106     die "skip\n" if !$hotplug || $opt =~ m/(ide|sata)(\d+)/;
4107     # hotplug new disks
4108     vm_deviceplug($storecfg, $conf, $vmid, $opt, $drive);
4109 }
4110
4111 sub vm_start {
4112     my ($storecfg, $vmid, $statefile, $skiplock, $migratedfrom, $paused, $forcemachine, $spice_ticket) = @_;
4113
4114     lock_config($vmid, sub {
4115         my $conf = load_config($vmid, $migratedfrom);
4116
4117         die "you can't start a vm if it's a template\n" if is_template($conf);
4118
4119         check_lock($conf) if !$skiplock;
4120
4121         die "VM $vmid already running\n" if check_running($vmid, undef, $migratedfrom);
4122
4123         if (!$statefile && scalar(keys %{$conf->{pending}})) {
4124             vmconfig_apply_pending($vmid, $conf, $storecfg);
4125             $conf = load_config($vmid); # update/reload
4126         }
4127
4128         my $defaults = load_defaults();
4129
4130         # set environment variable useful inside network script
4131         $ENV{PVE_MIGRATED_FROM} = $migratedfrom if $migratedfrom;
4132
4133         my ($cmd, $vollist, $spice_port) = config_to_command($storecfg, $vmid, $conf, $defaults, $forcemachine);
4134
4135         my $migrate_port = 0;
4136         my $migrate_uri;
4137         if ($statefile) {
4138             if ($statefile eq 'tcp') {
4139                 my $localip = "localhost";
4140                 my $datacenterconf = PVE::Cluster::cfs_read_file('datacenter.cfg');
4141                 if ($datacenterconf->{migration_unsecure}) {
4142                         my $nodename = PVE::INotify::nodename();
4143                         $localip = PVE::Cluster::remote_node_ip($nodename, 1);
4144                 }
4145                 $migrate_port = PVE::Tools::next_migrate_port();
4146                 $migrate_uri = "tcp:${localip}:${migrate_port}";
4147                 push @$cmd, '-incoming', $migrate_uri;
4148                 push @$cmd, '-S';
4149             } else {
4150                 push @$cmd, '-loadstate', $statefile;
4151             }
4152         } elsif ($paused) {
4153             push @$cmd, '-S';
4154         }
4155
4156         # host pci devices
4157         for (my $i = 0; $i < $MAX_HOSTPCI_DEVICES; $i++)  {
4158           my $d = parse_hostpci($conf->{"hostpci$i"});
4159           next if !$d;
4160           my $pcidevices = $d->{pciid};
4161           foreach my $pcidevice (@$pcidevices) {
4162                 my $pciid = $pcidevice->{id}.".".$pcidevice->{function};
4163
4164                 my $info = pci_device_info("0000:$pciid");
4165                 die "IOMMU not present\n" if !check_iommu_support();
4166                 die "no pci device info for device '$pciid'\n" if !$info;
4167
4168                 if ($d->{driver} && $d->{driver} eq "vfio") {
4169                     die "can't unbind/bind pci group to vfio '$pciid'\n" if !pci_dev_group_bind_to_vfio($pciid);
4170                 } else {
4171                     die "can't unbind/bind to stub pci device '$pciid'\n" if !pci_dev_bind_to_stub($info);
4172                 }
4173
4174                 die "can't reset pci device '$pciid'\n" if $info->{has_fl_reset} and !pci_dev_reset($info);
4175           }
4176         }
4177
4178         PVE::Storage::activate_volumes($storecfg, $vollist);
4179
4180         eval  { run_command($cmd, timeout => $statefile ? undef : 30,
4181                     umask => 0077); };
4182         my $err = $@;
4183         die "start failed: $err" if $err;
4184
4185         print "migration listens on $migrate_uri\n" if $migrate_uri;
4186
4187         if ($statefile && $statefile ne 'tcp')  {
4188             eval { vm_mon_cmd_nocheck($vmid, "cont"); };
4189             warn $@ if $@;
4190         }
4191
4192         if ($migratedfrom) {
4193
4194             eval {
4195                 set_migration_caps($vmid);
4196             };
4197             warn $@ if $@;
4198
4199             if ($spice_port) {
4200                 print "spice listens on port $spice_port\n";
4201                 if ($spice_ticket) {
4202                     vm_mon_cmd_nocheck($vmid, "set_password", protocol => 'spice', password => $spice_ticket);
4203                     vm_mon_cmd_nocheck($vmid, "expire_password", protocol => 'spice', time => "+30");
4204                 }
4205             }
4206
4207         } else {
4208
4209             if (!$statefile && (!defined($conf->{balloon}) || $conf->{balloon})) {
4210                 vm_mon_cmd_nocheck($vmid, "balloon", value => $conf->{balloon}*1024*1024)
4211                     if $conf->{balloon};
4212             }
4213
4214             foreach my $opt (keys %$conf) {
4215                 next if $opt !~  m/^net\d+$/;
4216                 my $nicconf = parse_net($conf->{$opt});
4217                 qemu_set_link_status($vmid, $opt, 0) if $nicconf->{link_down};
4218             }
4219         }
4220
4221         vm_mon_cmd_nocheck($vmid, 'qom-set',
4222                     path => "machine/peripheral/balloon0",
4223                     property => "guest-stats-polling-interval",
4224                     value => 2) if (!defined($conf->{balloon}) || $conf->{balloon});
4225
4226     });
4227 }
4228
4229 sub vm_mon_cmd {
4230     my ($vmid, $execute, %params) = @_;
4231
4232     my $cmd = { execute => $execute, arguments => \%params };
4233     vm_qmp_command($vmid, $cmd);
4234 }
4235
4236 sub vm_mon_cmd_nocheck {
4237     my ($vmid, $execute, %params) = @_;
4238
4239     my $cmd = { execute => $execute, arguments => \%params };
4240     vm_qmp_command($vmid, $cmd, 1);
4241 }
4242
4243 sub vm_qmp_command {
4244     my ($vmid, $cmd, $nocheck) = @_;
4245
4246     my $res;
4247
4248     my $timeout;
4249     if ($cmd->{arguments} && $cmd->{arguments}->{timeout}) {
4250         $timeout = $cmd->{arguments}->{timeout};
4251         delete $cmd->{arguments}->{timeout};
4252     }
4253
4254     eval {
4255         die "VM $vmid not running\n" if !check_running($vmid, $nocheck);
4256         my $sname = qmp_socket($vmid);
4257         if (-e $sname) { # test if VM is reasonambe new and supports qmp/qga
4258             my $qmpclient = PVE::QMPClient->new();
4259
4260             $res = $qmpclient->cmd($vmid, $cmd, $timeout);
4261         } elsif (-e "${var_run_tmpdir}/$vmid.mon") {
4262             die "can't execute complex command on old monitor - stop/start your vm to fix the problem\n"
4263                 if scalar(%{$cmd->{arguments}});
4264             vm_monitor_command($vmid, $cmd->{execute}, $nocheck);
4265         } else {
4266             die "unable to open monitor socket\n";
4267         }
4268     };
4269     if (my $err = $@) {
4270         syslog("err", "VM $vmid qmp command failed - $err");
4271         die $err;
4272     }
4273
4274     return $res;
4275 }
4276
4277 sub vm_human_monitor_command {
4278     my ($vmid, $cmdline) = @_;
4279
4280     my $res;
4281
4282     my $cmd = {
4283         execute => 'human-monitor-command',
4284         arguments => { 'command-line' => $cmdline},
4285     };
4286
4287     return vm_qmp_command($vmid, $cmd);
4288 }
4289
4290 sub vm_commandline {
4291     my ($storecfg, $vmid) = @_;
4292
4293     my $conf = load_config($vmid);
4294
4295     my $defaults = load_defaults();
4296
4297     my $cmd = config_to_command($storecfg, $vmid, $conf, $defaults);
4298
4299     return join(' ', @$cmd);
4300 }
4301
4302 sub vm_reset {
4303     my ($vmid, $skiplock) = @_;
4304
4305     lock_config($vmid, sub {
4306
4307         my $conf = load_config($vmid);
4308
4309         check_lock($conf) if !$skiplock;
4310
4311         vm_mon_cmd($vmid, "system_reset");
4312     });
4313 }
4314
4315 sub get_vm_volumes {
4316     my ($conf) = @_;
4317
4318     my $vollist = [];
4319     foreach_volid($conf, sub {
4320         my ($volid, $is_cdrom) = @_;
4321
4322         return if $volid =~ m|^/|;
4323
4324         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
4325         return if !$sid;
4326
4327         push @$vollist, $volid;
4328     });
4329
4330     return $vollist;
4331 }
4332
4333 sub vm_stop_cleanup {
4334     my ($storecfg, $vmid, $conf, $keepActive, $apply_pending_changes) = @_;
4335
4336     eval {
4337         fairsched_rmnod($vmid); # try to destroy group
4338
4339         if (!$keepActive) {
4340             my $vollist = get_vm_volumes($conf);
4341             PVE::Storage::deactivate_volumes($storecfg, $vollist);
4342         }
4343
4344         foreach my $ext (qw(mon qmp pid vnc qga)) {
4345             unlink "/var/run/qemu-server/${vmid}.$ext";
4346         }
4347
4348         vmconfig_apply_pending($vmid, $conf, $storecfg) if $apply_pending_changes;
4349     };
4350     warn $@ if $@; # avoid errors - just warn
4351 }
4352
4353 # Note: use $nockeck to skip tests if VM configuration file exists.
4354 # We need that when migration VMs to other nodes (files already moved)
4355 # Note: we set $keepActive in vzdump stop mode - volumes need to stay active
4356 sub vm_stop {
4357     my ($storecfg, $vmid, $skiplock, $nocheck, $timeout, $shutdown, $force, $keepActive, $migratedfrom) = @_;
4358
4359     $force = 1 if !defined($force) && !$shutdown;
4360
4361     if ($migratedfrom){
4362         my $pid = check_running($vmid, $nocheck, $migratedfrom);
4363         kill 15, $pid if $pid;
4364         my $conf = load_config($vmid, $migratedfrom);
4365         vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive, 0);
4366         return;
4367     }
4368
4369     lock_config($vmid, sub {
4370
4371         my $pid = check_running($vmid, $nocheck);
4372         return if !$pid;
4373
4374         my $conf;
4375         if (!$nocheck) {
4376             $conf = load_config($vmid);
4377             check_lock($conf) if !$skiplock;
4378             if (!defined($timeout) && $shutdown && $conf->{startup}) {
4379                 my $opts = parse_startup($conf->{startup});
4380                 $timeout = $opts->{down} if $opts->{down};
4381             }
4382         }
4383
4384         $timeout = 60 if !defined($timeout);
4385
4386         eval {
4387             if ($shutdown) {
4388                 if (defined($conf) && $conf->{agent}) {
4389                     vm_qmp_command($vmid, { execute => "guest-shutdown" }, $nocheck);
4390                 } else {
4391                     vm_qmp_command($vmid, { execute => "system_powerdown" }, $nocheck);
4392                 }
4393             } else {
4394                 vm_qmp_command($vmid, { execute => "quit" }, $nocheck);
4395             }
4396         };
4397         my $err = $@;
4398
4399         if (!$err) {
4400             my $count = 0;
4401             while (($count < $timeout) && check_running($vmid, $nocheck)) {
4402                 $count++;
4403                 sleep 1;
4404             }
4405
4406             if ($count >= $timeout) {
4407                 if ($force) {
4408                     warn "VM still running - terminating now with SIGTERM\n";
4409                     kill 15, $pid;
4410                 } else {
4411                     die "VM quit/powerdown failed - got timeout\n";
4412                 }
4413             } else {
4414                 vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive, 1) if $conf;
4415                 return;
4416             }
4417         } else {
4418             if ($force) {
4419                 warn "VM quit/powerdown failed - terminating now with SIGTERM\n";
4420                 kill 15, $pid;
4421             } else {
4422                 die "VM quit/powerdown failed\n";
4423             }
4424         }
4425
4426         # wait again
4427         $timeout = 10;
4428
4429         my $count = 0;
4430         while (($count < $timeout) && check_running($vmid, $nocheck)) {
4431             $count++;
4432             sleep 1;
4433         }
4434
4435         if ($count >= $timeout) {
4436             warn "VM still running - terminating now with SIGKILL\n";
4437             kill 9, $pid;
4438             sleep 1;
4439         }
4440
4441         vm_stop_cleanup($storecfg, $vmid, $conf, $keepActive, 1) if $conf;
4442    });
4443 }
4444
4445 sub vm_suspend {
4446     my ($vmid, $skiplock) = @_;
4447
4448     lock_config($vmid, sub {
4449
4450         my $conf = load_config($vmid);
4451
4452         check_lock($conf) if !($skiplock || ($conf->{lock} && $conf->{lock} eq 'backup'));
4453
4454         vm_mon_cmd($vmid, "stop");
4455     });
4456 }
4457
4458 sub vm_resume {
4459     my ($vmid, $skiplock) = @_;
4460
4461     lock_config($vmid, sub {
4462
4463         my $conf = load_config($vmid);
4464
4465         check_lock($conf) if !($skiplock || ($conf->{lock} && $conf->{lock} eq 'backup'));
4466
4467         vm_mon_cmd($vmid, "cont");
4468     });
4469 }
4470
4471 sub vm_sendkey {
4472     my ($vmid, $skiplock, $key) = @_;
4473
4474     lock_config($vmid, sub {
4475
4476         my $conf = load_config($vmid);
4477
4478         # there is no qmp command, so we use the human monitor command
4479         vm_human_monitor_command($vmid, "sendkey $key");
4480     });
4481 }
4482
4483 sub vm_destroy {
4484     my ($storecfg, $vmid, $skiplock) = @_;
4485
4486     lock_config($vmid, sub {
4487
4488         my $conf = load_config($vmid);
4489
4490         check_lock($conf) if !$skiplock;
4491
4492         if (!check_running($vmid)) {
4493             fairsched_rmnod($vmid); # try to destroy group
4494             destroy_vm($storecfg, $vmid);
4495         } else {
4496             die "VM $vmid is running - destroy failed\n";
4497         }
4498     });
4499 }
4500
4501 # pci helpers
4502
4503 sub file_write {
4504     my ($filename, $buf) = @_;
4505
4506     my $fh = IO::File->new($filename, "w");
4507     return undef if !$fh;
4508
4509     my $res = print $fh $buf;
4510
4511     $fh->close();
4512
4513     return $res;
4514 }
4515
4516 sub pci_device_info {
4517     my ($name) = @_;
4518
4519     my $res;
4520
4521     return undef if $name !~ m/^([a-f0-9]{4}):([a-f0-9]{2}):([a-f0-9]{2})\.([a-f0-9])$/;
4522     my ($domain, $bus, $slot, $func) = ($1, $2, $3, $4);
4523
4524     my $irq = file_read_firstline("$pcisysfs/devices/$name/irq");
4525     return undef if !defined($irq) || $irq !~ m/^\d+$/;
4526
4527     my $vendor = file_read_firstline("$pcisysfs/devices/$name/vendor");
4528     return undef if !defined($vendor) || $vendor !~ s/^0x//;
4529
4530     my $product = file_read_firstline("$pcisysfs/devices/$name/device");
4531     return undef if !defined($product) || $product !~ s/^0x//;
4532
4533     $res = {
4534         name => $name,
4535         vendor => $vendor,
4536         product => $product,
4537         domain => $domain,
4538         bus => $bus,
4539         slot => $slot,
4540         func => $func,
4541         irq => $irq,
4542         has_fl_reset => -f "$pcisysfs/devices/$name/reset" || 0,
4543     };
4544
4545     return $res;
4546 }
4547
4548 sub pci_dev_reset {
4549     my ($dev) = @_;
4550
4551     my $name = $dev->{name};
4552
4553     my $fn = "$pcisysfs/devices/$name/reset";
4554
4555     return file_write($fn, "1");
4556 }
4557
4558 sub pci_dev_bind_to_stub {
4559     my ($dev) = @_;
4560
4561     my $name = $dev->{name};
4562
4563     my $testdir = "$pcisysfs/drivers/pci-stub/$name";
4564     return 1 if -d $testdir;
4565
4566     my $data = "$dev->{vendor} $dev->{product}";
4567     return undef if !file_write("$pcisysfs/drivers/pci-stub/new_id", $data);
4568
4569     my $fn = "$pcisysfs/devices/$name/driver/unbind";
4570     if (!file_write($fn, $name)) {
4571         return undef if -f $fn;
4572     }
4573
4574     $fn = "$pcisysfs/drivers/pci-stub/bind";
4575     if (! -d $testdir) {
4576         return undef if !file_write($fn, $name);
4577     }
4578
4579     return -d $testdir;
4580 }
4581
4582 sub pci_dev_bind_to_vfio {
4583     my ($dev) = @_;
4584
4585     my $name = $dev->{name};
4586
4587     my $vfio_basedir = "$pcisysfs/drivers/vfio-pci";
4588
4589     if (!-d $vfio_basedir) {
4590         system("/sbin/modprobe vfio-pci >/dev/null 2>/dev/null");
4591     }
4592     die "Cannot find vfio-pci module!\n" if !-d $vfio_basedir;
4593
4594     my $testdir = "$vfio_basedir/$name";
4595     return 1 if -d $testdir;
4596
4597     my $data = "$dev->{vendor} $dev->{product}";
4598     return undef if !file_write("$vfio_basedir/new_id", $data);
4599
4600     my $fn = "$pcisysfs/devices/$name/driver/unbind";
4601     if (!file_write($fn, $name)) {
4602         return undef if -f $fn;
4603     }
4604
4605     $fn = "$vfio_basedir/bind";
4606     if (! -d $testdir) {
4607         return undef if !file_write($fn, $name);
4608     }
4609
4610     return -d $testdir;
4611 }
4612
4613 sub pci_dev_group_bind_to_vfio {
4614     my ($pciid) = @_;
4615
4616     my $vfio_basedir = "$pcisysfs/drivers/vfio-pci";
4617
4618     if (!-d $vfio_basedir) {
4619         system("/sbin/modprobe vfio-pci >/dev/null 2>/dev/null");
4620     }
4621     die "Cannot find vfio-pci module!\n" if !-d $vfio_basedir;
4622
4623     # get IOMMU group devices
4624     opendir(my $D, "$pcisysfs/devices/0000:$pciid/iommu_group/devices/") || die "Cannot open iommu_group: $!\n";
4625       my @devs = grep /^0000:/, readdir($D);
4626     closedir($D);
4627
4628     foreach my $pciid (@devs) {
4629         $pciid =~ m/^([:\.\da-f]+)$/ or die "PCI ID $pciid not valid!\n";
4630
4631         # pci bridges, switches or root ports are not supported
4632         # they have a pci_bus subdirectory so skip them
4633         next if (-e "$pcisysfs/devices/$pciid/pci_bus");
4634
4635         my $info = pci_device_info($1);
4636         pci_dev_bind_to_vfio($info) || die "Cannot bind $pciid to vfio\n";
4637     }
4638
4639     return 1;
4640 }
4641
4642 sub print_pci_addr {
4643     my ($id, $bridges) = @_;
4644
4645     my $res = '';
4646     my $devices = {
4647         piix3 => { bus => 0, addr => 1 },
4648         #addr2 : first videocard
4649         balloon0 => { bus => 0, addr => 3 },
4650         watchdog => { bus => 0, addr => 4 },
4651         scsihw0 => { bus => 0, addr => 5 },
4652         scsihw1 => { bus => 0, addr => 6 },
4653         ahci0 => { bus => 0, addr => 7 },
4654         qga0 => { bus => 0, addr => 8 },
4655         spice => { bus => 0, addr => 9 },
4656         virtio0 => { bus => 0, addr => 10 },
4657         virtio1 => { bus => 0, addr => 11 },
4658         virtio2 => { bus => 0, addr => 12 },
4659         virtio3 => { bus => 0, addr => 13 },
4660         virtio4 => { bus => 0, addr => 14 },
4661         virtio5 => { bus => 0, addr => 15 },
4662         hostpci0 => { bus => 0, addr => 16 },
4663         hostpci1 => { bus => 0, addr => 17 },
4664         net0 => { bus => 0, addr => 18 },
4665         net1 => { bus => 0, addr => 19 },
4666         net2 => { bus => 0, addr => 20 },
4667         net3 => { bus => 0, addr => 21 },
4668         net4 => { bus => 0, addr => 22 },
4669         net5 => { bus => 0, addr => 23 },
4670         vga1 => { bus => 0, addr => 24 },
4671         vga2 => { bus => 0, addr => 25 },
4672         vga3 => { bus => 0, addr => 26 },
4673         hostpci2 => { bus => 0, addr => 27 },
4674         hostpci3 => { bus => 0, addr => 28 },
4675         #addr29 : usb-host (pve-usb.cfg)
4676         'pci.1' => { bus => 0, addr => 30 },
4677         'pci.2' => { bus => 0, addr => 31 },
4678         'net6' => { bus => 1, addr => 1 },
4679         'net7' => { bus => 1, addr => 2 },
4680         'net8' => { bus => 1, addr => 3 },
4681         'net9' => { bus => 1, addr => 4 },
4682         'net10' => { bus => 1, addr => 5 },
4683         'net11' => { bus => 1, addr => 6 },
4684         'net12' => { bus => 1, addr => 7 },
4685         'net13' => { bus => 1, addr => 8 },
4686         'net14' => { bus => 1, addr => 9 },
4687         'net15' => { bus => 1, addr => 10 },
4688         'net16' => { bus => 1, addr => 11 },
4689         'net17' => { bus => 1, addr => 12 },
4690         'net18' => { bus => 1, addr => 13 },
4691         'net19' => { bus => 1, addr => 14 },
4692         'net20' => { bus => 1, addr => 15 },
4693         'net21' => { bus => 1, addr => 16 },
4694         'net22' => { bus => 1, addr => 17 },
4695         'net23' => { bus => 1, addr => 18 },
4696         'net24' => { bus => 1, addr => 19 },
4697         'net25' => { bus => 1, addr => 20 },
4698         'net26' => { bus => 1, addr => 21 },
4699         'net27' => { bus => 1, addr => 22 },
4700         'net28' => { bus => 1, addr => 23 },
4701         'net29' => { bus => 1, addr => 24 },
4702         'net30' => { bus => 1, addr => 25 },
4703         'net31' => { bus => 1, addr => 26 },
4704         'virtio6' => { bus => 2, addr => 1 },
4705         'virtio7' => { bus => 2, addr => 2 },
4706         'virtio8' => { bus => 2, addr => 3 },
4707         'virtio9' => { bus => 2, addr => 4 },
4708         'virtio10' => { bus => 2, addr => 5 },
4709         'virtio11' => { bus => 2, addr => 6 },
4710         'virtio12' => { bus => 2, addr => 7 },
4711         'virtio13' => { bus => 2, addr => 8 },
4712         'virtio14' => { bus => 2, addr => 9 },
4713         'virtio15' => { bus => 2, addr => 10 },
4714     };
4715
4716     if (defined($devices->{$id}->{bus}) && defined($devices->{$id}->{addr})) {
4717            my $addr = sprintf("0x%x", $devices->{$id}->{addr});
4718            my $bus = $devices->{$id}->{bus};
4719            $res = ",bus=pci.$bus,addr=$addr";
4720            $bridges->{$bus} = 1 if $bridges;
4721     }
4722     return $res;
4723
4724 }
4725
4726 sub print_pcie_addr {
4727     my ($id) = @_;
4728
4729     my $res = '';
4730     my $devices = {
4731         hostpci0 => { bus => "ich9-pcie-port-1", addr => 0 },
4732         hostpci1 => { bus => "ich9-pcie-port-2", addr => 0 },
4733         hostpci2 => { bus => "ich9-pcie-port-3", addr => 0 },
4734         hostpci3 => { bus => "ich9-pcie-port-4", addr => 0 },
4735     };
4736
4737     if (defined($devices->{$id}->{bus}) && defined($devices->{$id}->{addr})) {
4738            my $addr = sprintf("0x%x", $devices->{$id}->{addr});
4739            my $bus = $devices->{$id}->{bus};
4740            $res = ",bus=$bus,addr=$addr";
4741     }
4742     return $res;
4743
4744 }
4745
4746 # vzdump restore implementaion
4747
4748 sub tar_archive_read_firstfile {
4749     my $archive = shift;
4750
4751     die "ERROR: file '$archive' does not exist\n" if ! -f $archive;
4752
4753     # try to detect archive type first
4754     my $pid = open (TMP, "tar tf '$archive'|") ||
4755         die "unable to open file '$archive'\n";
4756     my $firstfile = <TMP>;
4757     kill 15, $pid;
4758     close TMP;
4759
4760     die "ERROR: archive contaions no data\n" if !$firstfile;
4761     chomp $firstfile;
4762
4763     return $firstfile;
4764 }
4765
4766 sub tar_restore_cleanup {
4767     my ($storecfg, $statfile) = @_;
4768
4769     print STDERR "starting cleanup\n";
4770
4771     if (my $fd = IO::File->new($statfile, "r")) {
4772         while (defined(my $line = <$fd>)) {
4773             if ($line =~ m/vzdump:([^\s:]*):(\S+)$/) {
4774                 my $volid = $2;
4775                 eval {
4776                     if ($volid =~ m|^/|) {
4777                         unlink $volid || die 'unlink failed\n';
4778                     } else {
4779                         PVE::Storage::vdisk_free($storecfg, $volid);
4780                     }
4781                     print STDERR "temporary volume '$volid' sucessfuly removed\n";
4782                 };
4783                 print STDERR "unable to cleanup '$volid' - $@" if $@;
4784             } else {
4785                 print STDERR "unable to parse line in statfile - $line";
4786             }
4787         }
4788         $fd->close();
4789     }
4790 }
4791
4792 sub restore_archive {
4793     my ($archive, $vmid, $user, $opts) = @_;
4794
4795     my $format = $opts->{format};
4796     my $comp;
4797
4798     if ($archive =~ m/\.tgz$/ || $archive =~ m/\.tar\.gz$/) {
4799         $format = 'tar' if !$format;
4800         $comp = 'gzip';
4801     } elsif ($archive =~ m/\.tar$/) {
4802         $format = 'tar' if !$format;
4803     } elsif ($archive =~ m/.tar.lzo$/) {
4804         $format = 'tar' if !$format;
4805         $comp = 'lzop';
4806     } elsif ($archive =~ m/\.vma$/) {
4807         $format = 'vma' if !$format;
4808     } elsif ($archive =~ m/\.vma\.gz$/) {
4809         $format = 'vma' if !$format;
4810         $comp = 'gzip';
4811     } elsif ($archive =~ m/\.vma\.lzo$/) {
4812         $format = 'vma' if !$format;
4813         $comp = 'lzop';
4814     } else {
4815         $format = 'vma' if !$format; # default
4816     }
4817
4818     # try to detect archive format
4819     if ($format eq 'tar') {
4820         return restore_tar_archive($archive, $vmid, $user, $opts);
4821     } else {
4822         return restore_vma_archive($archive, $vmid, $user, $opts, $comp);
4823     }
4824 }
4825
4826 sub restore_update_config_line {
4827     my ($outfd, $cookie, $vmid, $map, $line, $unique) = @_;
4828
4829     return if $line =~ m/^\#qmdump\#/;
4830     return if $line =~ m/^\#vzdump\#/;
4831     return if $line =~ m/^lock:/;
4832     return if $line =~ m/^unused\d+:/;
4833     return if $line =~ m/^parent:/;
4834     return if $line =~ m/^template:/; # restored VM is never a template
4835
4836     if (($line =~ m/^(vlan(\d+)):\s*(\S+)\s*$/)) {
4837         # try to convert old 1.X settings
4838         my ($id, $ind, $ethcfg) = ($1, $2, $3);
4839         foreach my $devconfig (PVE::Tools::split_list($ethcfg)) {
4840             my ($model, $macaddr) = split(/\=/, $devconfig);
4841             $macaddr = PVE::Tools::random_ether_addr() if !$macaddr || $unique;
4842             my $net = {
4843                 model => $model,
4844                 bridge => "vmbr$ind",
4845                 macaddr => $macaddr,
4846             };
4847             my $netstr = print_net($net);
4848
4849             print $outfd "net$cookie->{netcount}: $netstr\n";
4850             $cookie->{netcount}++;
4851         }
4852     } elsif (($line =~ m/^(net\d+):\s*(\S+)\s*$/) && $unique) {
4853         my ($id, $netstr) = ($1, $2);
4854         my $net = parse_net($netstr);
4855         $net->{macaddr} = PVE::Tools::random_ether_addr() if $net->{macaddr};
4856         $netstr = print_net($net);
4857         print $outfd "$id: $netstr\n";
4858     } elsif ($line =~ m/^((ide|scsi|virtio|sata)\d+):\s*(\S+)\s*$/) {
4859         my $virtdev = $1;
4860         my $value = $3;
4861         if ($line =~ m/backup=no/) {
4862             print $outfd "#$line";
4863         } elsif ($virtdev && $map->{$virtdev}) {
4864             my $di = parse_drive($virtdev, $value);
4865             delete $di->{format}; # format can change on restore
4866             $di->{file} = $map->{$virtdev};
4867             $value = print_drive($vmid, $di);
4868             print $outfd "$virtdev: $value\n";
4869         } else {
4870             print $outfd $line;
4871         }
4872     } else {
4873         print $outfd $line;
4874     }
4875 }
4876
4877 sub scan_volids {
4878     my ($cfg, $vmid) = @_;
4879
4880     my $info = PVE::Storage::vdisk_list($cfg, undef, $vmid);
4881
4882     my $volid_hash = {};
4883     foreach my $storeid (keys %$info) {
4884         foreach my $item (@{$info->{$storeid}}) {
4885             next if !($item->{volid} && $item->{size});
4886             $item->{path} = PVE::Storage::path($cfg, $item->{volid});
4887             $volid_hash->{$item->{volid}} = $item;
4888         }
4889     }
4890
4891     return $volid_hash;
4892 }
4893
4894 sub get_used_paths {
4895     my ($vmid, $storecfg, $conf, $scan_snapshots, $skip_drive) = @_;
4896
4897     my $used_path = {};
4898
4899     my $scan_config = sub {
4900         my ($cref, $snapname) = @_;
4901
4902         foreach my $key (keys %$cref) {
4903             my $value = $cref->{$key};
4904             if (valid_drivename($key)) {
4905                 next if $skip_drive && $key eq $skip_drive;
4906                 my $drive = parse_drive($key, $value);
4907                 next if !$drive || !$drive->{file} || drive_is_cdrom($drive);
4908                 if ($drive->{file} =~ m!^/!) {
4909                     $used_path->{$drive->{file}}++; # = 1;
4910                 } else {
4911                     my ($storeid, $volname) = PVE::Storage::parse_volume_id($drive->{file}, 1);
4912                     next if !$storeid;
4913                     my $scfg = PVE::Storage::storage_config($storecfg, $storeid, 1);
4914                     next if !$scfg;
4915                     my $path = PVE::Storage::path($storecfg, $drive->{file}, $snapname);
4916                     $used_path->{$path}++; # = 1;
4917                 }
4918             }
4919         }
4920     };
4921
4922     &$scan_config($conf);
4923
4924     undef $skip_drive;
4925
4926     if ($scan_snapshots) {
4927         foreach my $snapname (keys %{$conf->{snapshots}}) {
4928             &$scan_config($conf->{snapshots}->{$snapname}, $snapname);
4929         }
4930     }
4931
4932     return $used_path;
4933 }
4934
4935 sub update_disksize {
4936     my ($vmid, $conf, $volid_hash) = @_;
4937
4938     my $changes;
4939
4940     my $used = {};
4941
4942     # Note: it is allowed to define multiple storages with same path (alias), so
4943     # we need to check both 'volid' and real 'path' (two different volid can point
4944     # to the same path).
4945
4946     my $usedpath = {};
4947
4948     # update size info
4949     foreach my $opt (keys %$conf) {
4950         if (valid_drivename($opt)) {
4951             my $drive = parse_drive($opt, $conf->{$opt});
4952             my $volid = $drive->{file};
4953             next if !$volid;
4954
4955             $used->{$volid} = 1;
4956             if ($volid_hash->{$volid} &&
4957                 (my $path = $volid_hash->{$volid}->{path})) {
4958                 $usedpath->{$path} = 1;
4959             }
4960
4961             next if drive_is_cdrom($drive);
4962             next if !$volid_hash->{$volid};
4963
4964             $drive->{size} = $volid_hash->{$volid}->{size};
4965             my $new = print_drive($vmid, $drive);
4966             if ($new ne $conf->{$opt}) {
4967                 $changes = 1;
4968                 $conf->{$opt} = $new;
4969             }
4970         }
4971     }
4972
4973     # remove 'unusedX' entry if volume is used
4974     foreach my $opt (keys %$conf) {
4975         next if $opt !~ m/^unused\d+$/;
4976         my $volid = $conf->{$opt};
4977         my $path = $volid_hash->{$volid}->{path} if $volid_hash->{$volid};
4978         if ($used->{$volid} || ($path && $usedpath->{$path})) {
4979             $changes = 1;
4980             delete $conf->{$opt};
4981         }
4982     }
4983
4984     foreach my $volid (sort keys %$volid_hash) {
4985         next if $volid =~ m/vm-$vmid-state-/;
4986         next if $used->{$volid};
4987         my $path = $volid_hash->{$volid}->{path};
4988         next if !$path; # just to be sure
4989         next if $usedpath->{$path};
4990         $changes = 1;
4991         add_unused_volume($conf, $volid);
4992         $usedpath->{$path} = 1; # avoid to add more than once (aliases)
4993     }
4994
4995     return $changes;
4996 }
4997
4998 sub rescan {
4999     my ($vmid, $nolock) = @_;
5000
5001     my $cfg = PVE::Cluster::cfs_read_file("storage.cfg");
5002
5003     my $volid_hash = scan_volids($cfg, $vmid);
5004
5005     my $updatefn =  sub {
5006         my ($vmid) = @_;
5007
5008         my $conf = load_config($vmid);
5009
5010         check_lock($conf);
5011
5012         my $vm_volids = {};
5013         foreach my $volid (keys %$volid_hash) {
5014             my $info = $volid_hash->{$volid};
5015             $vm_volids->{$volid} = $info if $info->{vmid} && $info->{vmid} == $vmid;
5016         }
5017
5018         my $changes = update_disksize($vmid, $conf, $vm_volids);
5019
5020         update_config_nolock($vmid, $conf, 1) if $changes;
5021     };
5022
5023     if (defined($vmid)) {
5024         if ($nolock) {
5025             &$updatefn($vmid);
5026         } else {
5027             lock_config($vmid, $updatefn, $vmid);
5028         }
5029     } else {
5030         my $vmlist = config_list();
5031         foreach my $vmid (keys %$vmlist) {
5032             if ($nolock) {
5033                 &$updatefn($vmid);
5034             } else {
5035                 lock_config($vmid, $updatefn, $vmid);
5036             }
5037         }
5038     }
5039 }
5040
5041 sub restore_vma_archive {
5042     my ($archive, $vmid, $user, $opts, $comp) = @_;
5043
5044     my $input = $archive eq '-' ? "<&STDIN" : undef;
5045     my $readfrom = $archive;
5046
5047     my $uncomp = '';
5048     if ($comp) {
5049         $readfrom = '-';
5050         my $qarchive = PVE::Tools::shellquote($archive);
5051         if ($comp eq 'gzip') {
5052             $uncomp = "zcat $qarchive|";
5053         } elsif ($comp eq 'lzop') {
5054             $uncomp = "lzop -d -c $qarchive|";
5055         } else {
5056             die "unknown compression method '$comp'\n";
5057         }
5058
5059     }
5060
5061     my $tmpdir = "/var/tmp/vzdumptmp$$";
5062     rmtree $tmpdir;
5063
5064     # disable interrupts (always do cleanups)
5065     local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = sub {
5066         warn "got interrupt - ignored\n";
5067     };
5068
5069     my $mapfifo = "/var/tmp/vzdumptmp$$.fifo";
5070     POSIX::mkfifo($mapfifo, 0600);
5071     my $fifofh;
5072
5073     my $openfifo = sub {
5074         open($fifofh, '>', $mapfifo) || die $!;
5075     };
5076
5077     my $cmd = "${uncomp}vma extract -v -r $mapfifo $readfrom $tmpdir";
5078
5079     my $oldtimeout;
5080     my $timeout = 5;
5081
5082     my $devinfo = {};
5083
5084     my $rpcenv = PVE::RPCEnvironment::get();
5085
5086     my $conffile = config_file($vmid);
5087     my $tmpfn = "$conffile.$$.tmp";
5088
5089     # Note: $oldconf is undef if VM does not exists
5090     my $oldconf = PVE::Cluster::cfs_read_file(cfs_config_path($vmid));
5091
5092     my $print_devmap = sub {
5093         my $virtdev_hash = {};
5094
5095         my $cfgfn = "$tmpdir/qemu-server.conf";
5096
5097         # we can read the config - that is already extracted
5098         my $fh = IO::File->new($cfgfn, "r") ||
5099             "unable to read qemu-server.conf - $!\n";
5100
5101         while (defined(my $line = <$fh>)) {
5102             if ($line =~ m/^\#qmdump\#map:(\S+):(\S+):(\S*):(\S*):$/) {
5103                 my ($virtdev, $devname, $storeid, $format) = ($1, $2, $3, $4);
5104                 die "archive does not contain data for drive '$virtdev'\n"
5105                     if !$devinfo->{$devname};
5106                 if (defined($opts->{storage})) {
5107                     $storeid = $opts->{storage} || 'local';
5108                 } elsif (!$storeid) {
5109                     $storeid = 'local';
5110                 }
5111                 $format = 'raw' if !$format;
5112                 $devinfo->{$devname}->{devname} = $devname;
5113                 $devinfo->{$devname}->{virtdev} = $virtdev;
5114                 $devinfo->{$devname}->{format} = $format;
5115                 $devinfo->{$devname}->{storeid} = $storeid;
5116
5117                 # check permission on storage
5118                 my $pool = $opts->{pool}; # todo: do we need that?
5119                 if ($user ne 'root@pam') {
5120                     $rpcenv->check($user, "/storage/$storeid", ['Datastore.AllocateSpace']);
5121                 }
5122
5123                 $virtdev_hash->{$virtdev} = $devinfo->{$devname};
5124             }
5125         }
5126
5127         foreach my $devname (keys %$devinfo) {
5128             die "found no device mapping information for device '$devname'\n"
5129                 if !$devinfo->{$devname}->{virtdev};
5130         }
5131
5132         my $cfg = cfs_read_file('storage.cfg');
5133
5134         # create empty/temp config
5135         if ($oldconf) {
5136             PVE::Tools::file_set_contents($conffile, "memory: 128\n");
5137             foreach_drive($oldconf, sub {
5138                 my ($ds, $drive) = @_;
5139
5140                 return if drive_is_cdrom($drive);
5141
5142                 my $volid = $drive->{file};
5143
5144                 return if !$volid || $volid =~ m|^/|;
5145
5146                 my ($path, $owner) = PVE::Storage::path($cfg, $volid);
5147                 return if !$path || !$owner || ($owner != $vmid);
5148
5149                 # Note: only delete disk we want to restore
5150                 # other volumes will become unused
5151                 if ($virtdev_hash->{$ds}) {
5152                     PVE::Storage::vdisk_free($cfg, $volid);
5153                 }
5154             });
5155         }
5156
5157         my $map = {};
5158         foreach my $virtdev (sort keys %$virtdev_hash) {
5159             my $d = $virtdev_hash->{$virtdev};
5160             my $alloc_size = int(($d->{size} + 1024 - 1)/1024);
5161             my $scfg = PVE::Storage::storage_config($cfg, $d->{storeid});
5162
5163             # test if requested format is supported
5164             my ($defFormat, $validFormats) = PVE::Storage::storage_default_format($cfg, $d->{storeid});
5165             my $supported = grep { $_ eq $d->{format} } @$validFormats;
5166             $d->{format} = $defFormat if !$supported;
5167
5168             my $volid = PVE::Storage::vdisk_alloc($cfg, $d->{storeid}, $vmid,
5169                                                   $d->{format}, undef, $alloc_size);
5170             print STDERR "new volume ID is '$volid'\n";
5171             $d->{volid} = $volid;
5172             my $path = PVE::Storage::path($cfg, $volid);
5173
5174             my $write_zeros = 1;
5175             # fixme: what other storages types initialize volumes with zero?
5176             if ($scfg->{type} eq 'dir' || $scfg->{type} eq 'nfs' || $scfg->{type} eq 'glusterfs' ||
5177                 $scfg->{type} eq 'sheepdog' || $scfg->{type} eq 'rbd') {
5178                 $write_zeros = 0;
5179             }
5180
5181             print $fifofh "${write_zeros}:$d->{devname}=$path\n";
5182
5183             print "map '$d->{devname}' to '$path' (write zeros = ${write_zeros})\n";
5184             $map->{$virtdev} = $volid;
5185         }
5186
5187         $fh->seek(0, 0) || die "seek failed - $!\n";
5188
5189         my $outfd = new IO::File ($tmpfn, "w") ||
5190             die "unable to write config for VM $vmid\n";
5191
5192         my $cookie = { netcount => 0 };
5193         while (defined(my $line = <$fh>)) {
5194             restore_update_config_line($outfd, $cookie, $vmid, $map, $line, $opts->{unique});
5195         }
5196
5197         $fh->close();
5198         $outfd->close();
5199     };
5200
5201     eval {
5202         # enable interrupts
5203         local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = $SIG{PIPE} = sub {
5204             die "interrupted by signal\n";
5205         };
5206         local $SIG{ALRM} = sub { die "got timeout\n"; };
5207
5208         $oldtimeout = alarm($timeout);
5209
5210         my $parser = sub {
5211             my $line = shift;
5212
5213             print "$line\n";
5214
5215             if ($line =~ m/^DEV:\sdev_id=(\d+)\ssize:\s(\d+)\sdevname:\s(\S+)$/) {
5216                 my ($dev_id, $size, $devname) = ($1, $2, $3);
5217                 $devinfo->{$devname} = { size => $size, dev_id => $dev_id };
5218             } elsif ($line =~ m/^CTIME: /) {
5219                 # we correctly received the vma config, so we can disable
5220                 # the timeout now for disk allocation (set to 10 minutes, so
5221                 # that we always timeout if something goes wrong)
5222                 alarm(600);
5223                 &$print_devmap();
5224                 print $fifofh "done\n";
5225                 my $tmp = $oldtimeout || 0;
5226                 $oldtimeout = undef;
5227                 alarm($tmp);
5228                 close($fifofh);
5229             }
5230         };
5231
5232         print "restore vma archive: $cmd\n";
5233         run_command($cmd, input => $input, outfunc => $parser, afterfork => $openfifo);
5234     };
5235     my $err = $@;
5236
5237     alarm($oldtimeout) if $oldtimeout;
5238
5239     unlink $mapfifo;
5240
5241     if ($err) {
5242         rmtree $tmpdir;
5243         unlink $tmpfn;
5244
5245         my $cfg = cfs_read_file('storage.cfg');
5246         foreach my $devname (keys %$devinfo) {
5247             my $volid = $devinfo->{$devname}->{volid};
5248             next if !$volid;
5249             eval {
5250                 if ($volid =~ m|^/|) {
5251                     unlink $volid || die 'unlink failed\n';
5252                 } else {
5253                     PVE::Storage::vdisk_free($cfg, $volid);
5254                 }
5255                 print STDERR "temporary volume '$volid' sucessfuly removed\n";
5256             };
5257             print STDERR "unable to cleanup '$volid' - $@" if $@;
5258         }
5259         die $err;
5260     }
5261
5262     rmtree $tmpdir;
5263
5264     rename($tmpfn, $conffile) ||
5265         die "unable to commit configuration file '$conffile'\n";
5266
5267     PVE::Cluster::cfs_update(); # make sure we read new file
5268
5269     eval { rescan($vmid, 1); };
5270     warn $@ if $@;
5271 }
5272
5273 sub restore_tar_archive {
5274     my ($archive, $vmid, $user, $opts) = @_;
5275
5276     if ($archive ne '-') {
5277         my $firstfile = tar_archive_read_firstfile($archive);
5278         die "ERROR: file '$archive' dos not lock like a QemuServer vzdump backup\n"
5279             if $firstfile ne 'qemu-server.conf';
5280     }
5281
5282     my $storecfg = cfs_read_file('storage.cfg');
5283
5284     # destroy existing data - keep empty config
5285     my $vmcfgfn = config_file($vmid);
5286     destroy_vm($storecfg, $vmid, 1) if -f $vmcfgfn;
5287
5288     my $tocmd = "/usr/lib/qemu-server/qmextract";
5289
5290     $tocmd .= " --storage " . PVE::Tools::shellquote($opts->{storage}) if $opts->{storage};
5291     $tocmd .= " --pool " . PVE::Tools::shellquote($opts->{pool}) if $opts->{pool};
5292     $tocmd .= ' --prealloc' if $opts->{prealloc};
5293     $tocmd .= ' --info' if $opts->{info};
5294
5295     # tar option "xf" does not autodetect compression when read from STDIN,
5296     # so we pipe to zcat
5297     my $cmd = "zcat -f|tar xf " . PVE::Tools::shellquote($archive) . " " .
5298         PVE::Tools::shellquote("--to-command=$tocmd");
5299
5300     my $tmpdir = "/var/tmp/vzdumptmp$$";
5301     mkpath $tmpdir;
5302
5303     local $ENV{VZDUMP_TMPDIR} = $tmpdir;
5304     local $ENV{VZDUMP_VMID} = $vmid;
5305     local $ENV{VZDUMP_USER} = $user;
5306
5307     my $conffile = config_file($vmid);
5308     my $tmpfn = "$conffile.$$.tmp";
5309
5310     # disable interrupts (always do cleanups)
5311     local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = sub {
5312         print STDERR "got interrupt - ignored\n";
5313     };
5314
5315     eval {
5316         # enable interrupts
5317         local $SIG{INT} = $SIG{TERM} = $SIG{QUIT} = $SIG{HUP} = $SIG{PIPE} = sub {
5318             die "interrupted by signal\n";
5319         };
5320
5321         if ($archive eq '-') {
5322             print "extracting archive from STDIN\n";
5323             run_command($cmd, input => "<&STDIN");
5324         } else {
5325             print "extracting archive '$archive'\n";
5326             run_command($cmd);
5327         }
5328
5329         return if $opts->{info};
5330
5331         # read new mapping
5332         my $map = {};
5333         my $statfile = "$tmpdir/qmrestore.stat";
5334         if (my $fd = IO::File->new($statfile, "r")) {
5335             while (defined (my $line = <$fd>)) {
5336                 if ($line =~ m/vzdump:([^\s:]*):(\S+)$/) {
5337                     $map->{$1} = $2 if $1;
5338                 } else {
5339                     print STDERR "unable to parse line in statfile - $line\n";
5340                 }
5341             }
5342             $fd->close();
5343         }
5344
5345         my $confsrc = "$tmpdir/qemu-server.conf";
5346
5347         my $srcfd = new IO::File($confsrc, "r") ||
5348             die "unable to open file '$confsrc'\n";
5349
5350         my $outfd = new IO::File ($tmpfn, "w") ||
5351             die "unable to write config for VM $vmid\n";
5352
5353         my $cookie = { netcount => 0 };
5354         while (defined (my $line = <$srcfd>)) {
5355             restore_update_config_line($outfd, $cookie, $vmid, $map, $line, $opts->{unique});
5356         }
5357
5358         $srcfd->close();
5359         $outfd->close();
5360     };
5361     my $err = $@;
5362
5363     if ($err) {
5364
5365         unlink $tmpfn;
5366
5367         tar_restore_cleanup($storecfg, "$tmpdir/qmrestore.stat") if !$opts->{info};
5368
5369         die $err;
5370     }
5371
5372     rmtree $tmpdir;
5373
5374     rename $tmpfn, $conffile ||
5375         die "unable to commit configuration file '$conffile'\n";
5376
5377     PVE::Cluster::cfs_update(); # make sure we read new file
5378
5379     eval { rescan($vmid, 1); };
5380     warn $@ if $@;
5381 };
5382
5383
5384 # Internal snapshots
5385
5386 # NOTE: Snapshot create/delete involves several non-atomic
5387 # action, and can take a long time.
5388 # So we try to avoid locking the file and use 'lock' variable
5389 # inside the config file instead.
5390
5391 my $snapshot_copy_config = sub {
5392     my ($source, $dest) = @_;
5393
5394     foreach my $k (keys %$source) {
5395         next if $k eq 'snapshots';
5396         next if $k eq 'snapstate';
5397         next if $k eq 'snaptime';
5398         next if $k eq 'vmstate';
5399         next if $k eq 'lock';
5400         next if $k eq 'digest';
5401         next if $k eq 'description';
5402         next if $k =~ m/^unused\d+$/;
5403
5404         $dest->{$k} = $source->{$k};
5405     }
5406 };
5407
5408 my $snapshot_apply_config = sub {
5409     my ($conf, $snap) = @_;
5410
5411     # copy snapshot list
5412     my $newconf = {
5413         snapshots => $conf->{snapshots},
5414     };
5415
5416     # keep description and list of unused disks
5417     foreach my $k (keys %$conf) {
5418         next if !($k =~ m/^unused\d+$/ || $k eq 'description');
5419         $newconf->{$k} = $conf->{$k};
5420     }
5421
5422     &$snapshot_copy_config($snap, $newconf);
5423
5424     return $newconf;
5425 };
5426
5427 sub foreach_writable_storage {
5428     my ($conf, $func) = @_;
5429
5430     my $sidhash = {};
5431
5432     foreach my $ds (keys %$conf) {
5433         next if !valid_drivename($ds);
5434
5435         my $drive = parse_drive($ds, $conf->{$ds});
5436         next if !$drive;
5437         next if drive_is_cdrom($drive);
5438
5439         my $volid = $drive->{file};
5440
5441         my ($sid, $volname) = PVE::Storage::parse_volume_id($volid, 1);
5442         $sidhash->{$sid} = $sid if $sid;
5443     }
5444
5445     foreach my $sid (sort keys %$sidhash) {
5446         &$func($sid);
5447     }
5448 }
5449
5450 my $alloc_vmstate_volid = sub {
5451     my ($storecfg, $vmid, $conf, $snapname) = @_;
5452
5453     # Note: we try to be smart when selecting a $target storage
5454
5455     my $target;
5456
5457     # search shared storage first
5458     foreach_writable_storage($conf, sub {
5459         my ($sid) = @_;
5460         my $scfg = PVE::Storage::storage_config($storecfg, $sid);
5461         return if !$scfg->{shared};
5462
5463         $target = $sid if !$target || $scfg->{path}; # prefer file based storage
5464     });
5465
5466     if (!$target) {
5467         # now search local storage
5468         foreach_writable_storage($conf, sub {
5469             my ($sid) = @_;
5470             my $scfg = PVE::Storage::storage_config($storecfg, $sid);
5471             return if $scfg->{shared};
5472
5473             $target = $sid if !$target || $scfg->{path}; # prefer file based storage;
5474         });
5475     }
5476
5477     $target = 'local' if !$target;
5478
5479     my $driver_state_size = 500; # assume 32MB is enough to safe all driver state;
5480     # we abort live save after $conf->{memory}, so we need at max twice that space
5481     my $size = $conf->{memory}*2 + $driver_state_size;
5482
5483     my $name = "vm-$vmid-state-$snapname";
5484     my $scfg = PVE::Storage::storage_config($storecfg, $target);
5485     $name .= ".raw" if $scfg->{path}; # add filename extension for file base storage
5486     my $volid = PVE::Storage::vdisk_alloc($storecfg, $target, $vmid, 'raw', $name, $size*1024);
5487
5488     return $volid;
5489 };
5490
5491 my $snapshot_prepare = sub {
5492     my ($vmid, $snapname, $save_vmstate, $comment) = @_;
5493
5494     my $snap;
5495
5496     my $updatefn =  sub {
5497
5498         my $conf = load_config($vmid);
5499
5500         die "you can't take a snapshot if it's a template\n"
5501             if is_template($conf);
5502
5503         check_lock($conf);
5504
5505         $conf->{lock} = 'snapshot';
5506
5507         die "snapshot name '$snapname' already used\n"
5508             if defined($conf->{snapshots}->{$snapname});
5509
5510         my $storecfg = PVE::Storage::config();
5511         die "snapshot feature is not available" if !has_feature('snapshot', $conf, $storecfg);
5512
5513         $snap = $conf->{snapshots}->{$snapname} = {};
5514
5515         if ($save_vmstate && check_running($vmid)) {
5516             $snap->{vmstate} = &$alloc_vmstate_volid($storecfg, $vmid, $conf, $snapname);
5517         }
5518
5519         &$snapshot_copy_config($conf, $snap);
5520
5521         $snap->{snapstate} = "prepare";
5522         $snap->{snaptime} = time();
5523         $snap->{description} = $comment if $comment;
5524
5525         # always overwrite machine if we save vmstate. This makes sure we
5526         # can restore it later using correct machine type
5527         $snap->{machine} = get_current_qemu_machine($vmid) if $snap->{vmstate};
5528
5529         update_config_nolock($vmid, $conf, 1);
5530     };
5531
5532     lock_config($vmid, $updatefn);
5533
5534     return $snap;
5535 };
5536
5537 my $snapshot_commit = sub {
5538     my ($vmid, $snapname) = @_;
5539
5540     my $updatefn = sub {
5541
5542         my $conf = load_config($vmid);
5543
5544         die "missing snapshot lock\n"
5545             if !($conf->{lock} && $conf->{lock} eq 'snapshot');
5546
5547         my $has_machine_config = defined($conf->{machine});
5548
5549         my $snap = $conf->{snapshots}->{$snapname};
5550
5551         die "snapshot '$snapname' does not exist\n" if !defined($snap);
5552
5553         die "wrong snapshot state\n"
5554             if !($snap->{snapstate} && $snap->{snapstate} eq "prepare");
5555
5556         delete $snap->{snapstate};
5557         delete $conf->{lock};
5558
5559         my $newconf = &$snapshot_apply_config($conf, $snap);
5560
5561         delete $newconf->{machine} if !$has_machine_config;
5562
5563         $newconf->{parent} = $snapname;
5564
5565         update_config_nolock($vmid, $newconf, 1);
5566     };
5567
5568     lock_config($vmid, $updatefn);
5569 };
5570
5571 sub snapshot_rollback {
5572     my ($vmid, $snapname) = @_;
5573
5574     my $prepare = 1;
5575
5576     my $storecfg = PVE::Storage::config();
5577
5578     my $conf = load_config($vmid);
5579
5580     my $get_snapshot_config = sub {
5581
5582         die "you can't rollback if vm is a template\n" if is_template($conf);
5583
5584         my $res = $conf->{snapshots}->{$snapname};
5585
5586         die "snapshot '$snapname' does not exist\n" if !defined($res);
5587
5588         return $res;
5589     };
5590
5591     my $snap = &$get_snapshot_config();
5592
5593     foreach_drive($snap, sub {
5594         my ($ds, $drive) = @_;
5595
5596         return if drive_is_cdrom($drive);
5597
5598         my $volid = $drive->{file};
5599
5600         PVE::Storage::volume_rollback_is_possible($storecfg, $volid, $snapname);
5601     });
5602
5603     my $updatefn = sub {
5604
5605         $conf = load_config($vmid);
5606
5607         $snap = &$get_snapshot_config();
5608
5609         die "unable to rollback to incomplete snapshot (snapstate = $snap->{snapstate})\n"
5610             if $snap->{snapstate};
5611
5612         if ($prepare) {
5613             check_lock($conf);
5614             vm_stop($storecfg, $vmid, undef, undef, 5, undef, undef);
5615         }
5616
5617         die "unable to rollback vm $vmid: vm is running\n"
5618             if check_running($vmid);
5619
5620         if ($prepare) {
5621             $conf->{lock} = 'rollback';
5622         } else {
5623             die "got wrong lock\n" if !($conf->{lock} && $conf->{lock} eq 'rollback');
5624             delete $conf->{lock};
5625         }
5626
5627         my $forcemachine;
5628
5629         if (!$prepare) {
5630             my $has_machine_config = defined($conf->{machine});
5631
5632             # copy snapshot config to current config
5633             $conf = &$snapshot_apply_config($conf, $snap);
5634             $conf->{parent} = $snapname;
5635
5636             # Note: old code did not store 'machine', so we try to be smart
5637             # and guess the snapshot was generated with kvm 1.4 (pc-i440fx-1.4).
5638             $forcemachine = $conf->{machine} || 'pc-i440fx-1.4';
5639             # we remove the 'machine' configuration if not explicitly specified
5640             # in the original config.
5641             delete $conf->{machine} if $snap->{vmstate} && !$has_machine_config;
5642         }
5643
5644         update_config_nolock($vmid, $conf, 1);
5645
5646         if (!$prepare && $snap->{vmstate}) {
5647             my $statefile = PVE::Storage::path($storecfg, $snap->{vmstate});
5648             vm_start($storecfg, $vmid, $statefile, undef, undef, undef, $forcemachine);
5649         }
5650     };
5651
5652     lock_config($vmid, $updatefn);
5653
5654     foreach_drive($snap, sub {
5655         my ($ds, $drive) = @_;
5656
5657         return if drive_is_cdrom($drive);
5658
5659         my $volid = $drive->{file};
5660         my $device = "drive-$ds";
5661
5662         PVE::Storage::volume_snapshot_rollback($storecfg, $volid, $snapname);
5663     });
5664
5665     $prepare = 0;
5666     lock_config($vmid, $updatefn);
5667 }
5668
5669 my $savevm_wait = sub {
5670     my ($vmid) = @_;
5671
5672     for(;;) {
5673         my $stat = vm_mon_cmd_nocheck($vmid, "query-savevm");
5674         if (!$stat->{status}) {
5675             die "savevm not active\n";
5676         } elsif ($stat->{status} eq 'active') {
5677             sleep(1);
5678             next;
5679         } elsif ($stat->{status} eq 'completed') {
5680             last;
5681         } else {
5682             die "query-savevm returned status '$stat->{status}'\n";
5683         }
5684     }
5685 };
5686
5687 sub snapshot_create {
5688     my ($vmid, $snapname, $save_vmstate, $comment) = @_;
5689
5690     my $snap = &$snapshot_prepare($vmid, $snapname, $save_vmstate, $comment);
5691
5692     $save_vmstate = 0 if !$snap->{vmstate}; # vm is not running
5693
5694     my $config = load_config($vmid);
5695
5696     my $running = check_running($vmid);
5697
5698     my $freezefs = $running && $config->{agent};
5699     $freezefs = 0 if $snap->{vmstate}; # not needed if we save RAM
5700
5701     my $drivehash = {};
5702
5703     if ($freezefs) {
5704         eval { vm_mon_cmd($vmid, "guest-fsfreeze-freeze"); };
5705         warn "guest-fsfreeze-freeze problems - $@" if $@;
5706     }
5707
5708     eval {
5709         # create internal snapshots of all drives
5710
5711         my $storecfg = PVE::Storage::config();
5712
5713         if ($running) {
5714             if ($snap->{vmstate}) {
5715                 my $path = PVE::Storage::path($storecfg, $snap->{vmstate});
5716                 vm_mon_cmd($vmid, "savevm-start", statefile => $path);
5717                 &$savevm_wait($vmid);
5718             } else {
5719                 vm_mon_cmd($vmid, "savevm-start");
5720             }
5721         };
5722
5723         foreach_drive($snap, sub {
5724             my ($ds, $drive) = @_;
5725
5726             return if drive_is_cdrom($drive);
5727
5728             my $volid = $drive->{file};
5729             my $device = "drive-$ds";
5730
5731             qemu_volume_snapshot($vmid, $device, $storecfg, $volid, $snapname);
5732             $drivehash->{$ds} = 1;
5733        });
5734     };
5735     my $err = $@;
5736
5737     if ($running) {
5738         eval { vm_mon_cmd($vmid, "savevm-end")  };
5739         warn $@ if $@;
5740
5741         if ($freezefs) {
5742             eval { vm_mon_cmd($vmid, "guest-fsfreeze-thaw"); };
5743             warn "guest-fsfreeze-thaw problems - $@" if $@;
5744         }
5745
5746         # savevm-end is async, we need to wait
5747         for (;;) {
5748             my $stat = vm_mon_cmd_nocheck($vmid, "query-savevm");
5749             if (!$stat->{bytes}) {
5750                 last;
5751             } else {
5752                 print "savevm not yet finished\n";
5753                 sleep(1);
5754                 next;
5755             }
5756         }
5757     }
5758
5759     if ($err) {
5760         warn "snapshot create failed: starting cleanup\n";
5761         eval { snapshot_delete($vmid, $snapname, 0, $drivehash); };
5762         warn $@ if $@;
5763         die $err;
5764     }
5765
5766     &$snapshot_commit($vmid, $snapname);
5767 }
5768
5769 # Note: $drivehash is only set when called from snapshot_create.
5770 sub snapshot_delete {
5771     my ($vmid, $snapname, $force, $drivehash) = @_;
5772
5773     my $prepare = 1;
5774
5775     my $snap;
5776     my $unused = [];
5777
5778     my $unlink_parent = sub {
5779         my ($confref, $new_parent) = @_;
5780
5781         if ($confref->{parent} && $confref->{parent} eq $snapname) {
5782             if ($new_parent) {
5783                 $confref->{parent} = $new_parent;
5784             } else {
5785                 delete $confref->{parent};
5786             }
5787         }
5788     };
5789
5790     my $updatefn =  sub {
5791         my ($remove_drive) = @_;
5792
5793         my $conf = load_config($vmid);
5794
5795         if (!$drivehash) {
5796             check_lock($conf);
5797             die "you can't delete a snapshot if vm is a template\n"
5798                 if is_template($conf);
5799         }
5800
5801         $snap = $conf->{snapshots}->{$snapname};
5802
5803         die "snapshot '$snapname' does not exist\n" if !defined($snap);
5804
5805         # remove parent refs
5806         if (!$prepare) {
5807             &$unlink_parent($conf, $snap->{parent});
5808             foreach my $sn (keys %{$conf->{snapshots}}) {
5809                 next if $sn eq $snapname;
5810                 &$unlink_parent($conf->{snapshots}->{$sn}, $snap->{parent});
5811             }
5812         }
5813
5814         if ($remove_drive) {
5815             if ($remove_drive eq 'vmstate') {
5816                 delete $snap->{$remove_drive};
5817             } else {
5818                 my $drive = parse_drive($remove_drive, $snap->{$remove_drive});
5819                 my $volid = $drive->{file};
5820                 delete $snap->{$remove_drive};
5821                 add_unused_volume($conf, $volid);
5822             }
5823         }
5824
5825         if ($prepare) {
5826             $snap->{snapstate} = 'delete';
5827         } else {
5828             delete $conf->{snapshots}->{$snapname};
5829             delete $conf->{lock} if $drivehash;
5830             foreach my $volid (@$unused) {
5831                 add_unused_volume($conf, $volid);
5832             }
5833         }
5834
5835         update_config_nolock($vmid, $conf, 1);
5836     };
5837
5838     lock_config($vmid, $updatefn);
5839
5840     # now remove vmstate file
5841
5842     my $storecfg = PVE::Storage::config();
5843
5844     if ($snap->{vmstate}) {
5845         eval {  PVE::Storage::vdisk_free($storecfg, $snap->{vmstate}); };
5846         if (my $err = $@) {
5847             die $err if !$force;
5848             warn $err;
5849         }
5850         # save changes (remove vmstate from snapshot)
5851         lock_config($vmid, $updatefn, 'vmstate') if !$force;
5852     };
5853
5854     # now remove all internal snapshots
5855     foreach_drive($snap, sub {
5856         my ($ds, $drive) = @_;
5857
5858         return if drive_is_cdrom($drive);
5859
5860         my $volid = $drive->{file};
5861         my $device = "drive-$ds";
5862
5863         if (!$drivehash || $drivehash->{$ds}) {
5864             eval { qemu_volume_snapshot_delete($vmid, $device, $storecfg, $volid, $snapname); };
5865             if (my $err = $@) {
5866                 die $err if !$force;
5867                 warn $err;
5868             }
5869         }
5870
5871         # save changes (remove drive fron snapshot)
5872         lock_config($vmid, $updatefn, $ds) if !$force;
5873         push @$unused, $volid;
5874     });
5875
5876     # now cleanup config
5877     $prepare = 0;
5878     lock_config($vmid, $updatefn);
5879 }
5880
5881 sub has_feature {
5882     my ($feature, $conf, $storecfg, $snapname, $running) = @_;
5883
5884     my $err;
5885     foreach_drive($conf, sub {
5886         my ($ds, $drive) = @_;
5887
5888         return if drive_is_cdrom($drive);
5889         my $volid = $drive->{file};
5890         $err = 1 if !PVE::Storage::volume_has_feature($storecfg, $feature, $volid, $snapname, $running);
5891     });
5892
5893     return $err ? 0 : 1;
5894 }
5895
5896 sub template_create {
5897     my ($vmid, $conf, $disk) = @_;
5898
5899     my $storecfg = PVE::Storage::config();
5900
5901     foreach_drive($conf, sub {
5902         my ($ds, $drive) = @_;
5903
5904         return if drive_is_cdrom($drive);
5905         return if $disk && $ds ne $disk;
5906
5907         my $volid = $drive->{file};
5908         return if !PVE::Storage::volume_has_feature($storecfg, 'template', $volid);
5909
5910         my $voliddst = PVE::Storage::vdisk_create_base($storecfg, $volid);
5911         $drive->{file} = $voliddst;
5912         $conf->{$ds} = print_drive($vmid, $drive);
5913         update_config_nolock($vmid, $conf, 1);
5914     });
5915 }
5916
5917 sub is_template {
5918     my ($conf) = @_;
5919
5920     return 1 if defined $conf->{template} && $conf->{template} == 1;
5921 }
5922
5923 sub qemu_img_convert {
5924     my ($src_volid, $dst_volid, $size, $snapname) = @_;
5925
5926     my $storecfg = PVE::Storage::config();
5927     my ($src_storeid, $src_volname) = PVE::Storage::parse_volume_id($src_volid, 1);
5928     my ($dst_storeid, $dst_volname) = PVE::Storage::parse_volume_id($dst_volid, 1);
5929
5930     if ($src_storeid && $dst_storeid) {
5931         my $src_scfg = PVE::Storage::storage_config($storecfg, $src_storeid);
5932         my $dst_scfg = PVE::Storage::storage_config($storecfg, $dst_storeid);
5933
5934         my $src_format = qemu_img_format($src_scfg, $src_volname);
5935         my $dst_format = qemu_img_format($dst_scfg, $dst_volname);
5936
5937         my $src_path = PVE::Storage::path($storecfg, $src_volid, $snapname);
5938         my $dst_path = PVE::Storage::path($storecfg, $dst_volid);
5939
5940         my $cmd = [];
5941         push @$cmd, '/usr/bin/qemu-img', 'convert', '-t', 'writeback', '-p', '-n';
5942         push @$cmd, '-s', $snapname if($snapname && $src_format eq "qcow2");
5943         push @$cmd, '-f', $src_format, '-O', $dst_format, $src_path, $dst_path;
5944
5945         my $parser = sub {
5946             my $line = shift;
5947             if($line =~ m/\((\S+)\/100\%\)/){
5948                 my $percent = $1;
5949                 my $transferred = int($size * $percent / 100);
5950                 my $remaining = $size - $transferred;
5951
5952                 print "transferred: $transferred bytes remaining: $remaining bytes total: $size bytes progression: $percent %\n";
5953             }
5954
5955         };
5956
5957         eval  { run_command($cmd, timeout => undef, outfunc => $parser); };
5958         my $err = $@;
5959         die "copy failed: $err" if $err;
5960     }
5961 }
5962
5963 sub qemu_img_format {
5964     my ($scfg, $volname) = @_;
5965
5966     if ($scfg->{path} && $volname =~ m/\.(raw|qcow2|qed|vmdk)$/) {
5967         return $1;
5968     } elsif ($scfg->{type} eq 'iscsi') {
5969         return "host_device";
5970     } else {
5971         return "raw";
5972     }
5973 }
5974
5975 sub qemu_drive_mirror {
5976     my ($vmid, $drive, $dst_volid, $vmiddst) = @_;
5977
5978     my $count = 0;
5979     my $old_len = 0;
5980     my $frozen = undef;
5981     my $maxwait = 120;
5982
5983     my $storecfg = PVE::Storage::config();
5984     my ($dst_storeid, $dst_volname) = PVE::Storage::parse_volume_id($dst_volid);
5985
5986     my $dst_scfg = PVE::Storage::storage_config($storecfg, $dst_storeid);
5987
5988     my $format;
5989     if ($dst_volname =~ m/\.(raw|qcow2)$/){
5990         $format = $1;
5991     }
5992
5993     my $dst_path = PVE::Storage::path($storecfg, $dst_volid);
5994
5995     my $opts = { timeout => 10, device => "drive-$drive", mode => "existing", sync => "full", target => $dst_path };
5996     $opts->{format} = $format if $format;
5997
5998     #fixme : sometime drive-mirror timeout, but works fine after.
5999     # (I have see the problem with big volume > 200GB), so we need to eval
6000     eval { vm_mon_cmd($vmid, "drive-mirror", %$opts); };
6001     # ignore errors here
6002
6003     eval {
6004         while (1) {
6005             my $stats = vm_mon_cmd($vmid, "query-block-jobs");
6006             my $stat = @$stats[0];
6007             die "mirroring job seem to have die. Maybe do you have bad sectors?" if !$stat;
6008             die "error job is not mirroring" if $stat->{type} ne "mirror";
6009
6010             my $busy = $stat->{busy};
6011
6012             if (my $total = $stat->{len}) {
6013                 my $transferred = $stat->{offset} || 0;
6014                 my $remaining = $total - $transferred;
6015                 my $percent = sprintf "%.2f", ($transferred * 100 / $total);
6016
6017                 print "transferred: $transferred bytes remaining: $remaining bytes total: $total bytes progression: $percent % busy: $busy\n";
6018             }
6019
6020             if ($stat->{len} == $stat->{offset}) {
6021                 if ($busy eq 'false') {
6022
6023                     last if $vmiddst != $vmid;
6024
6025                     # try to switch the disk if source and destination are on the same guest
6026                     eval { vm_mon_cmd($vmid, "block-job-complete", device => "drive-$drive") };
6027                     last if !$@;
6028                     die $@ if $@ !~ m/cannot be completed/;
6029                 }
6030
6031                 if ($count > $maxwait) {
6032                     # if too much writes to disk occurs at the end of migration
6033                     #the disk needs to be freezed to be able to complete the migration
6034                     vm_suspend($vmid,1);
6035                     $frozen = 1;
6036                 }
6037                 $count ++
6038             }
6039             $old_len = $stat->{offset};
6040             sleep 1;
6041         }
6042
6043         vm_resume($vmid, 1) if $frozen;
6044
6045     };
6046     my $err = $@;
6047
6048     my $cancel_job = sub {
6049         vm_mon_cmd($vmid, "block-job-cancel", device => "drive-$drive");
6050         while (1) {
6051             my $stats = vm_mon_cmd($vmid, "query-block-jobs");
6052             my $stat = @$stats[0];
6053             last if !$stat;
6054             sleep 1;
6055         }
6056     };
6057
6058     if ($err) {
6059         eval { &$cancel_job(); };
6060         die "mirroring error: $err";
6061     }
6062
6063     if ($vmiddst != $vmid) {
6064         # if we clone a disk for a new target vm, we don't switch the disk
6065         &$cancel_job(); # so we call block-job-cancel
6066     }
6067 }
6068
6069 sub clone_disk {
6070     my ($storecfg, $vmid, $running, $drivename, $drive, $snapname,
6071         $newvmid, $storage, $format, $full, $newvollist) = @_;
6072
6073     my $newvolid;
6074
6075     if (!$full) {
6076         print "create linked clone of drive $drivename ($drive->{file})\n";
6077         $newvolid = PVE::Storage::vdisk_clone($storecfg,  $drive->{file}, $newvmid, $snapname);
6078         push @$newvollist, $newvolid;
6079     } else {
6080         my ($storeid, $volname) = PVE::Storage::parse_volume_id($drive->{file});
6081         $storeid = $storage if $storage;
6082
6083         my ($defFormat, $validFormats) = PVE::Storage::storage_default_format($storecfg, $storeid);
6084         if (!$format) {
6085             $format = $drive->{format} || $defFormat;
6086         }
6087
6088         # test if requested format is supported - else use default
6089         my $supported = grep { $_ eq $format } @$validFormats;
6090         $format = $defFormat if !$supported;
6091
6092         my ($size) = PVE::Storage::volume_size_info($storecfg, $drive->{file}, 3);
6093
6094         print "create full clone of drive $drivename ($drive->{file})\n";
6095         $newvolid = PVE::Storage::vdisk_alloc($storecfg, $storeid, $newvmid, $format, undef, ($size/1024));
6096         push @$newvollist, $newvolid;
6097
6098         if (!$running || $snapname) {
6099             qemu_img_convert($drive->{file}, $newvolid, $size, $snapname);
6100         } else {
6101             qemu_drive_mirror($vmid, $drivename, $newvolid, $newvmid);
6102         }
6103     }
6104
6105     my ($size) = PVE::Storage::volume_size_info($storecfg, $newvolid, 3);
6106
6107     my $disk = $drive;
6108     $disk->{format} = undef;
6109     $disk->{file} = $newvolid;
6110     $disk->{size} = $size;
6111
6112     return $disk;
6113 }
6114
6115 # this only works if VM is running
6116 sub get_current_qemu_machine {
6117     my ($vmid) = @_;
6118
6119     my $cmd = { execute => 'query-machines', arguments => {} };
6120     my $res = vm_qmp_command($vmid, $cmd);
6121
6122     my ($current, $default);
6123     foreach my $e (@$res) {
6124         $default = $e->{name} if $e->{'is-default'};
6125         $current = $e->{name} if $e->{'is-current'};
6126     }
6127
6128     # fallback to the default machine if current is not supported by qemu
6129     return $current || $default || 'pc';
6130 }
6131
6132 sub qemu_machine_feature_enabled {
6133     my ($machine, $kvmver, $version_major, $version_minor) = @_;
6134
6135     my $current_major;
6136     my $current_minor;
6137
6138     if ($machine && $machine =~ m/^(pc(-i440fx|-q35)?-(\d+)\.(\d+))/) {
6139
6140         $current_major = $3;
6141         $current_minor = $4;
6142
6143     } elsif ($kvmver =~ m/^(\d+)\.(\d+)/) {
6144
6145         $current_major = $1;
6146         $current_minor = $2;
6147     }
6148
6149     return 1 if $current_major >= $version_major && $current_minor >= $version_minor;
6150
6151
6152 }
6153
6154 sub lspci {
6155
6156     my $devices = {};
6157
6158     dir_glob_foreach("$pcisysfs/devices", '[a-f0-9]{4}:([a-f0-9]{2}:[a-f0-9]{2})\.([0-9])', sub {
6159             my (undef, $id, $function) = @_;
6160             my $res = { id => $id, function => $function};
6161             push @{$devices->{$id}}, $res;
6162     });
6163
6164     return $devices;
6165 }
6166
6167 1;