]> git.proxmox.com Git - pve-common.git/blob - src/PVE/Network.pm
ac49536c27646602c0e60f239beb1dfecf567940
[pve-common.git] / src / PVE / Network.pm
1 package PVE::Network;
2
3 use strict;
4 use warnings;
5
6 use PVE::INotify;
7 use PVE::ProcFSTools;
8 use PVE::Tools qw(run_command lock_file);
9
10 use File::Basename;
11 use IO::Socket::IP;
12 use Net::IP;
13 use POSIX qw(ECONNREFUSED);
14 use Socket qw(NI_NUMERICHOST NI_NUMERICSERV);
15
16 # host network related utility functions
17
18 our $PHYSICAL_NIC_RE = qr/(?:eth\d+|en[^:.]+|ib\d+)/;
19
20 our $ipv4_reverse_mask = [
21 '0.0.0.0',
22 '128.0.0.0',
23 '192.0.0.0',
24 '224.0.0.0',
25 '240.0.0.0',
26 '248.0.0.0',
27 '252.0.0.0',
28 '254.0.0.0',
29 '255.0.0.0',
30 '255.128.0.0',
31 '255.192.0.0',
32 '255.224.0.0',
33 '255.240.0.0',
34 '255.248.0.0',
35 '255.252.0.0',
36 '255.254.0.0',
37 '255.255.0.0',
38 '255.255.128.0',
39 '255.255.192.0',
40 '255.255.224.0',
41 '255.255.240.0',
42 '255.255.248.0',
43 '255.255.252.0',
44 '255.255.254.0',
45 '255.255.255.0',
46 '255.255.255.128',
47 '255.255.255.192',
48 '255.255.255.224',
49 '255.255.255.240',
50 '255.255.255.248',
51 '255.255.255.252',
52 '255.255.255.254',
53 '255.255.255.255',
54 ];
55
56 our $ipv4_mask_hash_localnet = {
57 '255.0.0.0' => 8,
58 '255.128.0.0' => 9,
59 '255.192.0.0' => 10,
60 '255.224.0.0' => 11,
61 '255.240.0.0' => 12,
62 '255.248.0.0' => 13,
63 '255.252.0.0' => 14,
64 '255.254.0.0' => 15,
65 '255.255.0.0' => 16,
66 '255.255.128.0' => 17,
67 '255.255.192.0' => 18,
68 '255.255.224.0' => 19,
69 '255.255.240.0' => 20,
70 '255.255.248.0' => 21,
71 '255.255.252.0' => 22,
72 '255.255.254.0' => 23,
73 '255.255.255.0' => 24,
74 '255.255.255.128' => 25,
75 '255.255.255.192' => 26,
76 '255.255.255.224' => 27,
77 '255.255.255.240' => 28,
78 '255.255.255.248' => 29,
79 '255.255.255.252' => 30,
80 '255.255.255.254' => 31,
81 '255.255.255.255' => 32,
82 };
83
84 sub setup_tc_rate_limit {
85 my ($iface, $rate, $burst) = @_;
86
87 # these are allowed / expected to fail, e.g. when there is no previous rate limit to remove
88 eval { run_command("/sbin/tc class del dev $iface parent 1: classid 1:1 >/dev/null 2>&1"); };
89 eval { run_command("/sbin/tc filter del dev $iface parent ffff: protocol all pref 50 u32 >/dev/null 2>&1"); };
90 eval { run_command("/sbin/tc qdisc del dev $iface ingress >/dev/null 2>&1"); };
91 eval { run_command("/sbin/tc qdisc del dev $iface root >/dev/null 2>&1"); };
92
93 return if !$rate;
94
95 # tbf does not work for unknown reason
96 #$TC qdisc add dev $DEV root tbf rate $RATE latency 100ms burst $BURST
97 # so we use htb instead
98 run_command("/sbin/tc qdisc add dev $iface root handle 1: htb default 1");
99 run_command("/sbin/tc class add dev $iface parent 1: classid 1:1 " .
100 "htb rate ${rate}bps burst ${burst}b");
101
102 run_command("/sbin/tc qdisc add dev $iface handle ffff: ingress");
103 run_command("/sbin/tc filter add dev $iface parent ffff: " .
104 "prio 50 basic " .
105 "police rate ${rate}bps burst ${burst}b mtu 64kb " .
106 "drop");
107 }
108
109 sub tap_rate_limit {
110 my ($iface, $rate) = @_;
111
112 $rate = int($rate*1024*1024) if $rate;
113 my $burst = 1024*1024;
114
115 setup_tc_rate_limit($iface, $rate, $burst);
116 }
117
118 my $read_bridge_mtu = sub {
119 my ($bridge) = @_;
120
121 my $mtu = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/mtu");
122 die "bridge '$bridge' does not exist\n" if !$mtu;
123 # avoid insecure dependency;
124 die "unable to parse mtu value" if $mtu !~ /^(\d+)$/;
125 $mtu = int($1);
126
127 return $mtu;
128 };
129
130 my $parse_tap_device_name = sub {
131 my ($iface, $noerr) = @_;
132
133 my ($vmid, $devid);
134
135 if ($iface =~ m/^tap(\d+)i(\d+)$/) {
136 $vmid = $1;
137 $devid = $2;
138 } elsif ($iface =~ m/^veth(\d+)i(\d+)$/) {
139 $vmid = $1;
140 $devid = $2;
141 } else {
142 return undef if $noerr;
143 die "can't create firewall bridge for random interface name '$iface'\n";
144 }
145
146 return ($vmid, $devid);
147 };
148
149 my $compute_fwbr_names = sub {
150 my ($vmid, $devid) = @_;
151
152 my $fwbr = "fwbr${vmid}i${devid}";
153 # Note: the firewall use 'fwln+' to filter traffic to VMs
154 my $vethfw = "fwln${vmid}i${devid}";
155 my $vethfwpeer = "fwpr${vmid}p${devid}";
156 my $ovsintport = "fwln${vmid}o${devid}";
157
158 return ($fwbr, $vethfw, $vethfwpeer, $ovsintport);
159 };
160
161 sub iface_delete($) {
162 my ($iface) = @_;
163 run_command(['/sbin/ip', 'link', 'delete', 'dev', $iface], noerr => 1)
164 == 0 or die "failed to delete interface '$iface'\n";
165 }
166
167 sub iface_create($$@) {
168 my ($iface, $type, @args) = @_;
169 run_command(['/sbin/ip', 'link', 'add', $iface, 'type', $type, @args], noerr => 1)
170 == 0 or die "failed to create interface '$iface'\n";
171 }
172
173 sub iface_set($@) {
174 my ($iface, @opts) = @_;
175 run_command(['/sbin/ip', 'link', 'set', $iface, @opts], noerr => 1)
176 == 0 or die "failed to set interface options for '$iface' (".join(' ', @opts).")\n";
177 }
178
179 # helper for nicer error messages:
180 sub iface_set_master($$) {
181 my ($iface, $master) = @_;
182 if (defined($master)) {
183 eval { iface_set($iface, 'master', $master) };
184 die "can't enslave '$iface' to '$master'\n" if $@;
185 } else {
186 eval { iface_set($iface, 'nomaster') };
187 die "can't unenslave '$iface'\n" if $@;
188 }
189 }
190
191 my $cond_create_bridge = sub {
192 my ($bridge) = @_;
193
194 if (! -d "/sys/class/net/$bridge") {
195 iface_create($bridge, 'bridge');
196 disable_ipv6($bridge);
197 }
198 };
199
200 sub disable_ipv6 {
201 my ($iface) = @_;
202 return if !-d '/proc/sys/net/ipv6'; # ipv6 might be completely disabled
203 my $file = "/proc/sys/net/ipv6/conf/$iface/disable_ipv6";
204 open(my $fh, '>', $file) or die "failed to open $file for writing: $!\n";
205 print {$fh} "1\n" or die "failed to disable link-local ipv6 for $iface\n";
206 close($fh);
207 }
208
209 my $bridge_add_interface = sub {
210 my ($bridge, $iface, $tag, $trunks) = @_;
211
212 # drop link local address (it can't be used when on a bridge anyway)
213 disable_ipv6($iface);
214 iface_set_master($iface, $bridge);
215
216 my $vlan_aware = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/bridge/vlan_filtering");
217
218 if ($vlan_aware) {
219 if ($tag) {
220 eval { run_command(['/sbin/bridge', 'vlan', 'del', 'dev', $iface, 'vid', '1-4094']) };
221 die "failed to remove default vlan tags of $iface - $@\n" if $@;
222
223 eval { run_command(['/sbin/bridge', 'vlan', 'add', 'dev', $iface, 'vid', $tag, 'pvid', 'untagged']) };
224 die "unable to add vlan $tag to interface $iface - $@\n" if $@;
225
226 warn "Caution: Setting VLAN ID 1 on a VLAN aware bridge may be dangerous\n" if $tag == 1;
227 } else {
228 eval { run_command(['/sbin/bridge', 'vlan', 'add', 'dev', $iface, 'vid', '2-4094']) };
229 die "unable to add default vlan tags to interface $iface - $@\n"
230 if $@ && !$trunks;
231 }
232
233 if ($trunks) {
234 my @trunks_array = split /;/, $trunks;
235 foreach my $trunk (@trunks_array) {
236 eval { run_command(['/sbin/bridge', 'vlan', 'add', 'dev', $iface, 'vid', $trunk]) };
237 die "unable to add vlan $trunk to interface $iface - $@\n" if $@;
238 }
239 }
240 }
241 };
242
243 my $ovs_bridge_add_port = sub {
244 my ($bridge, $iface, $tag, $internal, $trunks) = @_;
245
246 $trunks =~ s/;/,/g if $trunks;
247
248 my $cmd = ['/usr/bin/ovs-vsctl'];
249 # first command
250 push @$cmd, '--', 'add-port', $bridge, $iface;
251 push @$cmd, "tag=$tag" if $tag;
252 push @$cmd, "trunks=". join(',', $trunks) if $trunks;
253 push @$cmd, "vlan_mode=native-untagged" if $tag && $trunks;
254
255 if ($internal) {
256 # second command
257 push @$cmd, '--', 'set', 'Interface', $iface, 'type=internal';
258 }
259
260 eval { run_command($cmd) };
261 die "can't add ovs port '$iface' - $@\n" if $@;
262
263 disable_ipv6($iface);
264 };
265
266 my $activate_interface = sub {
267 my ($iface) = @_;
268
269 eval { run_command(['/sbin/ip', 'link', 'set', $iface, 'up']) };
270 die "can't activate interface '$iface' - $@\n" if $@;
271 };
272
273 sub tap_create {
274 my ($iface, $bridge) = @_;
275
276 die "unable to get bridge setting\n" if !$bridge;
277
278 my $bridgemtu = &$read_bridge_mtu($bridge);
279
280 eval {
281 disable_ipv6($iface);
282 PVE::Tools::run_command(['/sbin/ip', 'link', 'set', $iface, 'up', 'promisc', 'on', 'mtu', $bridgemtu]);
283 };
284 die "interface activation failed\n" if $@;
285 }
286
287 sub veth_create {
288 my ($veth, $vethpeer, $bridge, $mac) = @_;
289
290 die "unable to get bridge setting\n" if !$bridge;
291
292 my $bridgemtu = &$read_bridge_mtu($bridge);
293
294 # create veth pair
295 if (! -d "/sys/class/net/$veth") {
296 my $cmd = ['/sbin/ip', 'link', 'add'];
297 # veth device + MTU
298 push @$cmd, 'name', $veth;
299 push @$cmd, 'mtu', $bridgemtu;
300 push @$cmd, 'type', 'veth';
301 # peer device + MTU
302 push @$cmd, 'peer', 'name', $vethpeer, 'mtu', $bridgemtu;
303
304 push @$cmd, 'addr', $mac if $mac;
305
306 eval { run_command($cmd) };
307 die "can't create interface $veth - $@\n" if $@;
308 }
309
310 # up vethpair
311 disable_ipv6($veth);
312 disable_ipv6($vethpeer);
313 &$activate_interface($veth);
314 &$activate_interface($vethpeer);
315 }
316
317 sub veth_delete {
318 my ($veth) = @_;
319
320 if (-d "/sys/class/net/$veth") {
321 iface_delete($veth);
322 }
323 eval { tap_unplug($veth) };
324 }
325
326 my $create_firewall_bridge_linux = sub {
327 my ($iface, $bridge, $tag, $trunks) = @_;
328
329 my ($vmid, $devid) = &$parse_tap_device_name($iface);
330 my ($fwbr, $vethfw, $vethfwpeer) = &$compute_fwbr_names($vmid, $devid);
331
332 &$cond_create_bridge($fwbr);
333 &$activate_interface($fwbr);
334
335 copy_bridge_config($bridge, $fwbr);
336 veth_create($vethfw, $vethfwpeer, $bridge);
337
338 &$bridge_add_interface($fwbr, $vethfw);
339 &$bridge_add_interface($bridge, $vethfwpeer, $tag, $trunks);
340
341 &$bridge_add_interface($fwbr, $iface);
342 };
343
344 my $create_firewall_bridge_ovs = sub {
345 my ($iface, $bridge, $tag, $trunks) = @_;
346
347 my ($vmid, $devid) = &$parse_tap_device_name($iface);
348 my ($fwbr, undef, undef, $ovsintport) = &$compute_fwbr_names($vmid, $devid);
349
350 my $bridgemtu = &$read_bridge_mtu($bridge);
351
352 &$cond_create_bridge($fwbr);
353 &$activate_interface($fwbr);
354
355 &$bridge_add_interface($fwbr, $iface);
356
357 &$ovs_bridge_add_port($bridge, $ovsintport, $tag, 1, $trunks);
358 &$activate_interface($ovsintport);
359
360 # set the same mtu for ovs int port
361 PVE::Tools::run_command(['/sbin/ip', 'link', 'set', $ovsintport, 'mtu', $bridgemtu]);
362
363 &$bridge_add_interface($fwbr, $ovsintport);
364 };
365
366 my $cleanup_firewall_bridge = sub {
367 my ($iface) = @_;
368
369 my ($vmid, $devid) = &$parse_tap_device_name($iface, 1);
370 return if !defined($vmid);
371 my ($fwbr, $vethfw, $vethfwpeer, $ovsintport) = &$compute_fwbr_names($vmid, $devid);
372
373 # cleanup old port config from any openvswitch bridge
374 if (-d "/sys/class/net/$ovsintport") {
375 run_command("/usr/bin/ovs-vsctl del-port $ovsintport", outfunc => sub {}, errfunc => sub {});
376 }
377
378 # delete old vethfw interface
379 veth_delete($vethfw);
380
381 # cleanup fwbr bridge
382 if (-d "/sys/class/net/$fwbr") {
383 iface_delete($fwbr);
384 }
385 };
386
387 sub tap_plug {
388 my ($iface, $bridge, $tag, $firewall, $trunks, $rate) = @_;
389
390 #cleanup old port config from any openvswitch bridge
391 eval {run_command("/usr/bin/ovs-vsctl del-port $iface", outfunc => sub {}, errfunc => sub {}) };
392
393 if (-d "/sys/class/net/$bridge/bridge") {
394 &$cleanup_firewall_bridge($iface); # remove stale devices
395
396 my $vlan_aware = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/bridge/vlan_filtering");
397
398 if (!$vlan_aware) {
399 die "vlan aware feature need to be enabled to use trunks" if $trunks;
400 my $newbridge = activate_bridge_vlan($bridge, $tag);
401 copy_bridge_config($bridge, $newbridge) if $bridge ne $newbridge;
402 $bridge = $newbridge;
403 $tag = undef;
404 }
405
406 if ($firewall) {
407 &$create_firewall_bridge_linux($iface, $bridge, $tag, $trunks);
408 } else {
409 &$bridge_add_interface($bridge, $iface, $tag, $trunks);
410 }
411
412 } else {
413 &$cleanup_firewall_bridge($iface); # remove stale devices
414
415 if ($firewall) {
416 &$create_firewall_bridge_ovs($iface, $bridge, $tag, $trunks);
417 } else {
418 &$ovs_bridge_add_port($bridge, $iface, $tag, undef, $trunks);
419 }
420 }
421
422 tap_rate_limit($iface, $rate);
423 }
424
425 sub tap_unplug {
426 my ($iface) = @_;
427
428 my $path= "/sys/class/net/$iface/brport/bridge";
429 if (-l $path) {
430 my $bridge = basename(readlink($path));
431 #avoid insecure dependency;
432 ($bridge) = $bridge =~ /(\S+)/;
433
434 iface_set_master($iface, undef);
435 }
436
437 &$cleanup_firewall_bridge($iface);
438 #cleanup old port config from any openvswitch bridge
439 eval {run_command("/usr/bin/ovs-vsctl del-port $iface", outfunc => sub {}, errfunc => sub {}) };
440 }
441
442 sub copy_bridge_config {
443 my ($br0, $br1) = @_;
444
445 return if $br0 eq $br1;
446
447 my $br_configs = [ 'ageing_time', 'stp_state', 'priority', 'forward_delay',
448 'hello_time', 'max_age', 'multicast_snooping', 'multicast_querier'];
449
450 foreach my $sysname (@$br_configs) {
451 eval {
452 my $v0 = PVE::Tools::file_read_firstline("/sys/class/net/$br0/bridge/$sysname");
453 my $v1 = PVE::Tools::file_read_firstline("/sys/class/net/$br1/bridge/$sysname");
454 if ($v0 ne $v1) {
455 PVE::ProcFSTools::write_proc_entry("/sys/class/net/$br1/bridge/$sysname", $v0);
456 }
457 };
458 warn $@ if $@;
459 }
460 }
461
462 sub activate_bridge_vlan_slave {
463 my ($bridgevlan, $iface, $tag) = @_;
464 my $ifacevlan = "${iface}.$tag";
465
466 # create vlan on $iface is not already exist
467 if (! -d "/sys/class/net/$ifacevlan") {
468 eval {
469 my $cmd = ['/sbin/ip', 'link', 'add'];
470 push @$cmd, 'link', $iface;
471 push @$cmd, 'name', $ifacevlan;
472 push @$cmd, 'type', 'vlan', 'id', $tag;
473 run_command($cmd);
474 };
475 die "can't add vlan tag $tag to interface $iface - $@\n" if $@;
476
477 # remove ipv6 link-local address before activation
478 disable_ipv6($ifacevlan);
479 }
480
481 # be sure to have the $ifacevlan up
482 &$activate_interface($ifacevlan);
483
484 # test if $vlaniface is already enslaved in another bridge
485 my $path= "/sys/class/net/$ifacevlan/brport/bridge";
486 if (-l $path) {
487 my $tbridge = basename(readlink($path));
488 if ($tbridge ne $bridgevlan) {
489 die "interface $ifacevlan already exist in bridge $tbridge\n";
490 } else {
491 # Port already attached to bridge: do nothing.
492 return;
493 }
494 }
495
496 # add $ifacevlan to the bridge
497 &$bridge_add_interface($bridgevlan, $ifacevlan);
498 }
499
500 sub activate_bridge_vlan {
501 my ($bridge, $tag_param) = @_;
502
503 die "bridge '$bridge' is not active\n" if ! -d "/sys/class/net/$bridge";
504
505 return $bridge if !defined($tag_param); # no vlan, simply return
506
507 my $tag = int($tag_param);
508
509 die "got strange vlan tag '$tag_param'\n" if $tag < 1 || $tag > 4094;
510
511 my $bridgevlan = "${bridge}v$tag";
512
513 my @ifaces = ();
514 my $dir = "/sys/class/net/$bridge/brif";
515 PVE::Tools::dir_glob_foreach($dir, '(((eth|bond)\d+|en[^.]+)(\.\d+)?)', sub {
516 push @ifaces, $_[0];
517 });
518
519 die "no physical interface on bridge '$bridge'\n" if scalar(@ifaces) == 0;
520
521 lock_network(sub {
522 # add bridgevlan if it doesn't already exist
523 if (! -d "/sys/class/net/$bridgevlan") {
524 iface_create($bridgevlan, 'bridge');
525 }
526
527 # for each physical interface (eth or bridge) bind them to bridge vlan
528 foreach my $iface (@ifaces) {
529 activate_bridge_vlan_slave($bridgevlan, $iface, $tag);
530 }
531
532 #fixme: set other bridge flags
533
534 # remove ipv6 link-local address before activation
535 disable_ipv6($bridgevlan);
536 # be sure to have the bridge up
537 &$activate_interface($bridgevlan);
538 });
539 return $bridgevlan;
540 }
541
542 sub tcp_ping {
543 my ($host, $port, $timeout) = @_;
544
545 my $refused = 1;
546
547 $timeout = 3 if !$timeout; # sane default
548 if (!$port) {
549 # Net::Ping defaults to the echo port
550 $port = 7;
551 } else {
552 # Net::Ping's port_number() implies service_check(1)
553 $refused = 0;
554 }
555
556 my ($sock, $result);
557 eval {
558 $result = PVE::Tools::run_with_timeout($timeout, sub {
559 $sock = IO::Socket::IP->new(PeerHost => $host, PeerPort => $port, Type => SOCK_STREAM);
560 $result = $refused if $! == ECONNREFUSED;
561 });
562 };
563 if ($sock) {
564 $sock->close();
565 $result = 1;
566 }
567 return $result;
568 }
569
570 sub IP_from_cidr {
571 my ($cidr, $version) = @_;
572
573 return if $cidr !~ m!^(\S+?)/(\S+)$!;
574 my ($ip, $prefix) = ($1, $2);
575
576 my $ipobj = Net::IP->new($ip, $version);
577 return if !$ipobj;
578
579 $version = $ipobj->version();
580
581 my $binmask = Net::IP::ip_get_mask($prefix, $version);
582 return if !$binmask;
583
584 my $masked_binip = $ipobj->binip() & $binmask;
585 my $masked_ip = Net::IP::ip_bintoip($masked_binip, $version);
586 return Net::IP->new("$masked_ip/$prefix");
587 }
588
589 sub is_ip_in_cidr {
590 my ($ip, $cidr, $version) = @_;
591
592 my $cidr_obj = IP_from_cidr($cidr, $version);
593 return undef if !$cidr_obj;
594
595 my $ip_obj = Net::IP->new($ip, $version);
596 return undef if !$ip_obj;
597
598 return $cidr_obj->overlaps($ip_obj) == $Net::IP::IP_B_IN_A_OVERLAP;
599 }
600
601
602 sub get_local_ip_from_cidr {
603 my ($cidr) = @_;
604
605 my $IPs = [];
606 run_command(['/sbin/ip', 'address', 'show', 'to', $cidr, 'up'], outfunc => sub {
607 if ($_[0] =~ m!^\s*inet(?:6)?\s+($PVE::Tools::IPRE)(?:/\d+|\s+peer\s+)!) {
608 push @$IPs, $1;
609 }
610 });
611
612 return $IPs;
613 }
614
615 sub addr_to_ip {
616 my ($addr) = @_;
617 my ($err, $host, $port) = Socket::getnameinfo($addr, NI_NUMERICHOST | NI_NUMERICSERV);
618 die "failed to get numerical host address: $err\n" if $err;
619 return ($host, $port) if wantarray;
620 return $host;
621 }
622
623 sub get_ip_from_hostname {
624 my ($hostname, $noerr) = @_;
625
626 my @res = eval { PVE::Tools::getaddrinfo_all($hostname) };
627 if ($@) {
628 die "hostname lookup '$hostname' failed - $@" if !$noerr;
629 return undef;
630 }
631
632 my ($ip, $family);
633 for my $ai (@res) {
634 $family = $ai->{family};
635 my $tmpip = addr_to_ip($ai->{addr});
636 if ($tmpip !~ m/^127\.|^::1$/) {
637 $ip = $tmpip;
638 last;
639 }
640 }
641 if (!defined($ip) ) {
642 die "hostname lookup '$hostname' failed - got local IP address '$ip'\n" if !$noerr;
643 return undef;
644 }
645
646 return wantarray ? ($ip, $family) : $ip;
647 }
648
649 sub lock_network {
650 my ($code, @param) = @_;
651 my $res = lock_file('/var/lock/pve-network.lck', 10, $code, @param);
652 die $@ if $@;
653 return $res;
654 }
655
656 1;