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