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