]> git.proxmox.com Git - pve-common.git/blob - src/PVE/Network.pm
network: followup: move graceful require SDN out from subs
[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 my $have_sdn;
17 eval {
18 require PVE::Network::SDN::Zones;
19 $have_sdn = 1;
20 };
21
22 # host network related utility functions
23
24 our $PHYSICAL_NIC_RE = qr/(?:eth\d+|en[^:.]+|ib\d+)/;
25
26 our $ipv4_reverse_mask = [
27 '0.0.0.0',
28 '128.0.0.0',
29 '192.0.0.0',
30 '224.0.0.0',
31 '240.0.0.0',
32 '248.0.0.0',
33 '252.0.0.0',
34 '254.0.0.0',
35 '255.0.0.0',
36 '255.128.0.0',
37 '255.192.0.0',
38 '255.224.0.0',
39 '255.240.0.0',
40 '255.248.0.0',
41 '255.252.0.0',
42 '255.254.0.0',
43 '255.255.0.0',
44 '255.255.128.0',
45 '255.255.192.0',
46 '255.255.224.0',
47 '255.255.240.0',
48 '255.255.248.0',
49 '255.255.252.0',
50 '255.255.254.0',
51 '255.255.255.0',
52 '255.255.255.128',
53 '255.255.255.192',
54 '255.255.255.224',
55 '255.255.255.240',
56 '255.255.255.248',
57 '255.255.255.252',
58 '255.255.255.254',
59 '255.255.255.255',
60 ];
61
62 our $ipv4_mask_hash_localnet = {
63 '255.0.0.0' => 8,
64 '255.128.0.0' => 9,
65 '255.192.0.0' => 10,
66 '255.224.0.0' => 11,
67 '255.240.0.0' => 12,
68 '255.248.0.0' => 13,
69 '255.252.0.0' => 14,
70 '255.254.0.0' => 15,
71 '255.255.0.0' => 16,
72 '255.255.128.0' => 17,
73 '255.255.192.0' => 18,
74 '255.255.224.0' => 19,
75 '255.255.240.0' => 20,
76 '255.255.248.0' => 21,
77 '255.255.252.0' => 22,
78 '255.255.254.0' => 23,
79 '255.255.255.0' => 24,
80 '255.255.255.128' => 25,
81 '255.255.255.192' => 26,
82 '255.255.255.224' => 27,
83 '255.255.255.240' => 28,
84 '255.255.255.248' => 29,
85 '255.255.255.252' => 30,
86 '255.255.255.254' => 31,
87 '255.255.255.255' => 32,
88 };
89
90 sub setup_tc_rate_limit {
91 my ($iface, $rate, $burst, $debug) = @_;
92
93 # these are allowed / expected to fail, e.g. when there is no previous rate limit to remove
94 eval { run_command("/sbin/tc class del dev $iface parent 1: classid 1:1 >/dev/null 2>&1"); };
95 eval { run_command("/sbin/tc filter del dev $iface parent ffff: protocol all pref 50 u32 >/dev/null 2>&1"); };
96 eval { run_command("/sbin/tc qdisc del dev $iface ingress >/dev/null 2>&1"); };
97 eval { run_command("/sbin/tc qdisc del dev $iface root >/dev/null 2>&1"); };
98
99 return if !$rate;
100
101 # tbf does not work for unknown reason
102 #$TC qdisc add dev $DEV root tbf rate $RATE latency 100ms burst $BURST
103 # so we use htb instead
104 run_command("/sbin/tc qdisc add dev $iface root handle 1: htb default 1");
105 run_command("/sbin/tc class add dev $iface parent 1: classid 1:1 " .
106 "htb rate ${rate}bps burst ${burst}b");
107
108 run_command("/sbin/tc qdisc add dev $iface handle ffff: ingress");
109 run_command("/sbin/tc filter add dev $iface parent ffff: " .
110 "prio 50 basic " .
111 "police rate ${rate}bps burst ${burst}b mtu 64kb " .
112 "drop");
113
114 if ($debug) {
115 print "DEBUG tc settings\n";
116 system("/sbin/tc qdisc ls dev $iface");
117 system("/sbin/tc class ls dev $iface");
118 system("/sbin/tc filter ls dev $iface parent ffff:");
119 }
120 }
121
122 sub tap_rate_limit {
123 my ($iface, $rate) = @_;
124
125 my $debug = 0;
126 $rate = int($rate*1024*1024) if $rate;
127 my $burst = 1024*1024;
128
129 setup_tc_rate_limit($iface, $rate, $burst, $debug);
130 }
131
132 my $read_bridge_mtu = sub {
133 my ($bridge) = @_;
134
135 my $mtu = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/mtu");
136 die "bridge '$bridge' does not exist\n" if !$mtu;
137 # avoid insecure dependency;
138 die "unable to parse mtu value" if $mtu !~ /^(\d+)$/;
139 $mtu = int($1);
140
141 return $mtu;
142 };
143
144 my $parse_tap_device_name = sub {
145 my ($iface, $noerr) = @_;
146
147 my ($vmid, $devid);
148
149 if ($iface =~ m/^tap(\d+)i(\d+)$/) {
150 $vmid = $1;
151 $devid = $2;
152 } elsif ($iface =~ m/^veth(\d+)i(\d+)$/) {
153 $vmid = $1;
154 $devid = $2;
155 } else {
156 return undef if $noerr;
157 die "can't create firewall bridge for random interface name '$iface'\n";
158 }
159
160 return ($vmid, $devid);
161 };
162
163 my $compute_fwbr_names = sub {
164 my ($vmid, $devid) = @_;
165
166 my $fwbr = "fwbr${vmid}i${devid}";
167 # Note: the firewall use 'fwln+' to filter traffic to VMs
168 my $vethfw = "fwln${vmid}i${devid}";
169 my $vethfwpeer = "fwpr${vmid}p${devid}";
170 my $ovsintport = "fwln${vmid}o${devid}";
171
172 return ($fwbr, $vethfw, $vethfwpeer, $ovsintport);
173 };
174
175 sub iface_delete($) {
176 my ($iface) = @_;
177 run_command(['/sbin/ip', 'link', 'delete', 'dev', $iface], noerr => 1)
178 == 0 or die "failed to delete interface '$iface'\n";
179 }
180
181 sub iface_create($$@) {
182 my ($iface, $type, @args) = @_;
183 run_command(['/sbin/ip', 'link', 'add', $iface, 'type', $type, @args], noerr => 1)
184 == 0 or die "failed to create interface '$iface'\n";
185 }
186
187 sub iface_set($@) {
188 my ($iface, @opts) = @_;
189 run_command(['/sbin/ip', 'link', 'set', $iface, @opts], noerr => 1)
190 == 0 or die "failed to set interface options for '$iface' (".join(' ', @opts).")\n";
191 }
192
193 # helper for nicer error messages:
194 sub iface_set_master($$) {
195 my ($iface, $master) = @_;
196 if (defined($master)) {
197 eval { iface_set($iface, 'master', $master) };
198 die "can't enslave '$iface' to '$master'\n" if $@;
199 } else {
200 eval { iface_set($iface, 'nomaster') };
201 die "can't unenslave '$iface'\n" if $@;
202 }
203 }
204
205 my $cond_create_bridge = sub {
206 my ($bridge) = @_;
207
208 if (! -d "/sys/class/net/$bridge") {
209 iface_create($bridge, 'bridge');
210 disable_ipv6($bridge);
211 }
212 };
213
214 sub disable_ipv6 {
215 my ($iface) = @_;
216 return if !-d '/proc/sys/net/ipv6'; # ipv6 might be completely disabled
217 my $file = "/proc/sys/net/ipv6/conf/$iface/disable_ipv6";
218 open(my $fh, '>', $file) or die "failed to open $file for writing: $!\n";
219 print {$fh} "1\n" or die "failed to disable link-local ipv6 for $iface\n";
220 close($fh);
221 }
222
223 my $bridge_add_interface = sub {
224 my ($bridge, $iface, $tag, $trunks) = @_;
225
226 # drop link local address (it can't be used when on a bridge anyway)
227 disable_ipv6($iface);
228 iface_set_master($iface, $bridge);
229
230 my $vlan_aware = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/bridge/vlan_filtering");
231
232 if ($vlan_aware) {
233 if ($tag) {
234 system({'/sbin/bridge'} 'bridge', 'vlan', 'del', 'dev', $iface, 'vid', '1-4094') == 0
235 or die "failed to remove default vlan tags of $iface\n";
236 system({'/sbin/bridge'} 'bridge', 'vlan', 'add', 'dev', $iface, 'vid', $tag, 'pvid', 'untagged') == 0
237 or die "unable to add vlan $tag to interface $iface\n";
238
239 warn "Caution: Setting VLAN ID 1 on a VLAN aware bridge may be dangerous\n" if $tag == 1;
240 } else {
241 system("/sbin/bridge vlan add dev $iface vid 2-4094") == 0 ||
242 die "unable to add default vlan tags to interface $iface\n" if !$trunks;
243 }
244
245 if ($trunks) {
246 my @trunks_array = split /;/, $trunks;
247 foreach my $trunk (@trunks_array) {
248 system("/sbin/bridge vlan add dev $iface vid $trunk") == 0 ||
249 die "unable to add vlan $trunk to interface $iface\n";
250 }
251 }
252 }
253 };
254
255 my $ovs_bridge_add_port = sub {
256 my ($bridge, $iface, $tag, $internal, $trunks) = @_;
257
258 $trunks =~ s/;/,/g if $trunks;
259
260 my $cmd = "/usr/bin/ovs-vsctl add-port $bridge $iface";
261 $cmd .= " tag=$tag" if $tag;
262 $cmd .= " trunks=". join(',', $trunks) if $trunks;
263 $cmd .= " vlan_mode=native-untagged" if $tag && $trunks;
264
265 $cmd .= " -- set Interface $iface type=internal" if $internal;
266 system($cmd) == 0 ||
267 die "can't add ovs port '$iface'\n";
268 disable_ipv6($iface);
269 };
270
271 my $activate_interface = sub {
272 my ($iface) = @_;
273
274 system("/sbin/ip link set $iface up") == 0 ||
275 die "can't activate interface '$iface'\n";
276 };
277
278 sub tap_create {
279 my ($iface, $bridge) = @_;
280
281 if ($have_sdn) {
282 my ($bridgesdn, undef) = PVE::Network::SDN::Zones::get_bridge_vlan($bridge);
283 $bridge = $bridgesdn if $bridgesdn;
284 }
285
286 die "unable to get bridge setting\n" if !$bridge;
287
288 my $bridgemtu = &$read_bridge_mtu($bridge);
289
290 eval {
291 disable_ipv6($iface);
292 PVE::Tools::run_command(['/sbin/ip', 'link', 'set', $iface, 'up', 'promisc', 'on', 'mtu', $bridgemtu]);
293 };
294 die "interface activation failed\n" if $@;
295 }
296
297 sub veth_create {
298 my ($veth, $vethpeer, $bridge, $mac) = @_;
299
300 if ($have_sdn) {
301 my ($bridgesdn, undef) = PVE::Network::SDN::Zones::get_bridge_vlan($bridge);
302 $bridge = $bridgesdn if $bridgesdn;
303 }
304
305 die "unable to get bridge setting\n" if !$bridge;
306
307 my $bridgemtu = &$read_bridge_mtu($bridge);
308
309 # create veth pair
310 if (! -d "/sys/class/net/$veth") {
311 my $cmd = "/sbin/ip link add name $veth mtu $bridgemtu type veth peer name $vethpeer mtu $bridgemtu";
312 $cmd .= " addr $mac" if $mac;
313 system($cmd) == 0 || die "can't create interface $veth\n";
314 }
315
316 # up vethpair
317 disable_ipv6($veth);
318 disable_ipv6($vethpeer);
319 &$activate_interface($veth);
320 &$activate_interface($vethpeer);
321 }
322
323 sub veth_delete {
324 my ($veth) = @_;
325
326 if (-d "/sys/class/net/$veth") {
327 iface_delete($veth);
328 }
329 eval { tap_unplug($veth) };
330 }
331
332 my $create_firewall_bridge_linux = sub {
333 my ($iface, $bridge, $tag, $trunks) = @_;
334
335 my ($vmid, $devid) = &$parse_tap_device_name($iface);
336 my ($fwbr, $vethfw, $vethfwpeer) = &$compute_fwbr_names($vmid, $devid);
337
338 &$cond_create_bridge($fwbr);
339 &$activate_interface($fwbr);
340
341 copy_bridge_config($bridge, $fwbr);
342 veth_create($vethfw, $vethfwpeer, $bridge);
343
344 &$bridge_add_interface($fwbr, $vethfw);
345 &$bridge_add_interface($bridge, $vethfwpeer, $tag, $trunks);
346
347 &$bridge_add_interface($fwbr, $iface);
348 };
349
350 my $create_firewall_bridge_ovs = sub {
351 my ($iface, $bridge, $tag, $trunks) = @_;
352
353 my ($vmid, $devid) = &$parse_tap_device_name($iface);
354 my ($fwbr, undef, undef, $ovsintport) = &$compute_fwbr_names($vmid, $devid);
355
356 my $bridgemtu = &$read_bridge_mtu($bridge);
357
358 &$cond_create_bridge($fwbr);
359 &$activate_interface($fwbr);
360
361 &$bridge_add_interface($fwbr, $iface);
362
363 &$ovs_bridge_add_port($bridge, $ovsintport, $tag, 1, $trunks);
364 &$activate_interface($ovsintport);
365
366 # set the same mtu for ovs int port
367 PVE::Tools::run_command(['/sbin/ip', 'link', 'set', $ovsintport, 'mtu', $bridgemtu]);
368
369 &$bridge_add_interface($fwbr, $ovsintport);
370 };
371
372 my $cleanup_firewall_bridge = sub {
373 my ($iface) = @_;
374
375 my ($vmid, $devid) = &$parse_tap_device_name($iface, 1);
376 return if !defined($vmid);
377 my ($fwbr, $vethfw, $vethfwpeer, $ovsintport) = &$compute_fwbr_names($vmid, $devid);
378
379 # cleanup old port config from any openvswitch bridge
380 if (-d "/sys/class/net/$ovsintport") {
381 run_command("/usr/bin/ovs-vsctl del-port $ovsintport", outfunc => sub {}, errfunc => sub {});
382 }
383
384 # delete old vethfw interface
385 veth_delete($vethfw);
386
387 # cleanup fwbr bridge
388 if (-d "/sys/class/net/$fwbr") {
389 iface_delete($fwbr);
390 }
391 };
392
393 sub tap_plug {
394 my ($iface, $bridge, $tag, $firewall, $trunks, $rate) = @_;
395
396 if ($have_sdn) {
397 my ($bridgesdn, $tag) = PVE::Network::SDN::Zones::get_bridge_vlan($bridge);
398 $bridge = $bridgesdn if $bridgesdn;
399 $trunks = undef;
400 }
401
402 #cleanup old port config from any openvswitch bridge
403 eval {run_command("/usr/bin/ovs-vsctl del-port $iface", outfunc => sub {}, errfunc => sub {}) };
404
405 if (-d "/sys/class/net/$bridge/bridge") {
406 &$cleanup_firewall_bridge($iface); # remove stale devices
407
408 my $vlan_aware = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/bridge/vlan_filtering");
409
410 if (!$vlan_aware) {
411 die "vlan aware feature need to be enabled to use trunks" if $trunks;
412 my $newbridge = activate_bridge_vlan($bridge, $tag);
413 copy_bridge_config($bridge, $newbridge) if $bridge ne $newbridge;
414 $bridge = $newbridge;
415 $tag = undef;
416 }
417
418 if ($firewall) {
419 &$create_firewall_bridge_linux($iface, $bridge, $tag, $trunks);
420 } else {
421 &$bridge_add_interface($bridge, $iface, $tag, $trunks);
422 }
423
424 } else {
425 &$cleanup_firewall_bridge($iface); # remove stale devices
426
427 if ($firewall) {
428 &$create_firewall_bridge_ovs($iface, $bridge, $tag, $trunks);
429 } else {
430 &$ovs_bridge_add_port($bridge, $iface, $tag, undef, $trunks);
431 }
432 }
433
434 tap_rate_limit($iface, $rate);
435 }
436
437 sub tap_unplug {
438 my ($iface) = @_;
439
440 my $path= "/sys/class/net/$iface/brport/bridge";
441 if (-l $path) {
442 my $bridge = basename(readlink($path));
443 #avoid insecure dependency;
444 ($bridge) = $bridge =~ /(\S+)/;
445
446 iface_set_master($iface, undef);
447 }
448
449 &$cleanup_firewall_bridge($iface);
450 #cleanup old port config from any openvswitch bridge
451 eval {run_command("/usr/bin/ovs-vsctl del-port $iface", outfunc => sub {}, errfunc => sub {}) };
452 }
453
454 sub copy_bridge_config {
455 my ($br0, $br1) = @_;
456
457 return if $br0 eq $br1;
458
459 my $br_configs = [ 'ageing_time', 'stp_state', 'priority', 'forward_delay',
460 'hello_time', 'max_age', 'multicast_snooping', 'multicast_querier'];
461
462 foreach my $sysname (@$br_configs) {
463 eval {
464 my $v0 = PVE::Tools::file_read_firstline("/sys/class/net/$br0/bridge/$sysname");
465 my $v1 = PVE::Tools::file_read_firstline("/sys/class/net/$br1/bridge/$sysname");
466 if ($v0 ne $v1) {
467 PVE::ProcFSTools::write_proc_entry("/sys/class/net/$br1/bridge/$sysname", $v0);
468 }
469 };
470 warn $@ if $@;
471 }
472 }
473
474 sub activate_bridge_vlan_slave {
475 my ($bridgevlan, $iface, $tag) = @_;
476 my $ifacevlan = "${iface}.$tag";
477
478 # create vlan on $iface is not already exist
479 if (! -d "/sys/class/net/$ifacevlan") {
480 system("/sbin/ip link add link $iface name $ifacevlan type vlan id $tag") == 0 ||
481 die "can't add vlan tag $tag to interface $iface\n";
482
483 # remove ipv6 link-local address before activation
484 disable_ipv6($ifacevlan);
485 }
486
487 # be sure to have the $ifacevlan up
488 &$activate_interface($ifacevlan);
489
490 # test if $vlaniface is already enslaved in another bridge
491 my $path= "/sys/class/net/$ifacevlan/brport/bridge";
492 if (-l $path) {
493 my $tbridge = basename(readlink($path));
494 if ($tbridge ne $bridgevlan) {
495 die "interface $ifacevlan already exist in bridge $tbridge\n";
496 } else {
497 # Port already attached to bridge: do nothing.
498 return;
499 }
500 }
501
502 # add $ifacevlan to the bridge
503 &$bridge_add_interface($bridgevlan, $ifacevlan);
504 }
505
506 sub activate_bridge_vlan {
507 my ($bridge, $tag_param) = @_;
508
509 die "bridge '$bridge' is not active\n" if ! -d "/sys/class/net/$bridge";
510
511 return $bridge if !defined($tag_param); # no vlan, simply return
512
513 my $tag = int($tag_param);
514
515 die "got strange vlan tag '$tag_param'\n" if $tag < 1 || $tag > 4094;
516
517 my $bridgevlan = "${bridge}v$tag";
518
519 my @ifaces = ();
520 my $dir = "/sys/class/net/$bridge/brif";
521 PVE::Tools::dir_glob_foreach($dir, '(((eth|bond)\d+|en[^.]+)(\.\d+)?)', sub {
522 push @ifaces, $_[0];
523 });
524
525 die "no physical interface on bridge '$bridge'\n" if scalar(@ifaces) == 0;
526
527 lock_network(sub {
528 # add bridgevlan if it doesn't already exist
529 if (! -d "/sys/class/net/$bridgevlan") {
530 iface_create($bridgevlan, 'bridge');
531 }
532
533 # for each physical interface (eth or bridge) bind them to bridge vlan
534 foreach my $iface (@ifaces) {
535 activate_bridge_vlan_slave($bridgevlan, $iface, $tag);
536 }
537
538 #fixme: set other bridge flags
539
540 # remove ipv6 link-local address before activation
541 disable_ipv6($bridgevlan);
542 # be sure to have the bridge up
543 &$activate_interface($bridgevlan);
544 });
545 return $bridgevlan;
546 }
547
548 sub tcp_ping {
549 my ($host, $port, $timeout) = @_;
550
551 my $refused = 1;
552
553 $timeout = 3 if !$timeout; # sane default
554 if (!$port) {
555 # Net::Ping defaults to the echo port
556 $port = 7;
557 } else {
558 # Net::Ping's port_number() implies service_check(1)
559 $refused = 0;
560 }
561
562 my ($sock, $result);
563 eval {
564 $result = PVE::Tools::run_with_timeout($timeout, sub {
565 $sock = IO::Socket::IP->new(PeerHost => $host, PeerPort => $port, Type => SOCK_STREAM);
566 $result = $refused if $! == ECONNREFUSED;
567 });
568 };
569 if ($sock) {
570 $sock->close();
571 $result = 1;
572 }
573 return $result;
574 }
575
576 sub IP_from_cidr {
577 my ($cidr, $version) = @_;
578
579 return if $cidr !~ m!^(\S+?)/(\S+)$!;
580 my ($ip, $prefix) = ($1, $2);
581
582 my $ipobj = Net::IP->new($ip, $version);
583 return if !$ipobj;
584
585 $version = $ipobj->version();
586
587 my $binmask = Net::IP::ip_get_mask($prefix, $version);
588 return if !$binmask;
589
590 my $masked_binip = $ipobj->binip() & $binmask;
591 my $masked_ip = Net::IP::ip_bintoip($masked_binip, $version);
592 return Net::IP->new("$masked_ip/$prefix");
593 }
594
595 sub is_ip_in_cidr {
596 my ($ip, $cidr, $version) = @_;
597
598 my $cidr_obj = IP_from_cidr($cidr, $version);
599 return undef if !$cidr_obj;
600
601 my $ip_obj = Net::IP->new($ip, $version);
602 return undef if !$ip_obj;
603
604 return $cidr_obj->overlaps($ip_obj) == $Net::IP::IP_B_IN_A_OVERLAP;
605 }
606
607
608 sub get_local_ip_from_cidr {
609 my ($cidr) = @_;
610
611 my $IPs = [];
612 run_command(['/sbin/ip', 'address', 'show', 'to', $cidr, 'up'], outfunc => sub {
613 if ($_[0] =~ m!^\s*inet(?:6)?\s+($PVE::Tools::IPRE)(?:/\d+|\s+peer\s+)!) {
614 push @$IPs, $1;
615 }
616 });
617
618 return $IPs;
619 }
620
621 sub addr_to_ip {
622 my ($addr) = @_;
623 my ($err, $host, $port) = Socket::getnameinfo($addr, NI_NUMERICHOST | NI_NUMERICSERV);
624 die "failed to get numerical host address: $err\n" if $err;
625 return ($host, $port) if wantarray;
626 return $host;
627 }
628
629 sub get_ip_from_hostname {
630 my ($hostname, $noerr) = @_;
631
632 my @res = eval { PVE::Tools::getaddrinfo_all($hostname) };
633 if ($@) {
634 die "hostname lookup '$hostname' failed - $@" if !$noerr;
635 return undef;
636 }
637
638 my ($ip, $family);
639 for my $ai (@res) {
640 $family = $ai->{family};
641 my $tmpip = addr_to_ip($ai->{addr});
642 if ($tmpip !~ m/^127\.|^::1$/) {
643 $ip = $tmpip;
644 last;
645 }
646 }
647 if (!defined($ip) ) {
648 die "hostname lookup '$hostname' failed - got local IP address '$ip'\n" if !$noerr;
649 return undef;
650 }
651
652 return wantarray ? ($ip, $family) : $ip;
653 }
654
655 sub lock_network {
656 my ($code, @param) = @_;
657 my $res = lock_file('/var/lock/pve-network.lck', 10, $code, @param);
658 die $@ if $@;
659 return $res;
660 }
661
662 1;