]> git.proxmox.com Git - pve-common.git/blob - src/PVE/Network.pm
support for predictable network interface device names
[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);
6 use PVE::ProcFSTools;
7 use PVE::INotify;
8 use File::Basename;
9 use IO::Socket::IP;
10 use POSIX qw(ECONNREFUSED);
11
12 use Net::IP;
13
14 use Socket qw(IPPROTO_IP);
15
16 use constant IFF_UP => 1;
17 use constant IFNAMSIZ => 16;
18 use constant SIOCGIFFLAGS => 0x8913;
19
20 # host network related utility functions
21
22 our $ipv4_reverse_mask = [
23 '0.0.0.0',
24 '128.0.0.0',
25 '192.0.0.0',
26 '224.0.0.0',
27 '240.0.0.0',
28 '248.0.0.0',
29 '252.0.0.0',
30 '254.0.0.0',
31 '255.0.0.0',
32 '255.128.0.0',
33 '255.192.0.0',
34 '255.224.0.0',
35 '255.240.0.0',
36 '255.248.0.0',
37 '255.252.0.0',
38 '255.254.0.0',
39 '255.255.0.0',
40 '255.255.128.0',
41 '255.255.192.0',
42 '255.255.224.0',
43 '255.255.240.0',
44 '255.255.248.0',
45 '255.255.252.0',
46 '255.255.254.0',
47 '255.255.255.0',
48 '255.255.255.128',
49 '255.255.255.192',
50 '255.255.255.224',
51 '255.255.255.240',
52 '255.255.255.248',
53 '255.255.255.252',
54 '255.255.255.254',
55 '255.255.255.255',
56 ];
57
58 our $ipv4_mask_hash_localnet = {
59 '255.255.0.0' => 16,
60 '255.255.128.0' => 17,
61 '255.255.192.0' => 18,
62 '255.255.224.0' => 19,
63 '255.255.240.0' => 20,
64 '255.255.248.0' => 21,
65 '255.255.252.0' => 22,
66 '255.255.254.0' => 23,
67 '255.255.255.0' => 24,
68 '255.255.255.128' => 25,
69 '255.255.255.192' => 26,
70 '255.255.255.224' => 27,
71 '255.255.255.240' => 28,
72 '255.255.255.248' => 29,
73 '255.255.255.252' => 30,
74 };
75
76 sub setup_tc_rate_limit {
77 my ($iface, $rate, $burst, $debug) = @_;
78
79 # these are allowed / expected to fail, e.g. when there is no previous rate limit to remove
80 eval { run_command("/sbin/tc class del dev $iface parent 1: classid 1:1 >/dev/null 2>&1"); };
81 eval { run_command("/sbin/tc filter del dev $iface parent ffff: protocol all pref 50 u32 >/dev/null 2>&1"); };
82 eval { run_command("/sbin/tc qdisc del dev $iface ingress >/dev/null 2>&1"); };
83 eval { run_command("/sbin/tc qdisc del dev $iface root >/dev/null 2>&1"); };
84
85 return if !$rate;
86
87 # tbf does not work for unknown reason
88 #$TC qdisc add dev $DEV root tbf rate $RATE latency 100ms burst $BURST
89 # so we use htb instead
90 run_command("/sbin/tc qdisc add dev $iface root handle 1: htb default 1");
91 run_command("/sbin/tc class add dev $iface parent 1: classid 1:1 " .
92 "htb rate ${rate}bps burst ${burst}b");
93
94 run_command("/sbin/tc qdisc add dev $iface handle ffff: ingress");
95 run_command("/sbin/tc filter add dev $iface parent ffff: " .
96 "prio 50 basic " .
97 "police rate ${rate}bps burst ${burst}b mtu 64kb " .
98 "drop flowid :1");
99
100 if ($debug) {
101 print "DEBUG tc settings\n";
102 system("/sbin/tc qdisc ls dev $iface");
103 system("/sbin/tc class ls dev $iface");
104 system("/sbin/tc filter ls dev $iface parent ffff:");
105 }
106 }
107
108 sub tap_rate_limit {
109 my ($iface, $rate) = @_;
110
111 my $debug = 0;
112 $rate = int($rate*1024*1024) if $rate;
113 my $burst = 1024*1024;
114
115 setup_tc_rate_limit($iface, $rate, $burst, $debug);
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 my $cond_create_bridge = sub {
162 my ($bridge) = @_;
163
164 if (! -d "/sys/class/net/$bridge") {
165 system("/sbin/brctl addbr $bridge") == 0 ||
166 die "can't add bridge '$bridge'\n";
167 }
168 };
169
170 my $bridge_add_interface = sub {
171 my ($bridge, $iface, $tag, $trunks) = @_;
172
173 system("/sbin/brctl addif $bridge $iface") == 0 ||
174 die "can't add interface 'iface' to bridge '$bridge'\n";
175
176 my $vlan_aware = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/bridge/vlan_filtering");
177
178 if ($vlan_aware) {
179 if ($tag) {
180 system("/sbin/bridge vlan add dev $iface vid $tag pvid untagged") == 0 ||
181 die "unable to add vlan $tag to interface $iface\n";
182 } else {
183 system("/sbin/bridge vlan add dev $iface vid 2-4094") == 0 ||
184 die "unable to add default vlan tags to interface $iface\n" if !$trunks;
185 }
186
187 if ($trunks) {
188 my @trunks_array = split /;/, $trunks;
189 foreach my $trunk (@trunks_array) {
190 system("/sbin/bridge vlan add dev $iface vid $trunk") == 0 ||
191 die "unable to add vlan $trunk to interface $iface\n";
192 }
193 }
194 }
195 };
196
197 my $ovs_bridge_add_port = sub {
198 my ($bridge, $iface, $tag, $internal, $trunks) = @_;
199
200 $trunks =~ s/;/,/g if $trunks;
201
202 my $cmd = "/usr/bin/ovs-vsctl add-port $bridge $iface";
203 $cmd .= " tag=$tag" if $tag;
204 $cmd .= " trunks=". join(',', $trunks) if $trunks;
205 $cmd .= " vlan_mode=native-untagged" if $tag && $trunks;
206
207 $cmd .= " -- set Interface $iface type=internal" if $internal;
208 system($cmd) == 0 ||
209 die "can't add ovs port '$iface'\n";
210 };
211
212 my $activate_interface = sub {
213 my ($iface) = @_;
214
215 system("/sbin/ip link set $iface up") == 0 ||
216 die "can't activate interface '$iface'\n";
217 };
218
219 sub tap_create {
220 my ($iface, $bridge) = @_;
221
222 die "unable to get bridge setting\n" if !$bridge;
223
224 my $bridgemtu = &$read_bridge_mtu($bridge);
225
226 eval {
227 PVE::Tools::run_command("/sbin/ifconfig $iface 0.0.0.0 promisc up mtu $bridgemtu");
228 };
229 die "interface activation failed\n" if $@;
230 }
231
232 sub veth_create {
233 my ($veth, $vethpeer, $bridge, $mac) = @_;
234
235 die "unable to get bridge setting\n" if !$bridge;
236
237 my $bridgemtu = &$read_bridge_mtu($bridge);
238
239 # create veth pair
240 if (! -d "/sys/class/net/$veth") {
241 my $cmd = "/sbin/ip link add name $veth type veth peer name $vethpeer mtu $bridgemtu";
242 $cmd .= " addr $mac" if $mac;
243 system($cmd) == 0 || die "can't create interface $veth\n";
244 }
245
246 # up vethpair
247 &$activate_interface($veth);
248 &$activate_interface($vethpeer);
249 }
250
251 sub veth_delete {
252 my ($veth) = @_;
253
254 if (-d "/sys/class/net/$veth") {
255 run_command("/sbin/ip link delete dev $veth", outfunc => sub {}, errfunc => sub {});
256 }
257
258 }
259
260 my $create_firewall_bridge_linux = sub {
261 my ($iface, $bridge, $tag, $trunks) = @_;
262
263 my ($vmid, $devid) = &$parse_tap_device_name($iface);
264 my ($fwbr, $vethfw, $vethfwpeer) = &$compute_fwbr_names($vmid, $devid);
265
266 &$cond_create_bridge($fwbr);
267 &$activate_interface($fwbr);
268
269 copy_bridge_config($bridge, $fwbr);
270 veth_create($vethfw, $vethfwpeer, $bridge);
271
272 &$bridge_add_interface($fwbr, $vethfw);
273 &$bridge_add_interface($bridge, $vethfwpeer, $tag, $trunks);
274
275 &$bridge_add_interface($fwbr, $iface);
276 };
277
278 my $create_firewall_bridge_ovs = sub {
279 my ($iface, $bridge, $tag, $trunks) = @_;
280
281 my ($vmid, $devid) = &$parse_tap_device_name($iface);
282 my ($fwbr, undef, undef, $ovsintport) = &$compute_fwbr_names($vmid, $devid);
283
284 my $bridgemtu = &$read_bridge_mtu($bridge);
285
286 &$cond_create_bridge($fwbr);
287 &$activate_interface($fwbr);
288
289 &$bridge_add_interface($fwbr, $iface);
290
291 &$ovs_bridge_add_port($bridge, $ovsintport, $tag, 1, $trunks);
292 &$activate_interface($ovsintport);
293
294 # set the same mtu for ovs int port
295 PVE::Tools::run_command("/sbin/ifconfig $ovsintport mtu $bridgemtu");
296
297 &$bridge_add_interface($fwbr, $ovsintport);
298 };
299
300 my $cleanup_firewall_bridge = sub {
301 my ($iface) = @_;
302
303 my ($vmid, $devid) = &$parse_tap_device_name($iface, 1);
304 return if !defined($vmid);
305 my ($fwbr, $vethfw, $vethfwpeer, $ovsintport) = &$compute_fwbr_names($vmid, $devid);
306
307 # cleanup old port config from any openvswitch bridge
308 if (-d "/sys/class/net/$ovsintport") {
309 run_command("/usr/bin/ovs-vsctl del-port $ovsintport", outfunc => sub {}, errfunc => sub {});
310 }
311
312 # delete old vethfw interface
313 veth_delete($vethfw);
314
315 # cleanup fwbr bridge
316 if (-d "/sys/class/net/$fwbr") {
317 run_command("/sbin/ip link set dev $fwbr down", outfunc => sub {}, errfunc => sub {});
318 run_command("/sbin/brctl delbr $fwbr", outfunc => sub {}, errfunc => sub {});
319 }
320 };
321
322 sub tap_plug {
323 my ($iface, $bridge, $tag, $firewall, $trunks, $rate) = @_;
324
325 #cleanup old port config from any openvswitch bridge
326 eval {run_command("/usr/bin/ovs-vsctl del-port $iface", outfunc => sub {}, errfunc => sub {}) };
327
328 if (-d "/sys/class/net/$bridge/bridge") {
329 &$cleanup_firewall_bridge($iface); # remove stale devices
330
331 my $vlan_aware = PVE::Tools::file_read_firstline("/sys/class/net/$bridge/bridge/vlan_filtering");
332
333 if (!$vlan_aware) {
334 die "vlan aware feature need to be enabled to use trunks" if $trunks;
335 my $newbridge = activate_bridge_vlan($bridge, $tag);
336 copy_bridge_config($bridge, $newbridge) if $bridge ne $newbridge;
337 $bridge = $newbridge;
338 $tag = undef;
339 }
340
341 if ($firewall) {
342 &$create_firewall_bridge_linux($iface, $bridge, $tag, $trunks);
343 } else {
344 &$bridge_add_interface($bridge, $iface, $tag, $trunks);
345 }
346
347 } else {
348 &$cleanup_firewall_bridge($iface); # remove stale devices
349
350 if ($firewall) {
351 &$create_firewall_bridge_ovs($iface, $bridge, $tag, $trunks);
352 } else {
353 &$ovs_bridge_add_port($bridge, $iface, $tag, undef, $trunks);
354 }
355 }
356
357 tap_rate_limit($iface, $rate);
358 }
359
360 sub tap_unplug {
361 my ($iface) = @_;
362
363 my $path= "/sys/class/net/$iface/brport/bridge";
364 if (-l $path) {
365 my $bridge = basename(readlink($path));
366 #avoid insecure dependency;
367 ($bridge) = $bridge =~ /(\S+)/;
368
369 system("/sbin/brctl delif $bridge $iface") == 0 ||
370 die "can't del interface '$iface' from bridge '$bridge'\n";
371
372 }
373
374 &$cleanup_firewall_bridge($iface);
375 #cleanup old port config from any openvswitch bridge
376 eval {run_command("/usr/bin/ovs-vsctl del-port $iface", outfunc => sub {}, errfunc => sub {}) };
377 }
378
379 sub copy_bridge_config {
380 my ($br0, $br1) = @_;
381
382 return if $br0 eq $br1;
383
384 my $br_configs = [ 'ageing_time', 'stp_state', 'priority', 'forward_delay',
385 'hello_time', 'max_age', 'multicast_snooping', 'multicast_querier'];
386
387 foreach my $sysname (@$br_configs) {
388 eval {
389 my $v0 = PVE::Tools::file_read_firstline("/sys/class/net/$br0/bridge/$sysname");
390 my $v1 = PVE::Tools::file_read_firstline("/sys/class/net/$br1/bridge/$sysname");
391 if ($v0 ne $v1) {
392 PVE::ProcFSTools::write_proc_entry("/sys/class/net/$br1/bridge/$sysname", $v0);
393 }
394 };
395 warn $@ if $@;
396 }
397 }
398
399 sub activate_bridge_vlan_slave {
400 my ($bridgevlan, $iface, $tag) = @_;
401 my $ifacevlan = "${iface}.$tag";
402
403 # create vlan on $iface is not already exist
404 if (! -d "/sys/class/net/$ifacevlan") {
405 system("/sbin/ip link add link $iface name ${iface}.${tag} type vlan id $tag") == 0 ||
406 die "can't add vlan tag $tag to interface $iface\n";
407 }
408
409 # be sure to have the $ifacevlan up
410 &$activate_interface($ifacevlan);
411
412 # test if $vlaniface is already enslaved in another bridge
413 my $path= "/sys/class/net/$ifacevlan/brport/bridge";
414 if (-l $path) {
415 my $tbridge = basename(readlink($path));
416 if ($tbridge ne $bridgevlan) {
417 die "interface $ifacevlan already exist in bridge $tbridge\n";
418 } else {
419 # Port already attached to bridge: do nothing.
420 return;
421 }
422 }
423
424 # add $ifacevlan to the bridge
425 &$bridge_add_interface($bridgevlan, $ifacevlan);
426 }
427
428 sub activate_bridge_vlan {
429 my ($bridge, $tag_param) = @_;
430
431 die "bridge '$bridge' is not active\n" if ! -d "/sys/class/net/$bridge";
432
433 return $bridge if !defined($tag_param); # no vlan, simply return
434
435 my $tag = int($tag_param);
436
437 die "got strange vlan tag '$tag_param'\n" if $tag < 1 || $tag > 4094;
438
439 my $bridgevlan = "${bridge}v$tag";
440
441 my @ifaces = ();
442 my $dir = "/sys/class/net/$bridge/brif";
443 PVE::Tools::dir_glob_foreach($dir, '(((eth|bond)\d+|en[^.]+)(\.\d+)?)', sub {
444 push @ifaces, $_[0];
445 });
446
447 die "no physical interface on bridge '$bridge'\n" if scalar(@ifaces) == 0;
448
449 # add bridgevlan if it doesn't already exist
450 if (! -d "/sys/class/net/$bridgevlan") {
451 system("/sbin/brctl addbr $bridgevlan") == 0 ||
452 die "can't add bridge $bridgevlan\n";
453 }
454
455 # for each physical interface (eth or bridge) bind them to bridge vlan
456 foreach my $iface (@ifaces) {
457 activate_bridge_vlan_slave($bridgevlan, $iface, $tag);
458 }
459
460 #fixme: set other bridge flags
461
462 # be sure to have the bridge up
463 system("/sbin/ip link set $bridgevlan up") == 0 ||
464 die "can't up bridge $bridgevlan\n";
465
466 return $bridgevlan;
467 }
468
469 sub tcp_ping {
470 my ($host, $port, $timeout) = @_;
471
472 my $refused = 1;
473
474 $timeout = 3 if !$timeout; # sane default
475 if (!$port) {
476 # Net::Ping defaults to the echo port
477 $port = 7;
478 } else {
479 # Net::Ping's port_number() implies service_check(1)
480 $refused = 0;
481 }
482
483 my ($sock, $result);
484 eval {
485 $result = PVE::Tools::run_with_timeout($timeout, sub {
486 $sock = IO::Socket::IP->new(PeerHost => $host, PeerPort => $port, Type => SOCK_STREAM);
487 $result = $refused if $! == ECONNREFUSED;
488 });
489 };
490 if ($sock) {
491 $sock->close();
492 $result = 1;
493 }
494 return $result;
495 }
496
497 sub IP_from_cidr {
498 my ($cidr, $version) = @_;
499
500 return if $cidr !~ m!^(\S+?)/(\S+)$!;
501 my ($ip, $prefix) = ($1, $2);
502
503 my $ipobj = Net::IP->new($ip, $version);
504 return if !$ipobj;
505
506 $version = $ipobj->version();
507
508 my $binmask = Net::IP::ip_get_mask($prefix, $version);
509 return if !$binmask;
510
511 my $masked_binip = $ipobj->binip() & $binmask;
512 my $masked_ip = Net::IP::ip_bintoip($masked_binip, $version);
513 return Net::IP->new("$masked_ip/$prefix");
514 }
515
516 sub is_ip_in_cidr {
517 my ($ip, $cidr, $version) = @_;
518
519 my $cidr_obj = IP_from_cidr($cidr, $version);
520 return undef if !$cidr_obj;
521
522 my $ip_obj = Net::IP->new($ip, $version);
523 return undef if !$ip_obj;
524
525 return $cidr_obj->overlaps($ip_obj) == $Net::IP::IP_B_IN_A_OVERLAP;
526 }
527
528 # struct ifreq { // FOR SIOCGIFFLAGS:
529 # char ifrn_name[IFNAMSIZ]
530 # short ifru_flags
531 # };
532 my $STRUCT_IFREQ_SIOCGIFFLAGS = 'Z' . IFNAMSIZ . 's1';
533 sub get_active_interfaces {
534 # Use the interface name list from /proc/net/dev
535 open my $fh, '<', '/proc/net/dev'
536 or die "failed to open /proc/net/dev: $!\n";
537 # And filter by IFF_UP flag fetched via a PF_INET6 socket ioctl:
538 my $sock;
539 socket($sock, PF_INET6, SOCK_DGRAM, &IPPROTO_IP)
540 or socket($sock, PF_INET, SOCK_DGRAM, &IPPROTO_IP)
541 or return [];
542
543 my $ifaces = [];
544 while(defined(my $line = <$fh>)) {
545 next if $line !~ /^\s*([^:\s]+):/;
546 my $ifname = $1;
547 my $ifreq = pack($STRUCT_IFREQ_SIOCGIFFLAGS, $ifname, 0);
548 if (!defined(ioctl($sock, SIOCGIFFLAGS, $ifreq))) {
549 warn "failed to get interface flags for: $ifname\n";
550 next;
551 }
552 my ($name, $flags) = unpack($STRUCT_IFREQ_SIOCGIFFLAGS, $ifreq);
553 push @$ifaces, $ifname if ($flags & IFF_UP);
554 }
555 close $fh;
556 close $sock;
557 return $ifaces;
558 }
559
560 1;