Valgrind caught tinc reading free'd memory during a purge(). This was caused by
first removing it from the main node tree, which will already call free_node(),
and then removing it from the UDP tree. This might cause spurious segmentation
faults.
Although we select() before we call recvfrom(), it sometimes happens that
select() tells us we can read but a subsequent read fails anyway. This is
harmless.
If there is an outstanding MTU probe event for a node which is not reachable
anymore, a UDP packet would be sent to that node, which caused a key request to
be sent to that node, which triggered a NULL pointer dereference. Probes and
other UDP packets to unreachable nodes are now dropped.
When chrooted, we either need to force-initialize resolver
and/or nsswitch somehow (no clean way) or resolve all the
names we want before entering chroot jail. The latter
looks cleaner, easier and it is actually safe because
we still don't talk with the remote nodes there, only
initiating outgoing connections.
This option can be set to low, normal or high. On UNIX flavours, this changes
the nice value of the process by +10, 0 and -10 respectively. On Windows, it
sets the priority to BELOW_NORMAL_PRIORITY_CLASS, NORMAL_PRIORITY_CLASS and
HIGH_PRIORITY_CLASS respectively.
A high priority might help to reduce latency and packet loss on the VPN.
If a host has multiple addresses on an interface, the source address of the TCP
connection(s) was picked by the operating system while the UDP packets used a
bound socket, i. e. the source address was the address specified by the user.
This caused problems because the receiving code requires the TCP connection and
the UDP connection to originate from the same IP address.
This patch adds support for the `BindToInterface' and `BindToAddress' options
to the setup of outgoing TCP connections.
Tested with Debian Etch on x86 and Debian Lenny on x86_64.
Signed-off-by: Florian Forster <octo@verplant.org>
If running without `--net', the (global) variable `netname' is NULL. This
creates a segmentation fault because this NULL-pointer is passed to strdup:
Program terminated with signal 11, Segmentation fault.
#0 0xb7d30463 in strlen () from /lib/tls/i686/cmov/libc.so.6
(gdb) bt
#0 0xb7d30463 in strlen () from /lib/tls/i686/cmov/libc.so.6
#1 0xb7d30175 in strdup () from /lib/tls/i686/cmov/libc.so.6
#2 0x0805bf47 in xstrdup (s=0x0) at xmalloc.c:118 <---
#3 0x0805be33 in setup_device () at device.c:66
#4 0x0805072e in setup_myself () at net_setup.c:432
#5 0x08050db2 in setup_network () at net_setup.c:536
#6 0x0805b27f in main (argc=Cannot access memory at address 0x0) at tincd.c:580
This patch fixes this by checking `netname' in `setup_device'. An alternative
would be to check for NULL-pointers in `xstrdup' and return NULL in this case.
Signed-off-by: Florian Forster <octo@verplant.org>
Add some logging about refused ADD_SUBNET
(it causes subsequent client disconnect so it's
important to know which subnet was at fault).
Maybe we should just ignore it completely.
First of all, the idea behind the TunnelServer option is to hide all other
nodes from each other, so we shouldn't forward broadcast packets from them
anyway. The other reason is that since edges from other nodes are ignored, the
calculated minimum spanning tree might not be correct, which can result in
routing loops.
Since compression can either grow or shrink a packet, the size of an MTU probe
after decompression might not reflect the real path MTU. Now we use the size
before decompression, which is independent of the compression algorithm, and
substract a safety margin such that the calculated path MTU will be safe even
for packets which grow as much as possible after compression.
Instead of a single, global decryption context, each node has its own context.
However, in send_ans_key(), the global context was initialised. This commit
fixes that and removes the global context completely.
Also only set status.validkey after all checks have been evaluated.
In tunnelserver mode we're not interested to hear about
our client edges, just like in case of subnets. Just
ignore all requests which are not about our node or the
client node.
The fix is very similar to what was done for subnets.
Note that we don't need to add the "unknown" nodes to
the list in tunnelserver mode too, so move allocation
of new nodes down the line.
Similar changes as was in 2327d3f6eb
but for del_subnet_h().
Before, we vere returning false (and causing disconnect of the
client) in case of tunnelserver and the client sending DEL_SUBNET
for non-his subnet or for subnet which owner isn't in our connection
list.
After the mentioned change to add_subnet_h() that routine does not
add such indirect owners to the connection list anymore, so that
was ok (owner == NULL and we return true).
But if we too has a connection with the node about which the client
is sending DEL_SUBNET notification, say, because that client lost
connection with that other node, we'll disconnect this client from
us too, returning false for indirect DEL_SUBNET.
Fix that by allowing and ignoring indirect DEL_SUBNET in tunnelserver
mode.
Also rearranged the function a bit, to match add_subnet_h() (in
particular, syntax-check everything first, see if we've seen this
request before).
And also fix some comments.
Change formatting of error messages about failed syscalls
to be the same as in other places in tincd.
Also suggest a change in "$foo not supported on this platform"
message as it's now used more than once.
mlock()/mlockall() are not persistent across fork(), and it's
done in parent process before daemon() which does fork(). So
basically, current --mlock does nothing useful.
Move mlock() to after detach() so it works for child process
instead of parent.
Also, check if the platform supports mlock right when processing
options (since else we'll have to die after startup, not at
startup, the error message will be in log only).
Add two options, -R/--chroot and -U/--user=user, to chroot to the
config directory (where tinc.conf is located) and to perform
setuid to the user specified, after all the initialization is done.
What's left is handling of pid file since we can't remove it anymore.
In preparation of chroot/setuid operations, split out call to
try_outgoing_connections() from setup_network_connections()
(which was the last call in setup_network_connections()).
This is because dropping privileges should be done in-between
setup_network_connections() and try_outgoing_connections().
This patch renames setup_network_connections() to setup_network()
and moves call to try_outgoing_connections() into main routine.
No functional changes.
Previously, tinc used a fixed address and port for each node for UDP packet
exchange. The port was the one advertised by that node as its listening port.
However, due to NAT the port might be different. Now, tinc sends a different
session key to each node. This way, the sending node can be determined from
incoming packets by checking the MAC against all session keys. If a match is
found, the address and port for that node are updated.
In switch and hub modes, tinc does not generate ICMP packets in response to
packets that are larger than the path MTU. However, if PMTUDiscovery is
enabled, the IP_MTU_DISCOVER and IPV6_MTU_DISCOVER option is set on the UDP
sockets, which causes all UDP packets to be sent with the DF bit set, causing
large packets to be dropped, even if they would otherwise be routed fine.
Tinc allows multiple nodes to own the same Subnet, but did not have a sensible
way to decide which one to send packets to. Tinc also did not check the
reachability of nodes when deciding where to route packets to, so it would not
automatically fail over to a reachable node.
Tinc now assigns a weight to each Subnet. The default weight is 10, with lower
weights having higher priority. The Subnets are now internally sorted in the
same way as the kernel's routing table, and the Subnets are search linearly,
skipping those of unreachable nodes. A small cache of recently used addresses
is used to speed up the lookup functions.
Apparently FreeBSD likes to send out neighbor solicitation requests, even on a
tun interface where this is completely pointless. These requests do not have an
option header containing a link layer address, so the proxy-neighborsol code
was treating these requests as invalid. We now handle such requests, and send
back equally pointless replies, also without a link layer address. This seems
to satisfy FreeBSD.
In TunnelServer mode, tinc server disconnects any client if it announces
indirect subnets -- subnets that are not theirs (e.g. subnets for nodes
the CLIENT has connections now, even if those nodes are known to the server
too). Fix that by ignoring such (indirect) announces instead.
While we're at it, move check for such indirect subnet registration to
before allocating new node structure, as in TunnelServer mode we don't
really need to know that other node.
When generating an RSA keypair, the new public and private keys are appended to
files. However, when OpenSSL reads keys it only reads the first in a file, not
the last. Instead of printing an easily ignored warning, tinc now disables old
keys when appending new ones.
Previously an outgoing_t was maintained for each outgoing connection,
but the pointer to it was either stored in a connection_t or in an event_t.
This made it very hard to keep track of and to clean up.
Now a list is created when tinc starts and reads all the ConnectTo variables,
and which is recreated when tinc receives a HUP signal.
The former function made a totally bogus shallow copy of the event_tree, called
the handler of each event and then deleted the whole tree. This should've
caused tinc to crash when an ALARM signal was sent more than once, but for some
reason it didn't. It also behaved incorrectly when a handler added a new event.
The new function just moves the expiration time of all events to the past.
When no session key is known for a node, or when it is doing PMTU discovery but
no MTU probes have returned yet, packets are sent via TCP. Some logic is added
to make sure intermediate nodes continue forwarding via TCP. The per-node
packet queue is now no longer necessary and has been removed.
Due to a copy&paste bug, it tried to write a packet with the maximum size.
This was not a problem until the maximum size was increased to support VLANs.
It's a wonder it ever worked before. The socket that is created is not of a
datagram type, therefore packet boundaries were not preserved, which becomes
a problem as soon as the TAP-Win32 device receives packets in fast succession.
It has been available for years on any decent Linux distribution.
Although linux/if_tun.h is now required to compile tinc,
you can still run it on systems which only support Ethertap.
