If MTU probing discovered a node was not reachable via UDP, packets for it were
forwarded to the next hop, but always via TCP, even if the next hop was
reachable via UDP. This is now fixed by retrying to send the packet using
send_packet() if the destination is not the same as the nexthop.
This keeps NAT mappings for UDP alive, and will also detect when a node is not
reachable via UDP anymore or if the path MTU is decreasing. Tinc will fall back
to TCP if the node has become unreachable.
If UDP communication is impossible, we stop sending probes, but we retry if it
changes its keys.
We also decouple the UDP and TCP ping mechanisms completely, to ensure tinc
properly detects failure of either method.
This feature is not necessary anymore since we have tools like valgrind today
that can catch stack overflow errors before they make a backtrace in gdb
impossible.
During the path MTU discovery phase, we might not know the maximum MTU yet, but
we do know a safe minimum. If we encounter a packet that is larger than that
the minimum, we now send it via TCP instead to ensure it arrives. We also
allow large packets that we cannot fragment or create ICMP replies for to be
sent via TCP.
We used both rand() and random() in our code. Since it returns an int, we have
to use %x in our format strings instead of %lx. This fixes a crash under
Windows when cross-compiling tinc with a recent version of MinGW.
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.
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.
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.
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.
(The new code is still segfaulting for me, and I'd like to proceed with other
work.)
This largely rolls back to the revision 1545 state of the existing code
(new crypto layer is still there with no callers), though I reintroduced
the segfault fix of revision 1562.
This is a quick initial conversion that doesn't yet show much advantage:
- We roll our own timeouts.
- We roll our own signal handling.
- We build up the meta connection fd events on each loop rather than
on state changes.
This relieves some confusion and problems during the libevent transition.
In particular, "event_add" was defined by both.
(The 't' stands for 'timeout', 'tinc', 'temporary', or some such.)
