There are a number of ways a SPTPS tunnel can get into a corrupt state.
For example, during key regeneration, the KEX and SIG messages from
other nodes might arrive out of order, which confuses the hell out of
the SPTPS code. Another possible scenario is not noticing another node
crashed and restarted because there was no point in time where the node
was seen completely disconnected from *all* nodes; this could result in
using the wrong (old) key. There are probably other scenarios which have
not even been considered yet. Distributed systems are hard.
When SPTPS got confused by a packet, it used to crash the entire
process; fortunately that was fixed by commit
2e7f68ad2b. However, the error handling
(or lack thereof) leaves a lot to be desired. Currently, when SPTPS
encounters an error when receiving a packet, it just shrugs it off and
continues as if nothing happened. The problem is, sometimes getting
receive errors mean the tunnel is completely stuck and will not recover
on its own. In that case, the node will become unreachable - possibly
indefinitely.
The goal of this commit is to improve SPTPS error handling by taking
proactive action when an incoming packet triggers a failure, which is
often an indicator that the tunnel is stuck in some way. When that
happens, we simply restart SPTPS entirely, which should make the tunnel
recover quickly.
To prevent "storms" where two buggy nodes flood each other with invalid
packets and therefore spend all their time negotiating new tunnels, we
limit the frequency at which tunnel restarts happen to ten seconds.
It is likely this commit will solve the "Invalid KEX record length
during key regeneration" issue that has been seen in the wild. It is
difficult to be sure though because we do not have a full understanding
of all the possible conditions that can trigger this problem.
Commit 10c1f60c64 introduced a mechanism
by which a packet received by REQ_KEY could continue its journey over
UDP. This was based on the assumption that REQ_KEY messages would never
be used for handshake packets (which should never be sent over UDP,
because SPTPS currently doesn't handle lost handshake packets very
well).
Unfortunately, there is one case where handshake packets are sent using
REQ_KEY: when regenerating the SPTPS key for a pre-established channel.
With the current code, such packets risk getting relayed over UDP.
When processing a REQ_KEY message, it is impossible for the receiving
end to distinguish between a data SPTPS packet and a handshake packet,
because this information is stored in the type field which is encrypted
with the end-to-end key.
This commit fixes the issue by making tinc use ANS_KEY for all SPTPS
handshake messages. This works because ANS_KEY messages are never
forwarded using the SPTPS relay mechanisms, therefore they are
guaranteed to stick to TCP.
If the ADD_EDGE is for one of the edges we own, and if it is not the
same as we actually have, send a correcting ADD_EDGE back. Otherwise, if
the ADD_EDGE contains new information, update our idea of the local
address for that edge.
If the ADD_EDGE does not contain local address information, then we
never make a correction nor log a warning.
Currently, SPTPS packets are transported over TCP metaconnections using
extended REQ_KEY requests, in order for the packets to pass through
tinc-1.0 nodes unaltered. Unfortunately, this method presents two
significant downsides:
- An already encrypted SPTPS packet is decrypted and then encrypted
again every time it passes through a node, since it is transported
over the SPTPS channels of the metaconnections. This
double-encryption is unnecessary and wastes CPU cycles.
- More importantly, the only way to transport binary data over
standard metaconnection messages such as REQ_KEY is to encode it
in base64, which has a 33% encoding overhead. This wastes 25% of the
network bandwidth.
This commit introduces a new protocol message, SPTPS_PACKET, which can
be used to transport SPTPS packets over a TCP metaconnection in an
efficient way. The new message is appropriately protected through a
minor protocol version increment, and extended REQ_KEY messages are
still used with nodes that do not support the new message, as well as
for the intial handshake packets, for which efficiency is not a concern.
The way SPTPS_PACKET works is very similar to how the traditional PACKET
message works: after the SPTPS_PACKET message, the raw binary packet is
sent directly over the metaconnection. There is one important
difference, however: in the case of SPTPS_PACKET, the packet is sent
directly over the TCP stream completely bypassing the SPTPS channel of
the metaconnection itself for maximum efficiency. This is secure because
the SPTPS packet that is being sent is already encrypted with an
end-to-end key.
sptps_receive_data() always consumes the entire buffer passed to it,
which is somewhat inflexible. This commit improves the interface so that
sptps_receive_data() consumes at most one record. The goal is to allow
non-SPTPS stuff to be interleaved with SPTPS records in a single TCP
stream.
REQ_SPTPS implies the message has an ANS_ counterpart (like REQ_KEY,
ANS_KEY), but it doesn't. Therefore dropping the REQ_ seems more
appropriate, and we add a _PACKET suffix to reduce the likelihood of
naming conflicts.
Currently, when tinc receives a SPTPS packet over TCP via the REQ_KEY
encapsulation mechanism, it forwards it like any other TCP request. This
is inefficient, because even though we received the packet over TCP,
we might have an UDP link with the next hop, which means the packet
could be sent over UDP.
This commit removes that limitation by making sure SPTPS data packets
received through REQ_KEY requests are not forwarded as-is but passed
to send_sptps_data() instead, thereby using the same code path as if
the packet was received over UDP.
net_packet doesn't actually use send_sptps_data(); it only uses
send_sptps_data_priv(). In addition, the only user of send_sptps_data()
is protocol_key. Therefore it makes sense to expose
send_sptps_data_priv() directly, and move send_sptps_data() (which is
basically just boilerplate) as a local function in protocol_key.
Currently, when relaying SPTPS UDP packets, the code uses the direct
sender as the originator, instead of preserving the original source ID.
This wouldn't cause any issues in most cases because the originator and
the sender are the same in simple one-hop relay chains, but this will
break as soon as there is more than one relay.
This fixes some issues with the build system when building out of tree.
With this commit, it is now possible to do the following:
$ cd /tmp/build
$ /path/to/tinc/configure
$ make
This uses the output of "git describe" directly in configure.ac to
determine the version number to use, instead of hardcoding it.
With this change, current version information is completely removed
from the codebase itself, and is always fetched on-the-fly from git as
the single source of truth.
In order to ensure make dist always uses the current version number in
the contents of the packaged configure script as well as the package
name, a dependency is added to the dist target such that autoconf is
always run before dist to regenerate the version number. If this wasn't
the case, make dist would use the version number from when autoconf was
originally run, not the version number that make dist is running from.
That said, errors from that rule are ignored so that people can still
run make dist without a working autoconf.
In addition, the NEWS check is dropped, as it would then become annoying
because it would force make dist users to always have a line for the
current commit in the NEWS file.
Instead of using the hardcoded version number in configure.ac, this
makes tinc use the live version reported by "git describe",
queried on-the-fly during the build process and regenerated for every
build.
This makes tinc version output more useful, as tinc will now display the
number of commits since the last tag as well as the commit the binary is
built from, following the format described in git-describe(1).
Here's an example of tincd --version output:
tinc version release-1.1pre10-48-gc149315 (built Jun 29 2014 15:21:10, protocol 17.3)
When building directly from a release tag, this will look like the following:
tinc version release-1.1pre10 (built Jun 29 2014 15:21:10, protocol 17.3)
(Note that the format is slightly different - because of the way the
tags are named, it says "release-1.1pre10" instead of just "1.1pre10")
If git describe fails (for example when building from a release
tarball), the build automatically falls back to the autoconf-provided
VERSION macro (i.e. the old behavior).
read_rsa_public_key() was bailing out early if the given node already has an Ed25519 key, and
returned true even though c->rsa was NULL. The early bailout code isn't necessary anymore, so just
remove it.
This deals with the case where one node knows the Ed25519 key of another node, but not the other
way around. This was blocked by an overly paranoid check in id_h(). The upgrade_h() function already
handled this case, and the node that already knows the other's Ed25519 key checks that it has not
been changed, otherwise the connection will be aborted.
Unfortunately, glibc assumes that /etc/resolv.conf is a static file that
never changes. Even on servers, /etc/resolv.conf might be a dynamically
generated file, and we never know when it changes. So just call
res_init() every time, so glibc uses up-to-date nameserver information.
Conflicts:
src/have.h
src/net.c
src/net_setup.c
Testing has revealed that the newer series of Windows TAP drivers (i.e.
9.0.0.21 and later, also known as NDIS6, tap-windows6) suffer from
serious performance issues in the write path. Write operations seems to
take a very long time to complete, resulting in massive packet loss even
for throughputs as low as 10 Mbit/s.
I've made some attempts to alleviate the problem using parellelism. By
using custom code that allows up to 256 write operations at the same
time the results are much better, but it's still about 2 times worse
than the traditional 9.0.0.9 driver.
We need to investigate more and file a bug against tap-windows6, but in
the mean time, let's inform the user that he might not want to use the
latest drivers.
This is generally useful. We've seen issues that are specific to some
version of these drivers (especially the newer 9.0.0.21 version), so
it's relevant to log it, especially since that means it will be
copy-pasted by people posting their logs asking for help.
As a rule, it seems reasonable to make sure that tinc operates correctly
on at least 1G links, since these are pretty common. However, I have
observed replay window issues when operating at speeds of 600 Mbit/s and
above, especially when the receiving end is a Windows system (not sure
why). This commit increases the default so that this won't occur on
fresh setups.
It may not be obvious, but due to the way tinc operates (single-threaded
control loop with no intermediate packet buffer), UDP send and receive
buffers can have a massive impact on performance. It is therefore of
paramount importance that the buffers be large enough to prevent packet
drops that could occur while tinc is processing a packet.
Leaving that value to the OS default could be reasonable if we weren't
relying on it so much. Instead, this makes performance somewhat
unpredictable.
In practice, the worst case scenario occurs on Windows, where Microsoft
had the brillant idea of making the buffers 8K in size by default, no
matter what the link speed is. Considering that 8K flies past in a
matter of microseconds on >1G links, this is extremely inappropriate. On
these systems, changing the buffer size to 1M results in *obscene*
raw throughput improvements; I have observed a 10X jump from 40 Mbit/s
to 400 Mbit/s on my system.
In this commit, we stop trusting the OS to get this right and we use a
fixed 1M value instead, which should be enough for <=1G links.
Write operations to the Windows device do not necessarily complete
immediately; in fact, with the latest TAP-Win32 drivers, this never
seems to be the case.
write_packet() does not handle that case correctly, because the
OVERLAPPED structure and the packet data go out of scope before the
write operation completes, resulting in race conditions.
This commit fixes the issue by making sure these data structures are
kept in global scope, and by dropping any packets that may arrive while
the previous write operation is still pending.
On Windows, when disabling the device, tinc uses the CancelIo() to
cancel the pending read operation, and then proceeds to delete the event
handle immediately.
This assumes that CancelIo() blocks until the pending read request is
completely torn down and no references to it remain. While MSDN is not
completely clear on that subject, it does suggest that this is not the
case:
http://msdn.microsoft.com/en-us/library/windows/desktop/aa363791.aspx
If the function succeeds [...] the cancel operation for all pending
I/O operations issued by the calling thread for the specified file
handle was successfully requested.
This implies that cancellation was merely "requested", and that there
are no guarantees as to the state of the operation when CancelIo()
returns. Therefore, care must be taken not to close event handles
prematurely.
While I'm no aware of this potential race condition causing any problems
in practice, I don't want to take any chances.
Modern versions of GCC handle structure packing differently when
compiling for Windows, as reported in the following GCC bug report:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52991
In practice, this affects tinc because it uses packed structs as a
convenient way to populate packet headers. "struct ip" is especially
affected - on Linux, sizeof(struct ip) returns 20 as expected, while on
Windows, it returns 24 because of the broken alignment.
This in turn completely breaks code that has to populate an IP header.
Specifically, this breaks route_ipv4_unreachable() which is responsible,
among other things, for the generation of ICMP Fragmentation Needed
messages. On Windows, these messages are corrupted beyond hope because
of this alignment issue. For TCP connections that are established
before tinc obtains a fix on the MTU (and thus are not MSS clamped),
this can result in massive disruption.
This commit fixes the issue by forcing GCC to use standard alignment
for all packed structures in the tinc codebase instead of the MSVC
alignment.
HAVE_DECL_RES_INIT is generated using AC_CHECK_DECLS. tinc checks this
symbol using #ifdef, which is wrong because (according to autoconf docs)
the symbol is always defined, it's just set to zero if the check failed.
This broke the Windows build starting from
0b310bf406, because it introduced this
conditional in code that's not excluded from the Windows build.
Ironically, commit 0f8e2cc78c introduced
a regression on its own, since it accidently removed a return statement
that prevented try_tx_sptps() from sending UDP/MTU probes to nodes that
are past static relays.
This makes sure MTU_INFO messages are only sent at the maximum rate of
5 per second (by default). As usual with these "probe" mechanisms, the
rate of these messages cannot be higher than the rate of data packets
themselves, since they are sent from the RX path.
This makes sure UDP_INFO messages are only sent at the maximum rate of
5 per second (by default). As usual with these "probe" mechanisms, the
rate of these messages cannot be higher than the rate of data packets
themselves, since they are sent from the RX path.
In this commit, nodes use MTU_INFO messages to provide MTU information.
The issue this code is meant to address is the non-trivial problem of
finding the proper MTU when UDP SPTPS relays are involved. Currently,
tinc has no idea what the MTU looks like beyond the first relay, and
will arbitrarily use the first relay's MTU as the limit. This will fail
miserably if the MTU decreases after the first relay, forcing relays to
fall back to TCP. More generally, one should keep in mind that relay
paths can be arbitrarily complex, resulting in packets taking "epic
journeys" through the graph, switching back and forth between UDP (with
variable MTUs) and TCP multiple times along the path.
A solution that was considered consists in sending standard MTU probes
through the relays. This is inefficient (if there are 3 nodes on one
side of relay and 3 nodes on the other side, we end up with 3*3=9 MTU
discoveries taking place at the same time, while technically only
3+3=6 are needed) and would involve eyebrow-raising behaviors such as
probes being sent over TCP.
This commit implements an alternative solution, which consists in
the packet receiver sending MTU_INFO messages to the packet sender.
The message contains an MTU value which is set to maximum when the
message is originally sent. The message gets altered as it travels
through the metagraph, such that when the message arrives to the
destination, the MTU value contained in the message can be used to
send packets while making sure no relays will be forced to fall back to
TCP to deliver them.
The operating principles behind such a protocol message are similar to
how the UDP_INFO message works, but there is a key difference that
prevents us from simply reusing the same message: the UDP_INFO message
only cares about relay-to-relay links (i.e. it is sent between static
relays and the information it contains only makes sense between two
adjacent static relays), while the MTU_INFO cares about the end-to-end
MTU, including the entire relay path. Therefore, UDP_INFO messages stop
when they encounter static relays, while MTU_INFO messages don't stop
until they get to the original packet sender.
Note that, technically, the MTU that is obtained through this mechanism
can be slightly pessimistic, because it can be lowered by an
intermediate node that is not being used as a relay. Since nodes have no
way of knowing whether they'll be used as dynamic relays or not (and
have no say in the matter), this is not a trivial problem. That said,
this is highly unlikely to result in noticeable issues in realistic
scenarios.
In this commit, nodes use UDP_INFO messages to provide UDP address
information. The basic principle is that the node that receives packets
sends UDP_INFO messages to the node that's sending the packets. The
message originally contains no address information, and is (hopefully)
updated with relevant address information as it gets relayed through the
metagraph - specifically, each intermediate node will update the message
with its best guess as to what the address is while forwarding it.
When a node receives an UDP_INFO message, and it doesn't have a
confirmed UDP tunnel with the originator node, it will update its
records with the new address for that node, so that it always has the
best possible guess as to how to reach that node. This applies to the
destination node of course, but also to any intermediate nodes, because
there's no reason they should pass on the free intel, and because it
results in nice behavior in the presence of relay chains (multiple nodes
in a path all trying to reach the same destination).
If, on the other hand, the node does have a confirmed UDP tunnel, it
will ignore the address information contained in the message.
In all cases, if the node that receives the message is not the
destination node specified in the message, it will forward the message
but not before overriding the address information with the one from its
own records. If the node has a confirmed UDP tunnel, that means the
message is updated with the address of the confirmed tunnel; if not,
the message simply reflects the records of the intermediate node, which
just happen to be the contents of the UDP_INFO message it just got, so
it's simply forwarded with no modification.
This is similar to the way ANS_KEY messages are currently
overloaded to provide UDP address information, with two differences:
- UDP_INFO messages are sent way more often than ANS_KEY messages,
thereby keeping the address information fresh. Previously, if the UDP
situation were to change after the ANS_KEY message was sent, the
sender would virtually never get the updated information.
- Once a node puts address information in an ANS_KEY message, it is
never changed again as the message travels through the metagraph; in
contrast, UDP_INFO messages behave the opposite way, as they get
rewritten every time they travel through a node with a confirmed UDP
tunnel. The latter behavior seems more appropriate because UDP tunnel
information becomes more relevant as it moves closer to the
destination node. The ANS_KEY behavior is not satisfactory in some
cases such as multi-layered graphs where the first hop is located
before a NAT.
Ultimately, the rationale behind this whole process is to improve UDP
hole punching capabilities when port translation is in effect, and more
generally, to make tinc more reliable in (very) hostile network
conditions (such as multi-layered NAT).
