As we set the remote port to 0, we'd get a mapping change message with
every packet. Setting the threshold avoids all kernel messages after the
first, which we suppress explicitly as well.
Note that while PF_KEYv2 also uses sequence numbers to identify acquires,
which we currently don't use correctly by the way, it does not include
information about the packet that triggered an acquire. What we receive
in src and dst, and currently forward as traffic selectors, are actually
the designated endpoints of the SA. So especially in tunnel mode this is
useless to do narrowing on the responder (these addresses might not even
match the configured TS).
This enables GRO offload for inbound ESP-in-UDP packets if the
esp4|6_offload modules are loaded. Note that inbound ESP or ESP-in-UDP
packets won't be visible on layer 3 in Netfilter or tcpdump.
This reverts commit f717bb5249.
Causes issues in our testing environment. Default route via host is
preferred if no gateway is set in the installed routes. Needs some
investigation.
References strongswan/strongswan#2548
This doesn't really seem useful (perhaps it was before we started to
configure the outbound interface on our routes). And it can actually
cause the route installation to fail e.g. for routes over point-to-point
interfaces where we'd get "Error: Nexthop has invalid gateway" errors.
Closesstrongswan/strongswan#2548
The kernel now allows a 0 replay window with ESN for SAs that are
explicitly tagged as outbound SAs. But not just that, it actually
rejects outbound SAs with replay windows > 0. So we add a version check
to control the replay window size. Note that adding the attribute
unconditionally would be fine even for older kernels, but if somebody
backports the direction patches, the installation of outbound SAs might
fail if the replay window is not adjusted accordingly.
This reverts commit 8b9b11919d.
Since ESN was negotiated via proposal, just configuring the SA without
ESN won't work as the ICV will be incorrect if the peer enabled ESN
on its SA. While the Linux kernel currently doesn't support disabling
replay protection for SAs that use ESN, this at least gets users an
explicit error not just dropped packets, and it will automatically work
if the kernel supports this combination at some point.
References strongswan/strongswan#2117
A refcount variable is used to allocate sequential unique identifiers for
Netlink sequence numbers, subject to overflows. The risk of an overflow
has so far not been considered practical, as it requires 2^32 netlink
requests.
It seems that this issue is not only theoretical. A host with thousands
of tunnels doing aggressive rekeying and/or aggressive status checking
(via vici list-sas) may trigger the overflow after a few weeks uptime.
The consequences are rather devastating: Once the refcount overflows, a
Netlink request is sent with sequence number 0. This request is answered
by the kernel, but can't be matched to the request, resulting in the error:
"received unknown netlink seq 0, ignored". Without Netlink timeouts, the
thread indefinitely waits for a response while holding the Netlink mutex,
bringing all threads to a halt.
So at all costs avoid zero sequence numbers. Also, start at sequence number
1 instead of the arbitrary 201, so the same range is used on start and after
an overflow.
The kernel includes the XFRMA_REPLAY_ESN_VAL attribute when dumping
SAs since it was added with 2.6.39. So we basically added this attribute
twice to the message sent to the kernel, potentially exceeding the
message buffer if the window size is large.
The XFRMA_REPLAY_VAL attribute is only dumped since 3.19, so that might
still be relevant (Google seems to maintain a 3.18 kernel) and since we
have to query the current lifetime stats anyway, we can just avoid adding
this attribute twice.
Closesstrongswan/strongswan#1967
Also simplify how we try to exceed the system-wide maximum. We basically
just try to force the value and simply fall back to the regular call.
The kernel actually won't let the latter fail if the value is too big,
it just caps it at the internal maximum.
The calculation of the timeout is also shared now and the total
timeout in seconds is corrected in case retransmit_base is <= 1.
This could make it easier in the future to apply different retransmission
settings to messages/exchanges.
Since these might conflict with IKE traffic, this requires special care.
One option is to install bypass policies for the peer, which install
appropriate (throw) routes. However, that won't work if the traffic to
the gateway itself should be protected, in particular, for host-to-host
tunnels. So an alternative is to set a mark for the IKE socket and then
exclude that traffic from table 220 via the kernel-netlink plugin's
fwmark option.
The manager will allow charon-nm to create XFRM interfaces if supported
by the kernel instead of creating an unused dummy TUN interface.
The xfrmi tool is mostly obsolete nowadays as iproute2 supports creating
XFRM interfaces since 5.1.0 (2019-05). Older Debians don't ship that and
early versions didn't list the interface IDs. So there might still be
some uses for this tool.
Since 6.2 the Linux kernel updates the last use time per SA. In
previous releases the attribute was only updated and reported for
specific outbound IPv6 SAs.
Using this reduces the number of kernel queries per CHILD_SA: for DPDs
from two policy queries (IN/FWD) to a single query of the inbound SA,
and for status reports the three policy queries (IN/FWD/OUT) can be
omitted and only the two SAs have to be queried. For NAT keepalives the
number of queries doesn't change but a policy query (OUT) is replaced by
a query for the outbound SA.
While we could use the existence of the attribute as indicator for its
support, we don't know this until we queried an SA. By using a version
check we can announce the feature from the start.
While this uses the same mechanism, it's not necessary to explicitly
enable port_bypass, the regular socket policies work fine to bypass any
software policies.
While the kernel defines it since 2005 (2.6.14), some older versions of
socket.h shipped with C libraries might not. In particular, glibc only
added it with 2.24 in 2016.
Closesstrongswan/strongswan#1503
With newer kernels (basic support for extended ACKs is there since 4.12
but some messages for XFRM were only added with 6.1) this gives more
detailed error messages to the user than e.g. a generic EINVAL or ENOSYS
error would.
Also enabled omitting the request payload in NLMSG_ERROR messages.
When watching the output of `swanctl -l` during debugging, the debug
messages in query_sa/policy() cause a lot of noise in the logs (level 2
for DBG_KNL still has actually useful information that we want to see
in the logs) and they're not very useful.
Compared to the messages in the functions above, the ones in update_sa()
and get_replay_state() are not seen often. But since there already is a
log message on level 2 in update_sa(), they're kinda redundant.
Closesstrongswan/strongswan#1271
Normally, GCC sees that we terminate the destination with a zero byte.
However, when using `-fsanitize=address`, there seems to be additional
instrumentation code after strncpy() so GCC produces warnings like
these:
‘__builtin_strncpy’ specified bound 16 equals destination size [-Wstringop-truncation]
This was originally added with 1551d8b13d ("kernel-netlink: reject
policy refcount if the reqid differs"). Since then we added code to
allocate constant reqids for the same TS, which pretty much avoids the
previous issues.
However, the reqid might have to be changed due to MOBIKE updates. And
because reqids are allocated for a complete set of traffic selectors and
not individual pairs, this can create a problem with drop policies as
those will use the old reqid (they are installed with the same priority,
reqid etc. to replace the actual IPsec policies), while unmodified
replacement policies will use the new one. A similar issue exists for
CHILD_SAs with SELinux contexts as those all use duplicate policies (same
generic label) but can't all be updated concurrently.
If a policy with IPComp template triggers an acquire, we get two, one for
an IPComp, one for ESP/AH SA. However, the triggering template of the trap
policy (where we get the reqid from), will be the same in both acquires,
IPComp, which we ignore, so no acquire was actually forwarded.
Tobias Brunner