This is inspired by the logic in BIND [0], follow-up patches
will implement the reset of that scheme.
If we get a server error back, or if after several attempts we don't
get a reply at all, we switch from UDP to TCP for the given
server for the current and all subsequent requests. However, if
we ever successfully received a reply over UDP, we never fall
back to TCP, and once a grace-period has passed, we try to upgrade
again to using UDP. The grace-period starts off at five minutes
after the current feature level was verified and then grows
exponentially to six hours. This is to mitigate problems due
to temporary lack of network connectivity, but at the same time
avoid flooding the network with retries when the feature attempted
feature level genuinely does not work.
Note that UDP is likely much more commonly supported than TCP,
but depending on the path between the client and the server, we
may have more luck with TCP in case something is wrong. We really
do prefer UDP though, as that is much more lightweight, that is
why TCP is only the last resort.
[0]: <https://kb.isc.org/article/AA-01219/0/Refinements-to-EDNS-fallback-behavior-can-cause-different-outcomes-in-Recursive-Servers.html>
Let's track where the data came from: from the network, the cache or the
local zone. This is not only useful for debugging purposes, but is also
useful when the zone probing wants to ensure it's not reusing
transactions that were answered from the cache or the zone itself.
Previously we'd only store the DnsPacket in the DnsTransaction, and the
DnsQuery would then take the DnsPacket's DnsAnswer and return it. With
this change we already pull the DnsAnswer out inside the transaction.
We still store the DnsPacket in the transaction, if we have it, since we
still need to determine from which peer a response originates, to
implement caching properly. However, the DnsQuery logic doesn't care
anymore for the packet, it now only looks at answers and rcodes from the
successfuly candidate.
This also has the benefit of unifying how we propagate incoming packets,
data from the local zone or the local cache.
This adds support for searching single-label hostnames in a set of
configured search domains.
A new object DnsQueryCandidate is added that links queries to scopes.
It keeps track of the search domain last used for a query on a specific
link. Whenever a host name was unsuccessfuly resolved on a scope all its
transactions are flushed out and replaced by a new set, with the next
search domain appended.
This also adds a new flag SD_RESOLVED_NO_SEARCH to disable search domain
behaviour. The "systemd-resolve-host" tool is updated to make this
configurable via --search=.
Fixes#1697
This hopefully makes this a bit more expressive and clarifies that the
fd is not used for the DNS TCP socket. This also mimics how the LLMNR
UDP fd is named in the manager object.
Let's simplify things and only maintain a single RR key per transaction
object, instead of a full DnsQuestion. Unicast DNS and LLMNR don't
support multiple questions per packet anway, and Multicast DNS suggests
coalescing questions beyond a single dns query, across the whole system.
With the exponential backoff, we can perform more requests in the same amount of time,
so bump this a bit.
In case of large RTT this may be necessary in order not to regress, and in case
of large packet-loss it will make us more robust. The latter is particularly
relevant once we start probing for features (and hence may see packet-loss
until we settle on the right feature level).
Rather than fixing this to 5s for unicast DNS and 1s for LLMNR, start
at a tenth of those values and increase exponentially until the old
values are reached. For LLMNR the recommended timeout for IEEE802
networks (which basically means all of the ones we care about) is 100ms,
so that should be uncontroversial. For unicast DNS I have found no
recommended value. However, it seems vastly more likely that hitting a
500ms timeout is casued by a packet loss, rather than the RTT genuinely
being greater than 500ms, so taking this as a startnig value seems
reasonable to me.
In the common case this greatly reduces the latency due to normal packet
loss. Moreover, once we get support for probing for features, this means
that we can send more packets before degrading the feature level whilst
still allowing us to settle on the correct feature level in a reasonable
timeframe.
The timeouts are tracked per server (or per scope for the multicast
protocols), and once a server (or scope) receives a successfull package
the timeout is reset. We also track the largest RTT for the given
server/scope, and always start our timouts at twice the largest
observed RTT.
This function emits the UDP packet via the scope, but first it will
determine the current server (and connect to it) and store the
server in the transaction.
This should not change the behavior, but simplifies the code.
With access to the server when creating the socket, we can connect()
to the server and hence simplify message sending and receiving in
follow-up patches.
We used to have one global socket, use one per transaction instead. This
has the side-effect of giving us a random UDP port per transaction, and
hence increasing the entropy and making cache poisoining significantly
harder to achieve.
We still reuse the same port number for packets belonging to the same
transaction (resent packets).
This patch removes includes that are not used. The removals were found with
include-what-you-use which checks if any of the symbols from a header is
in use.
