sandy
README.md
DNS-over-TLS query forwarder design
Overview
The DNS-over-TLS query forwarder consists of five classes:
DnsTlsDispatcherDnsTlsTransportDnsTlsQueryMapDnsTlsSessionCacheDnsTlsSocket
DnsTlsDispatcher is a singleton class whose query method is the DnsTls's
only public interface. DnsTlsDispatcher is just a table holding the
DnsTlsTransport for each server (represented by a DnsTlsServer struct) and
network. DnsTlsDispatcher also blocks each query thread, waiting on a
std::future returned by DnsTlsTransport that represents the response.
DnsTlsTransport sends each query over a DnsTlsSocket, opening a
new one if necessary. It also has to listen for responses from the
DnsTlsSocket, which happen on a different thread.
IDnsTlsSocketObserver is an interface defining how DnsTlsSocket returns
responses to DnsTlsTransport.
DnsTlsQueryMap and DnsTlsSessionCache are helper classes owned by DnsTlsTransport.
DnsTlsQueryMap handles ID renumbering and query-response pairing.
DnsTlsSessionCache allows TLS session resumption.
DnsTlsSocket interleaves all queries onto a single socket, and reports all
responses to DnsTlsTransport (through the IDnsTlsObserver interface). It doesn't
know anything about which queries correspond to which responses, and does not retain
state to indicate whether there is an outstanding query.
Threading
Overall patterns
For clarity, each of the five classes in this design is thread-safe and holds one lock.
Classes that spawn a helper thread call thread::join() in their destructor to ensure
that it is cleaned up appropriately.
All the classes here make full use of Clang thread annotations (and also null-pointer annotations) to minimize the likelihood of a latent threading bug. The unit tests are also heavily threaded to exercise this functionality.
This code creates O(1) threads per socket, and does not create a new thread for each query or response. However, DnsProxyListener does create a thread for each query.
Threading in DnsTlsSocket
DnsTlsSocket can receive queries on any thread, and send them over a
"reliable datagram pipe" (socketpair() in SOCK_SEQPACKET mode).
The query method writes a struct (containing a pointer to the query) to the pipe
from its thread, and the loop thread (which owns the SSL socket)
reads off the other end of the pipe. The pipe doesn't actually have a queue "inside";
instead, any queueing happens by blocking the query thread until the
socket thread can read the datagram off the other end.
We need to pass messages between threads using a pipe, and not a condition variable
or a thread-safe queue, because the socket thread has to be blocked
in poll() waiting for data from the server, but also has to be woken
up on inputs from the query threads. Therefore, inputs from the query
threads have to arrive on a socket, so that poll() can listen for them.
(There can only be a single thread because you can't use different threads
to read and write in OpenSSL).
ID renumbering
DnsTlsDispatcher accepts queries that have colliding ID numbers and still sends them on
a single socket. To avoid confusion at the server, DnsTlsQueryMap assigns each
query a new ID for transmission, records the mapping from input IDs to sent IDs, and
applies the inverse mapping to responses before returning them to the caller.
DnsTlsQueryMap assigns each new query the ID number one greater than the largest
ID number of an outstanding query. This means that ID numbers are initially sequential
and usually small. If the largest possible ID number is already in use,
DnsTlsQueryMap will scan the ID space to find an available ID, or fail the query
if there are no available IDs. Queries will not block waiting for an ID number to
become available.
Time constants
DnsTlsSocket imposes a 20-second inactivity timeout. A socket that has been idle for
20 seconds will be closed. This sets the limit of tolerance for slow replies,
which could happen as a result of malfunctioning authoritative DNS servers.
If there are any pending queries, DnsTlsTransport will retry them.
DnsTlsQueryMap imposes a retry limit of 3. DnsTlsTransport will retry the query up
to 3 times before reporting failure to DnsTlsDispatcher.
This limit helps to ensure proper functioning in the case of a recursive resolver that
is malfunctioning or is flooded with requests that are stalled due to malfunctioning
authoritative servers.
DnsTlsDispatcher maintains a 5-minute timeout. Any DnsTlsTransport that has had no
outstanding queries for 5 minutes will be destroyed at the next query on a different
transport.
This sets the limit on how long session tickets will be preserved during idle periods,
because each DnsTlsTransport owns a DnsTlsSessionCache. Imposing this timeout
increases latency on the first query after an idle period, but also helps to avoid
unbounded memory usage.
DnsTlsSessionCache sets a limit of 5 sessions in each cache, expiring the oldest one
when the limit is reached. However, because the client code does not currently
reuse sessions more than once, it should not be possible to hit this limit.
Testing
Unit tests are in dns_tls_test.cpp. They cover all the classes except
DnsTlsSocket (which requires CAP_NET_ADMIN because it uses setsockopt(SO_MARK)) and
DnsTlsSessionCache (which requires integration with libssl). These classes are
exercised by the integration tests in ../tests/resolv_test.cpp.
Dependency Injection
For unit testing, we would like to be able to mock out DnsTlsSocket. This is
particularly required for unit testing of DnsTlsDispatcher and DnsTlsTransport.
To make these unit tests possible, this code uses a dependency injection pattern:
DnsTlsSocket is produced by a DnsTlsSocketFactory, and both of these have a
defined interface.
DnsTlsDispatcher's constructor takes an IDnsTlsSocketFactory,
which in production is a DnsTlsSocketFactory. However, in unit tests, we can
substitute a test factory that returns a fake socket, so that the unit tests can
run without actually connecting over TLS to a test server. (The integration tests
do actual TLS.)
Logging
This code uses ALOGV throughout for low-priority logging, and does not use
ALOGD. ALOGV is disabled by default, unless activated by #define LOG_NDEBUG 0.
(ALOGD is not disabled by default, requiring extra measures to avoid spamming the
system log in production builds.)