To prevent dropping the same message over and over again, the UDP proxy
test application programs/test/udp_proxy _logically_ maintains a mapping
from records to the number of times the record has already been dropped,
and stops dropping once a configurable threshold (currently 2) is passed.
However, the actual implementation deviates from this logical view
in two crucial respects:
- To keep the implementation simple and independent of
implementations of suitable map interfaces, it only counts how
many times a record of a given _size_ has been dropped, and
stops dropping further records of that size once the configurable
threshold is passed. Of course, this is not fail-proof, but a
good enough approximation for the proxy, and it allows to use
an inefficient but simple array for the required map.
- The implementation mixes datagram lengths and record lengths:
When deciding whether it is allowed to drop a datagram, it
uses the total datagram size as a lookup index into the map
counting the number of times a package has been dropped. However,
when updating this map, the UDP proxy traverses the datagram
record by record, and updates the mapping at the level of record
lengths.
Apart from this inconsistency, the introduction of the Connection ID
feature leads to yet another problem: The CID length is not part of
the record header but dynamically negotiated during (potentially
encrypted!) handshakes, and it is hence impossible for a passive traffic
analyzer (in this case our UDP proxy) to reliably parse record headers;
especially, it isn't possible to reliably infer the length of a record,
nor to dissect a datagram into records.
The previous implementation of the UDP proxy was not CID-aware and
assumed that the record length would always reside at offsets 11, 12
in the DTLS record header, which would allow it to iterate through
the datagram record by record. As mentioned, this is no longer possible
for CID-based records, and the current implementation can run into
a buffer overflow in this case (because it doesn't validate that
the record length is not larger than what remains in the datagram).
This commit removes the inconsistency in datagram vs. record length
and resolves the buffer overflow issue by not attempting any dissection
of datagrams into records, and instead only counting how often _datagrams_
of a particular size have been dropped.
There is only one practical situation where this makes a difference:
If datagram packing is used by default but disabled on retransmission
(which OpenSSL has been seen to do), it can happen that we drop a
datagram in its initial transmission, then also drop some of its records
when they retransmitted one-by-one afterwards, yet still keeping the
drop-counter at 1 instead of 2. However, even in this situation, we'll
correctly count the number of droppings from that point on and eventually
stop dropping, because the peer will not fall back to using packing
and hence use stable record lengths.
961e677fe3