Context: During a handshake, the SSL/TLS handshake logic constructs
an instance of ::mbedtls_ssl_session representing the SSL session
being established. This structure contains information such as the
session's master secret, the peer certificate, or the session ticket
issues by the server (if applicable).
During a renegotiation, the new session is constructed aside the existing
one and destroys and replaces the latter only when the renegotiation is
complete. While conceptually clear, this means that during the renegotiation,
large pieces of information such as the peer's CRT or the session ticket
exist twice in memory, even though the original versions are removed
eventually.
This commit removes the simultaneous presence of two peer CRT chains
in memory during renegotiation, in the following way:
- Unlike in the case of SessionTickets handled in the previous commit,
we cannot simply free the peer's CRT chain from the previous handshake
before parsing the new one, as we need to verify that the peer's end-CRT
hasn't changed to mitigate the 'Triple Handshake Attack'.
- Instead, we perform a binary comparison of the original peer end-CRT
with the one presented during renegotiation, and if it succeeds, we
avoid re-parsing CRT by moving the corresponding CRT pointer from the
old to the new session structure.
- The remaining CRTs in the peer's chain are not affected by the triple
handshake attack protection, and for them we may employ the canonical
approach of freeing them before parsing the remainder of the new chain.
Note that this commit intends to not change any observable behavior
of the stack. In particular:
- The peer's CRT chain is still verified during renegotiation.
- The tail of the peer's CRT chain may change during renegotiation.
Context: During a handshake, the SSL/TLS handshake logic constructs
an instance of ::mbedtls_ssl_session representing the SSL session
being established. This structure contains information such as the
session's master secret, the peer certificate, or the session ticket
issues by the server (if applicable).
During a renegotiation, the new session is constructed aside the existing
one and destroys and replaces the latter only when the renegotiation is
complete. While conceptually clear, this means that during the renegotiation,
large pieces of information such as the peer's CRT or the session ticket
exist twice in memory, even though the original versions are removed
eventually.
This commit starts removing this memory inefficiency by freeing the old
session's SessionTicket before the one for the new session is allocated.
The existing test `x509parse_crt()` for X.509 CRT parsing
so far used the generic parsing API `mbedtls_x509_crt_parse()`
capable of parsing both PEM encoded and DER encoded certficates,
but was actually only used with DER encoded input data. Moreover,
as the purpose of the test is the testing of the core DER X.509 parsing
functionality, not the PEM vs. DER dispatch (which is now already tested
in the various `x509_crt_info()` tests), the call can be replaced with a
direct call to `mbedtls_x509_parse_crt_der()`.
This commit does that, and further adds to the test an analogous
call to the new API `mbedtls_x509_parse_crt_der_nocopy()` to test
copyless parsing of X.509 certificates.
Context:
The existing API `mbedtls_x509_parse_crt_der()` for parsing DER
encoded X.509 CRTs unconditionally makes creates a copy of the
input buffer in RAM. While this comes at the benefit of easy use,
-- specifically: allowing the user to free or re-use the input
buffer right after the call -- it creates a significant memory
overhead, as the CRT is duplicated in memory (at least temporarily).
This might not be tolerable a resource constrained device.
As a remedy, this commit adds a new X.509 API call
`mbedtls_x509_parse_crt_der_nocopy()`
which has the same signature as `mbedtls_x509_parse_crt_der()`
and almost the same semantics, with one difference: The input
buffer must persist and be unmodified for the lifetime of the
established instance of `mbedtls_x509_crt`, that is, until
`mbedtls_x509_crt_free()` is called.
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.
* mbedtls-2.16:
test: Always use `make clean` by itself
list-symbols.sh: if the build fails, print the build transcript
Document "check-names.sh -v"
all.sh: invoke check-names.sh in print-trace-on-exit mode
Print a command trace if the check-names.sh exits unexpectedly
Only use submodule if present
Update change log
Reword ssl_conf_max_frag_len documentation for clarity
Ignore more generated files: seedfile, apidoc
Improve .gitignore grouping and documentation
Generate tags for Vi, for Emacs and with Global
This commit introduces a new SSL error code
`MBEDTLS_ERR_SSL_VERSION_MISMATCH`
which can be used to indicate operation failure due to a
mismatch of version or configuration.
It is put to use in the implementation of `mbedtls_ssl_session_load()`
to signal the attempt to de-serialize a session which has been serialized
in a build of Mbed TLS using a different version or configuration.
This commit improves the test exercising the behaviour of
session deserialization when facing an unexpected version
or config, by testing ver/cfg corruption at any bit in the
ver/cfg header of the serialized data; previously, it had
only tested the first bit of each byte.
This commit makes use of the added space in the session header to
encode the state of those parts of the compile-time configuration
which influence the structure of the serialized session in the
present version of Mbed TLS. Specifically, these are
- the options which influence the presence/omission of fields
from mbedtls_ssl_session (which is currently shallow-copied
into the serialized session)
- the setting of MBEDTLS_X509_CRT_PARSE_C, which determines whether
the serialized session contains a CRT-length + CRT-value pair after
the shallow-copied mbedtls_ssl_session instance.
- the setting of MBEDTLS_SSL_SESSION_TICKETS, which determines whether
the serialized session contains a session ticket.
This commit adds space for two bytes in the header of serizlied
SSL sessions which can be used to determine the structure of the
remaining serialized session in the respective version of Mbed TLS.
Specifically, if parts of the session depend on whether specific
compile-time options are set or not, the setting of these options
can be encoded in the added space.
This commit doesn't yet make use of the fields.
The format of serialized SSL sessions depends on the version and the
configuration of Mbed TLS; attempts to restore sessions established
in different versions and/or configurations lead to undefined behaviour.
This commit adds an 3-byte version header to the serialized session
generated and cleanly fails ticket parsing in case a session from a
non-matching version of Mbed TLS is presented.
The size of the ticket used in this test dropped from 192 to 143 bytes, so
move all sizes used in this test down 50 bytes. Also, we now need to adapt the
server response size as the default size would otherwise collide with the new
mtu value.
We have explicit recommendations to use US spelling for technical writing, so
let's apply this to code as well for uniformity. (My fingers tend to prefer UK
spelling, so this needs to be fixed in many places.)
sed -i 's/\([Ss]eriali\)s/\1z/g' **/*.[ch] **/*.function **/*.data ChangeLog
This test works regardless of the serialisation format and embedded pointers
in it, contrary to the load-save test, though it requires more maintenance of
the test code (sync the member list with the struct definition).
This uncovered a bug that led to a double-free (in practice, in general could
be free() on any invalid value): initially the session structure is loaded
with `memcpy()` which copies the previous values of pointers peer_cert and
ticket to heap-allocated buffers (or any other value if the input is
attacker-controlled). Now if we exit before we got a chance to replace those
invalid values with valid ones (for example because the input buffer is too
small, or because the second malloc() failed), then the next call to
session_free() is going to call free() on invalid pointers.
This bug is fixed in this commit by always setting the pointers to NULL right
after they've been read from the serialised state, so that the invalid values
can never be used.
(An alternative would be to NULL-ify them when writing, which was rejected
mostly because we need to do it when reading anyway (as the consequences of
free(invalid) are too severe to take any risk), so doing it when writing as
well is redundant and a waste of code size.)
Also, while thinking about what happens in case of errors, it became apparent
to me that it was bad practice to leave the session structure in an
half-initialised state and rely on the caller to call session_free(), so this
commit also ensures we always clear the structure when loading failed.
This test appeared to be passing for the wrong reason, it's not actually not
appropriate for the current implementation. The serialised data contains
values of pointers to heap-allocated buffers. There is no reason these should
be identical after a load-save pair. They just happened to be identical when I
first ran the test due to the place of session_free() in the test code and the
fact that the libc's malloc() reused the same buffers. The test no longer
passes if other malloc() implementations are used (for example, when compiling
with asan which avoids re-using the buffer, probably for better error
detection).
So, disable this test for now (we can re-enable it when we changed how
sessions are serialised, which will be done in a future PR, hence the name of
the dummy macro in depends_on). In the next commit we're going to add a test
that save-load is the identity instead - which will be more work in testing as
it will require checking each field manually, but at least is reliable.
This initial test ensures that a load-save function is the identity. It is so
far incomplete in that it only tests sessions without tickets or certificate.
This will be improved in the next commits.
This allows callers to discover what an appropriate size is. Otherwise they'd
have to either try repeatedly, or allocate an overly large buffer (or some
combination of those).
Adapt documentation an example usage in ssl_client2.
Avoid useless copy with mbedtls_ssl_get_session() before serialising.
Used in ssl_client2 for testing and demonstrating usage, but unfortunately
that means mbedtls_ssl_get_session() is no longer tested, which will be fixed
in the next commit.
This provides basic testing for the session (de)serialisation functions, as
well as an example of how to use them.
Tested locally with tests/ssl-opt.sh -f '^Session resume'.
On client side, this is required for the main use case where of serialising a
session for later resumption, in case tickets are used.
On server side, this doesn't change much as ticket_len will always be 0.
This unblocks testing the functions by using them in ssl_client2, which will
be done in the next commit.
