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 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.
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.
This finishes making these functions public. Next step is to get them tested,
but there's currently a blocker for that, see next commit (and the commit
after it for tests).
This commit modifies mbedtls_ssl_get_peer_cid() to also allow passing
NULL pointers in the arguments for the peer's CID value and length, in
case this information is needed.
For example, some users might only be interested in whether the use of
the CID was negotiated, in which case both CID value and length pointers
can be set to NULL. Other users might only be interested in confirming
that the use of CID was negotiated and the peer chose the empty CID,
in which case the CID value pointer only would be set to NULL.
It doesn't make sense to pass a NULL pointer for the CID length but a
non-NULL pointer for the CID value, as the caller has no way of telling
the length of the returned CID - and this case is therefore forbidden.
This commit modifies the CID configuration API mbedtls_ssl_conf_cid_len()
to allow the configuration of the stack's behaviour when receiving an
encrypted DTLS record with unexpected CID.
Currently, the stack silently ignores DTLS frames with an unexpected CID.
However, in a system which performs CID-based demultiplexing before passing
datagrams to the Mbed TLS stack, unexpected CIDs are a sign of something not
working properly, and users might want to know about it.
This commit introduces an SSL error code MBEDTLS_ERR_SSL_UNEXPECTED_CID
which the stack can return in response to an unexpected CID. It will
conditionally be put to use in subsequent commits.
There are two options:
1. Don't set it, and don't use it during record protection,
guarding the respective paths by a check whether TLS or
DTLS is used.
2. Set it to the default value even for TLS, and avoid the
protocol-dependent branch during record protection.
This commit picks option 2.
This commit changes the stack's behaviour when facing a record
with a non-matching CID. Previously, the stack failed in this
case, while now we silently skip over the current record.
Previously, ssl_get_next_record() would fetch 13 Bytes for the
record header and hand over to ssl_parse_record_header() to parse
and validate these. With the introduction of CID-based records, the
record length is not known in advance, and parsing and validating
must happen at the same time. ssl_parse_record_header() is therefore
rewritten in the following way:
1. Fetch and validate record content type and version.
2. If the record content type indicates a record including a CID,
adjust the record header pointers accordingly; here, we use the
statically configured length of incoming CIDs, avoiding any
elaborate CID parsing mechanism or dependency on the record
epoch, as explained in the previous commit.
3. Fetch the rest of the record header (note: this doesn't actually
fetch anything, but makes sure that the datagram fetched in the
earlier call to ssl_fetch_input() contains enough data).
4. Parse and validate the rest of the record header as before.
This commit modifies the code surrounding the invocations of
ssl_decrypt_buf() and ssl_encrypt_buf() to deal with a change
of record content type during CID-based record encryption/decryption.
mbedtls_ssl_context contains pointers in_buf, in_hdr, in_len, ...
which point to various parts of the header of an incoming TLS or
DTLS record; similarly, there are pointers out_buf, ... for
outgoing records.
This commit adds fields in_cid and out_cid which point to where
the CID of incoming/outgoing records should reside, if present,
namely prior to where the record length resides.
Quoting https://tools.ietf.org/html/draft-ietf-tls-dtls-connection-id-04:
The DTLSInnerPlaintext value is then encrypted and the CID added to
produce the final DTLSCiphertext.
struct {
ContentType special_type = tls12_cid; /* 25 */
ProtocolVersion version;
uint16 epoch;
uint48 sequence_number;
opaque cid[cid_length]; // New field
uint16 length;
opaque enc_content[DTLSCiphertext.length];
} DTLSCiphertext;
For outgoing records, out_cid is set in ssl_update_out_pointers()
based on the settings in the current outgoing transform.
For incoming records, ssl_update_in_pointers() sets in_cid as if no
CID was present, and it is the responsibility of ssl_parse_record_header()
to update the field (as well as in_len, in_msg and in_iv) when parsing
records that do contain a CID. This will be done in a subsequent commit.
Finally, the code around the invocations of ssl_decrypt_buf()
and ssl_encrypt_buf() is adapted to transfer the CID from the
input/output buffer to the CID field in the internal record
structure (which is what ssl_{encrypt/decrypt}_buf() uses).
Note that mbedtls_ssl_in_hdr_len() doesn't need change because
it infers the header length as in_iv - in_hdr, which will account
for the CID for records using such.
Using the Connection ID extension increases the maximum record expansion
because
- the real record content type is added to the plaintext
- the plaintext may be padded with an arbitrary number of
zero bytes, in order to prevent leakage of information
through package length analysis. Currently, we always
pad the plaintext in a minimal way so that its length
is a multiple of 16 Bytes.
This commit adapts the various parts of the library to account
for that additional source of record expansion.
Context:
The CID draft does not require that the length of CIDs used for incoming
records must not change in the course of a connection. Since the record
header does not contain a length field for the CID, this means that if
CIDs of varying lengths are used, the CID length must be inferred from
other aspects of the record header (such as the epoch) and/or by means
outside of the protocol, e.g. by coding its length in the CID itself.
Inferring the CID length from the record's epoch is theoretically possible
in DTLS 1.2, but it requires the information about the epoch to be present
even if the epoch is no longer used: That's because one should silently drop
records from old epochs, but not the entire datagrams to which they belong
(there might be entire flights in a single datagram, including a change of
epoch); however, in order to do so, one needs to parse the record's content
length, the position of which is only known once the CID length for the epoch
is known. In conclusion, it puts a significant burden on the implementation
to infer the CID length from the record epoch, which moreover mangles record
processing with the high-level logic of the protocol (determining which epochs
are in use in which flights, when they are changed, etc. -- this would normally
determine when we drop epochs).
Moreover, with DTLS 1.3, CIDs are no longer uniquely associated to epochs,
but every epoch may use a set of CIDs of varying lengths -- in that case,
it's even theoretically impossible to do record header parsing based on
the epoch configuration only.
We must therefore seek a way for standalone record header parsing, which
means that we must either (a) fix the CID lengths for incoming records,
or (b) allow the application-code to configure a callback to implement
an application-specific CID parsing which would somehow infer the length
of the CID from the CID itself.
Supporting multiple lengths for incoming CIDs significantly increases
complexity while, on the other hand, the restriction to a fixed CID length
for incoming CIDs (which the application controls - in contrast to the
lengths of the CIDs used when writing messages to the peer) doesn't
appear to severely limit the usefulness of the CID extension.
Therefore, the initial implementation of the CID feature will require
a fixed length for incoming CIDs, which is what this commit enforces,
in the following way:
In order to avoid a change of API in case support for variable lengths
CIDs shall be added at some point, we keep mbedtls_ssl_set_cid(), which
includes a CID length parameter, but add a new API mbedtls_ssl_conf_cid_len()
which applies to an SSL configuration, and which fixes the CID length that
any call to mbetls_ssl_set_cid() which applies to an SSL context that is bound
to the given SSL configuration must use.
While this creates a slight redundancy of parameters, it allows to
potentially add an API like mbedtls_ssl_conf_cid_len_cb() later which
could allow users to register a callback which dynamically infers the
length of a CID at record header parsing time, without changing the
rest of the API.
The function mbedtls_ssl_hdr_len() returns the length of the record
header (so far: always 13 Bytes for DTLS, and always 5 Bytes for TLS).
With the introduction of the CID extension, the lengths of record
headers depends on whether the records are incoming or outgoing,
and also on the current transform.
Preparing for this, this commit splits mbedtls_ssl_hdr_len() in two
-- so far unmodified -- functions mbedtls_ssl_in_hdr_len() and
mbedtls_ssl_out_hdr_len() and replaces the uses of mbedtls_ssl_hdr_len()
according to whether they are about incoming or outgoing records.
There is no need to change the signature of mbedtls_ssl_{in/out}_hdr_len()
in preparation for its dependency on the currently active transform,
since the SSL context is passed as an argument, and the currently
active transform is referenced from that.
With the introduction of the CID feature, the stack needs to be able
to handle a change of record content type during record protection,
which in particular means that the record content type check will
need to move or be duplicated.
This commit introduces a tiny static helper function which checks
the validity of record content types, which hopefully makes it
easier to subsequently move or duplicate this check.
With the introduction of the CID extension, the record content type
may change during decryption; we must therefore re-consider every
record content type check that happens before decryption, and either
move or duplicate it to ensure it also applies to records whose
real content type is only revealed during decryption.
This commit does this for the silent dropping of unexpected
ApplicationData records in DTLS. Previously, this was caught
in ssl_parse_record_header(), returning
MBEDTLS_ERR_SSL_UNEXPECTED_RECORD which in ssl_get_next_record()
would lead to silent skipping of the record.
When using CID, this check wouldn't trigger e.g. when delayed
encrypted ApplicationData records come on a CID-based connection
during a renegotiation.
This commit moves the check to mbedtls_ssl_handle_message_type()
and returns MBEDTLS_ERR_SSL_NON_FATAL if it triggers, which leads
so silent skipover in the caller mbedtls_ssl_read_record().
The SSL context structure mbedtls_ssl_context contains several pointers
ssl->in_hdr, ssl->in_len, ssl->in_iv, ssl->in_msg pointing to various
parts of the record header in an incoming record, and they are setup
in the static function ssl_update_in_pointers() based on the _expected_
transform for the next incoming record.
In particular, the pointer ssl->in_msg is set to where the record plaintext
should reside after record decryption, and an assertion double-checks this
after each call to ssl_decrypt_buf().
This commit removes the dependency of ssl_update_in_pointers() on the
expected incoming transform by setting ssl->in_msg to ssl->in_iv --
the beginning of the record content (potentially including the IV) --
and adjusting ssl->in_msg after calling ssl_decrypt_buf() on a protected
record.
Care has to be taken to not load ssl->in_msg before calling
mbedtls_ssl_read_record(), then, which was previously the
case in ssl_parse_server_hello(); the commit fixes that.
If a record exhibits an invalid feature only after successful
authenticated decryption, this is a protocol violation by the
peer and should hence lead to connection failure. The previous
code, however, would silently ignore such records. This commit
fixes this.
So far, the only case to which this applies is the non-acceptance
of empty non-AD records in TLS 1.2. With the present commit, such
records lead to connection failure, while previously, they were
silently ignored.
With the introduction of the Connection ID extension (or TLS 1.3),
this will also apply to records whose real content type -- which
is only revealed during authenticated decryption -- is invalid.
In contrast to other aspects of the Connection ID extension,
the CID-based additional data for MAC computations differs from
the non-CID case even if the CID length is 0, because it
includes the CID length.
Quoting the CID draft 04:
- Block Ciphers:
MAC(MAC_write_key, seq_num +
tls12_cid + // New input
DTLSPlaintext.version +
cid + // New input
cid_length + // New input
length_of_DTLSInnerPlaintext + // New input
DTLSInnerPlaintext.content + // New input
DTLSInnerPlaintext.real_type + // New input
DTLSInnerPlaintext.zeros // New input
)
And similar for AEAD and Encrypt-then-MAC.
