Flush following the CCS after an HRR. Only flush the HRR if middlebox
compat is turned off.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4701)
Normally we flush immediately after writing the ClientHello. However if
we are going to write a CCS immediately because we've got early_data to
come, then we should move the flush until after the CCS.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4701)
The CCS may be sent at different times based on whether or not we
sent an HRR earlier. In order to make that decision this commit
also updates things to make sure we remember whether an HRR was
used or not.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4701)
Since we no longer do version negotiation during the processing of an HRR
we do not need the TLSv1.3 specific write transition for ClientHello
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4701)
The new ServerHello format is essentially now the same as the old TLSv1.2
one, but it must additionally include supported_versions. The version
field is fixed at TLSv1.2, and the version negotiation happens solely via
supported_versions.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4701)
The check_fatal macro is supposed to only be called if we are already
expecting to be in the fatal state. The macro asserts that we are and
puts us into the fatal state if not.
This issue combined with the problem fixed in the previous commit meant
that the fuzzer detected a crash at a point in the processing when we
should have already been in the fatal state.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4847)
The recent SSL error overhaul left a case where an error occurs but
SSLfatal() is not called.
Credit to OSSfuzz for finding this issue.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4847)
The most likely explanation for us ending up at this point in the code
is that we were called by the user application incorrectly - so use an
appropriate error code.
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4778)
Follow up from the conversion to use SSLfatal() in the state machine to
clean things up a bit more.
[extended tests]
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4778)
We shouldn't call SSLfatal() multiple times for the same error condition.
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4778)
Sometimes at the top level of the state machine code we know we are
supposed to be in a fatal error condition. This commit adds some sanity
checks to ensure that SSLfatal() has been called.
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4778)
This is an initial step towards using SSLfatal() everywhere. Initially in
this commit and in subsequent commits we focus on the state machine code.
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4778)
Typically if a fatal error occurs three things need to happen:
- Put an error on the error queue
- Send an alert
- Put the state machine into the error state
Although all 3 of these things need to be done every time we hit a fatal
error the responsibilities for doing this are distributed throughout the
code. The place where the error goes on the queue, where the alert gets
sent and where the state machine goes into the error state are almost
invariably different. It has been a common pattern to pass alert codes up
and down the stack to get the alert information from the point in the code
where the error is detected to the point in the code where the alert gets
sent.
This commit provides an SSLfatal() macro (backed by an ossl_statem_fatal
function) that does all 3 of the above error tasks. This is largely a drop
in replacement for SSLerr, but takes a couple of extra parameters (the SSL
object, and an alert code).
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4778)
SNI needs to be consistent before we accept early_data. However a
server may choose to not acknowledge SNI. In that case we have to
expect that a client may send it anyway. We change the consistency
checks so that not acknowledging is treated more a like a "wild card",
accepting any SNI as being consistent.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4738)
We can only send early_data if the SNI is consistent. However it is valid
for the client to set SNI and the server to not use it. This would still be
counted as consistent. OpenSSL client was being overzealous in this check
and disallowing this scenario.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4738)
It's argued that /WX allows to keep better focus on new code, which
motivates its comeback...
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4721)
We currently increment the SSL_CTX stats.sess_accept field in
tls_setup_handshake(), which is invoked from the state machine well
before ClientHello processing would have had a chance to switch
the SSL_CTX attached to the SSL object due to a provided SNI value.
However, stats.sess_accept_good is incremented in tls_finish_handshake(),
and uses the s->ctx.stats field (i.e., the new SSL_CTX that was switched
to as a result of SNI processing). This leads to the confusing
(nonsensical) situation where stats.sess_accept_good is larger than
stats.sess_accept, as the "sess_accept" value was counted on the
s->session_ctx.
In order to provide some more useful numbers, increment
s->ctx.stats.sess_accept after SNI processing if the SNI processing
changed s->ctx to differ from s->session_ctx. To preserve the
property that any given accept is counted only once, make the
corresponding decrement to s->session_ctx.stats.sess_accept when
doing so.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/4549)
For client SSL objects and before any callbacks have had a chance
to be called, we can write the stats accesses using the session_ctx,
which makes sense given that these values are all prefixed with
"sess_".
For servers after a client_hello or servername callback has been
called, retain the existing behavior of modifying the statistics
for the current (non-session) context. This has some value,
in that it allows the statistics to be viewed on a per-vhost level.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/4549)
It is expected that SSL_CTX objects are shared across threads,
and as such we are responsible for ensuring coherent data accesses.
Aligned integer accesses ought to be atomic already on all supported
architectures, but we can be formally correct.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/4549)
Since return is inconsistent, I removed unnecessary parentheses and
unified them.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4541)
The previous commit removed version negotiation on an HRR. However we should
still sanity check the contents of the version field.
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4527)
Previously if a client received an HRR then we would do version negotiation
immediately - because we know we are going to get TLSv1.3. However this
causes a problem when we emit the 2nd ClientHello because we start changing
a whole load of stuff to ommit things that aren't relevant for < TLSv1.3.
The spec requires that the 2nd ClientHello is the same except for changes
required from the HRR. Therefore the simplest thing to do is to defer the
version negotiation until we receive the ServerHello.
Fixes#4292
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4527)
The function tls_check_curve is only called on clients and contains
almost identical functionaity to tls1_check_group_id when called from
a client. Merge the two.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4475)
When an SSL's context is swtiched from a ticket-enabled context to
a ticket-disabled context in the servername callback, no session-id
is generated, so the session can't be resumed.
If a servername callback changes the SSL_OP_NO_TICKET option, check
to see if it's changed to disable, and whether a session ticket is
expected (i.e. the client indicated ticket support and the SSL had
tickets enabled at the time), and whether we already have a previous
session (i.e. s->hit is set).
In this case, clear the ticket-expected flag, remove any ticket data
and generate a session-id in the session.
If the SSL hit (resumed) and switched to a ticket-disabled context,
assume that the resumption was via session-id, and don't bother to
update the session.
Before this fix, the updated unit-tests in 06-sni-ticket.conf would
fail test #4 (server1 = SNI, server2 = no SNI).
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/1529)
Rename tls1_get_curvelist to tls1_get_grouplist, change to void as
it can never fail and remove unnecessary return value checks. Clean
up the code.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/=4412)
Setup EVP_PKEY structure from a group ID in ssl_generate_param_group,
replace duplicate code with this function.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/=4412)
Replace tls1_ec_curve_id2nid() with tls_group_id_lookup() which returns
the TLS_GROUP_INFO for the group.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/=4412)
Instead of storing supported groups in on-the-wire format store
them as parsed uint16_t values. This simplifies handling of groups
as the values can be directly used instead of being converted.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4406)
It is otherwise unclear what all the magic numbers mean.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4349)
"Early callback" is a little ambiguous now that early data exists.
Perhaps "ClientHello callback"?
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4349)
In OpenSSL 1.1.0, when there were no extensions added to the ServerHello,
we did not write the extension data length bytes to the end of the
ServerHello; this is needed for compatibility with old client implementations
that do not support TLS extensions (such as the default configuration of
OpenSSL 0.9.8). When ServerHello extension construction was converted
to the new extensions framework in commit
7da160b0f4, this behavior was inadvertently
limited to cases when SSLv3 was negotiated (and similarly for ClientHellos),
presumably since extensions are not defined at all for SSLv3. However,
extensions for TLS prior to TLS 1.3 have been defined in separate
RFCs (6066, 4366, and 3546) from the TLS protocol specifications, and as such
should be considered an optional protocol feature in those cases.
Accordingly, be conservative in what we send, and skip the extensions block
when there are no extensions to be sent, regardless of the TLS/SSL version.
(TLS 1.3 requires extensions and can safely be treated differently.)
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/4296)
This is actually not all warnings, only return values.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4297)
SNI and ALPN must be set to be consistent with the PSK. Otherwise this is
an error.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/3926)
cryptilib.h is the second.
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4188)
Removed e_os.h from all bar three headers (apps/apps.h crypto/bio/bio_lcl.h and
ssl/ssl_locl.h).
Added e_os.h into the files that need it now.
Directly reference internal/nelem.h when required.
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4188)
Force non-empty padding extension.
When enabled, force the padding extension to be at least 1 byte long.
WebSphere application server cannot handle having an empty
extension (e.g. EMS/EtM) as the last extension in a client hello.
This moves the SigAlgs extension last for TLSv1.2 to avoid this
issue.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3921)
Commit 02f0274e8c moved ALPN processing
into an extension finalization function, as the only documented ordering
requirement from previous commits was that ALPN processing occur after
SNI processing, and SNI processing is performed before the extension
finalization step. However, it is useful for applications'
alpn_select callbacks to run after ciphersuite selection as well -- at
least one application protocol specification (HTTP/2) imposes restrictions
on which ciphersuites are usable with that protocol. Since it is generally
more preferrable to have a successful TLS connection with a default application
protocol than to fail the TLS connection and not be able to have the preferred
application protocol, it is good to give the alpn_select callback information
about the ciphersuite to be used, so that appropriate restrctions can be
enforced in application code.
Accordingly, split the ALPN handling out into a separate tls_handl_alpn()
function akin to tls_handle_status_request(), called from
tls_post_process_client_hello(). This is an alternative to resuscitating
ssl_check_clienthello_tlsext_late(), something of an awkwward name itself.
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4070)
If s->hit is set, s->session corresponds to a session created on
a previous connection, and is a data structure that is potentially
shared across other SSL objects. As such, there are thread-safety
issues with modifying the structure without taking its lock (and
of course all corresponding read accesses would also need to take
the lock as well), which have been observed to cause double-frees.
Regardless of thread-safety, the resumed session object is intended
to reflect parameters of the connection that created the session,
and modifying it to reflect the parameters from the current connection
is confusing. So, modifications to the session object during
ClientHello processing should only be performed on new connections,
i.e., those where s->hit is not set.
The code mostly got this right, providing such checks when processing
SNI and EC point formats, but the supported groups (formerly
supported curves) extension was missing it, which is fixed by this commit.
However, TLS 1.3 makes the suppported_groups extension mandatory
(when using (EC)DHE, which is the normal case), checking for the group
list in the key_share extension processing. But, TLS 1.3 only [0] supports
session tickets for session resumption, so the session object in question
is the output of d2i_SSL_SESSION(), and will not be shared across SSL
objects. Thus, it is safe to modify s->session for TLS 1.3 connections.
[0] A psk_find_session callback can also be used, but the restriction that
each callback execution must produce a distinct SSL_SESSION structure
can be documented when the psk_find_session callback documentation is
completed.
Reviewed-by: Andy Polyakov <appro@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4123)
Give each SSL object it's own DRBG, chained to the parent global
DRBG which is used only as a source of randomness into the per-SSL
DRBG. This is used for all session, ticket, and pre-master secret keys.
It is NOT used for ECDH key generation which use only the global
DRBG. (Doing that without changing the API is tricky, if not impossible.)
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/4050)
If RAND_add wraps around, XOR with existing. Add test to drbgtest that
does the wrap-around.
Re-order seeding and stop after first success.
Add RAND_poll_ex()
Use the DF and therefore lower RANDOMNESS_NEEDED. Also, for child DRBG's,
mix in the address as the personalization bits.
Centralize the entropy callbacks, from drbg_lib to rand_lib.
(Conceptually, entropy is part of the enclosing application.)
Thanks to Dr. Matthias St Pierre for the suggestion.
Various code cleanups:
-Make state an enum; inline RANDerr calls.
-Add RAND_POLL_RETRIES (thanks Pauli for the idea)
-Remove most RAND_seed calls from rest of library
-Rename DRBG_CTX to RAND_DRBG, etc.
-Move some code from drbg_lib to drbg_rand; drbg_lib is now only the
implementation of NIST DRBG.
-Remove blocklength
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/4019)
Move the definition of ossl_assert() out of e_os.h which is intended for OS
specific things. Instead it is moved into internal/cryptlib.h.
This also changes the definition to remove the (int) cast.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4073)
If a new_session_cb is set then it was only ever getting invoked if !s->hit
is true. This is sensible for <=TLSv1.2 but does not work for TLSv1.3.
Fixes#4045
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/4068)
Some extensions were being displayed twice, before they were parsed, and
again after they were parsed.
The supported_versions extension was not being fully displayed, as it
was processed differently than other extensions.
Move the debug callback to where the extensions are first collected, to
catch all the extensions as they come in, so they are ordered correctly.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3911)
Fixes: issue #3747
make SSL_CIPHER_standard_name globally available and introduce a new
function OPENSSL_cipher_name.
A new option '-convert' is also added to 'openssl ciphers' app.
Documentation and test cases are added.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/3859)
TLSv1.3 draft-21 requires the ticket nonce to be at least 1 byte in length.
However NSS sends a zero length nonce. This is actually ok because the next
draft will allow zero length nonces anyway, so we should tolerate this.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3957)
In most scenarios the length of the input data is the hashsize, or 0 if
the data is NULL. However with the new ticket_nonce changes the length can
be different.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/3852)
This just adds the processing for sending and receiving the newly added
ticket_nonce field. It doesn't actually use it yet.
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/3852)
Following on from the previous commit this fixes another instance where
we need to treat a -ve return from EVP_DigestVerify() as a bad signature.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3756)
Prior to 72ceb6a we treated all failures from the call to
EVP_DigestVerifyFinal() as if it were a bad signature, and failures in
EVP_DigestUpdate() as an internal error. After that commit we replaced
this with the one-shot function EVP_DigestVerify() and treated a 0 return
as a bad signature and a negative return as an internal error. However,
some signature errors can be negative (e.g. according to the docs if the
form of the signature is wrong). Therefore we should treat all <=0
returns as a bad signature.
This fixes a boringssl test failure.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3756)
initialize some local variables
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3741)
The value of BIO_CTRL_DGRAM_SET_PEEK_MODE was clashing with the value for
BIO_CTRL_DGRAM_SCTP_SET_IN_HANDSHAKE. In an SCTP enabled build
BIO_CTRL_DGRAM_SCTP_SET_IN_HANDSHAKE was used unconditionally with
the reasoning that it would be ignored if SCTP wasn't in use. Unfortunately
due to this clash, this wasn't the case. The BIO ended up going into peek
mode and was continually reading the same data over and over - throwing it
away as a replay.
Fixes#3723
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3724)