Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2747)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2747)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2747)

There are cases when, if you pass a NULL UI_METHOD, the called
function will use an internal default.  This is well and good, but
there may be cases when this is undesirable and one would rather send
in a UI that does absolutely nothing (sort of a /dev/null).  UI_null()
is the UI_METHOD for this purpose.
Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2744)

This is initial and minimal single-block implementation.
Reviewed-by: NRich Salz <rsalz@openssl.org>

Effectively it's minor size optimization, 5-6% per affected subroutine.
Reviewed-by: NRich Salz <rsalz@openssl.org>

On pre-Skylake best optimization strategy was balancing port-specific
instructions, while on Skylake minimizing the sheer amount appears
more sensible.
Reviewed-by: NRich Salz <rsalz@openssl.org>

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2739)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2739)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2739)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2739)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2738)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2738)

Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2738)

- Reject fractional seconds
- Reject offsets
- Check that the date/time digits are in valid range.
- Add documentation for X509_cmp_time

GH issue 2620
Reviewed-by: NRichard Levitte <levitte@openssl.org>
Reviewed-by: NRich Salz <rsalz@openssl.org>

Since CI is engaged on per merge request basis, it can be wasteful to
run each request through all the tests, especially those resource
consuming. Idea is to mark most of tests as "extended" and provide a
way to opt-in by marking last commit with [extended tests] tag. It's
still not as optimal as one could wish, as decision to skip a test
still requires machine time, and it's taken in configured environment,
i.e. with updates and additional packages installed...
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2292)

Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2292)

Avoid a -Wundef warning in refcount.h
Avoid a -Wundef warning in o_str.c
Avoid a -Wundef warning in testutil.h
Include internal/cryptlib.h before openssl/stack.h
to avoid use of undefined symbol OPENSSL_API_COMPAT.
Reviewed-by: NAndy Polyakov <appro@openssl.org>
Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2712)

The certificate types used to be held in a fixed length array or (if
it was too long) a malloced buffer. This was done to retain binary
compatibility. The code can be simplified now SSL is opaque by always
using a malloced buffer.
Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2733)

Reviewed-by: NRich Salz <rsalz@openssl.org>
Reviewed-by: NAndy Polyakov <appro@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2721)

If ret is allocated, it may be leaked on error.
Reviewed-by: NTim Hudson <tjh@openssl.org>
Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2666)

Reviewed-by: NMatt Caswell <matt@openssl.org>
GH: #2732

Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
GH: #1964

Make sure that we can stop handshake processing and resume it later.
Also check that the cipher list and compression methods are sane.
Unfortunately, we don't have the client-side APIs needed to force
a specific (known) session ID to be sent in the ClientHello, so
that accessor cannot be tested here.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Certain callback APIs allow the callback to request async processing
by trickling a particular error value up the stack to the application
as an error return from the handshake function.  In those cases,
SSL_want() returns a code specific to the type of async processing
needed.

The create_ssl_connection() helper function for the tests is very
helpful for several things, including creating API tests.  However,
it does not currently let us test the async processing functionality
of these callback interfaces, because the special SSL error codes
are treated as generic errors and the helper continues to loop until
it reaches its maximum iteration count.

Add a new parameter, 'want', that indicates an expected/desired
special SSL error code, so that the helper will terminate when
either side reports that error, giving control back to the calling
function and allowing the test to proceed.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

create_ssl_connection() prints out the results if SSL_accept() and/or
SSL_connect() fail, but was reusing the client return value when printing
about SSL_accept() failures.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Plumb things through in the same place as the SNI callback, since
we recommend that the early callback replace (and supplement) the
SNI callback, and add a few test cases.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Provide a callback interface that gives the application the ability
to adjust the nascent SSL object at the earliest stage of ClientHello
processing, immediately after extensions have been collected but
before they have been processed.

This is akin to BoringSSL's "select_certificate_cb" (though it is not
API compatible), and as the name indicates, one major use is to examine
the supplied server name indication and select what certificate to
present to the client.  However, it can also be used to make more
sweeping configuration changes to the SSL object according to the
selected server identity and configuration.  That may include adjusting
the permitted TLS versions, swapping out the SSL_CTX object (as is
traditionally done in a tlsext_servername_callback), changing the
server's cipher list, and more.

We also wish to allow an early callback to indicate that it needs to perform
additional work asynchronously and resume processing later.  To that effect,
refactor the second half of tls_process_client_hello() into a subroutine to be
called at the post-processing stage (including the early callback itself), to
allow the callback to result in remaining in the same work stage for a later
call to succeed.  This requires allocating for and storing the CLIENTHELLO_MSG
in the SSL object to be preserved across such calls, but the storage is
reclaimed after ClientHello processing finishes.

Information about the CliehtHello is available to the callback by means of
accessor functions that can only be used from the early callback.  This allows
extensions to make use of the existing internal parsing machinery without
exposing structure internals (e.g., of PACKET), so that applications do not
have to write fragile parsing code.

Applications are encouraged to utilize an early callback and not use
a servername_callback, in order to avoid unexpected behavior that
occurs due to the relative order of processing between things like
session resumption and the historical servername callback.

Also tidy up nearby style by removing unnecessary braces around one-line
conditional bodies.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Add the new enum value and case statements as appropriate.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Split off the portions that mutate the SSL object into a separate
function that the state machine calls, so that the public API can
be a pure function.  (It still needs the SSL parameter in order
to determine what SSL_METHOD's get_cipher_by_char() routine to use,
though.)

Instead of returning the stack of ciphers (functionality that was
not used internally), require using the output parameter, and add
a separate output parameter for the SCSVs contained in the supplied
octets, if desired.  This lets us move to the standard return value
convention.  Also make both output stacks optional parameters.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Move ssl_bytes_to_cipher_list() to ssl_lib.c and create a public
wrapper around it.  This lets application early callbacks easily get
SSL_CIPHER objects from the raw ciphers bytes without having to
reimplement the parsing code.  In particular, they do not need to
know the details of the sslv2 format ClientHello's ciphersuite
specifications.

Document the new public function, including the arguably buggy behavior
of modifying the supplied SSL object.  On the face of it, such a function
should be able to be pure, just a direct translation of wire octets to
internal data structures.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Now that we have made SCSVs into more of a first-class object, provide
a way for the bytes-to-SSL_CIPHER conversion to actually return them.
Add a flag 'all' to ssl_get_cipher_by_char to indicate that we want
all the known ciphers, not just the ones valid for encryption.  This will,
in practice, let the caller retrieve the SCSVs.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Just as we have a table of ssl3_ciphers, add a table of ssl3_scsvs, to contain
SSL_CIPHER objects for these non-valid ciphers.  This will allow for unified
handling of such indicators, especially as we are preparing to pass them around
between functions.

Since the 'valid' field is not set for the SCSVs, they should not be used
for anything requiring a cryptographic cipher (as opposed to something
being stuck in a cipher-shaped hole in the TLS wire protocol).
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

We'll be adding a field of this type to struct ssl_st in a subsequent
commit, and need the type definition to be in scope already.
Also move up the RAW_EXTENSION definition that it depends on.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Keep track of the length of the pre_proc_exts array.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Modify the API of tls_collect_extensions() to be able to output the number of
extensions that are known (i.e., the length of its 'res' output).  This number
can never be zero on a successful return due to the builtin extensions list,
but use a separate output variable so as to not overload the return value
semantics.

Having this value easily available will give consumers a way to avoid repeating
the calculation.
Reviewed-by: NMatt Caswell <matt@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2279)

Reviewed-by: NAndy Polyakov <appro@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2727)

The library files are built with symbol names as is, while the
application is built with the default uppercase-all-symbols mode.
That's fine for public APIs, because we have __DECC_INCLUDE_PROLOGUE.H
and __DECC_INCLUDE_EPILOGUE.H automatically telling the compiler how
to treat the public header files.  However, we don't have the same
setup for internal library APIs, since they are usually only used by
the libraries.

Because apps/rehash.c uses a library internal header file, we have to
surround that inclusion with the same kind of pragmas found in
__DECC_INCLUDE_PROLOGUE.H and __DECC_INCLUDE_EPILOGUE.H, or we get
unresolved symbols when building no-shared.
Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/2725)