The handshake hash is initialized from the buffer as soon as the cipher
is known. When adding a message to the transcript, independently update
the buffer and rolling hash, whichever is active. This avoids the
complications around dont_free_handshake_buffer and EMS.

BUG=492371

Change-Id: I3b1065796a50fd1be5d42ead7210c2f253ef0aca
Reviewed-on: https://boringssl-review.googlesource.com/5615
Reviewed-by: Adam Langley <agl@google.com>

The RSA key exchange needs decryption and is still unsupported.

Change-Id: I8c13b74e25a5424356afbe6e97b5f700a56de41f
Reviewed-on: https://boringssl-review.googlesource.com/5467
Reviewed-by: Adam Langley <agl@google.com>

Much of this was done automatically with
  find . -name '*.c' | xargs sed -E -i '' -e 's/(OPENSSL_PUT_ERROR\([a-zA-Z_0-9]+, )[a-zA-Z_0-9]+, ([a-zA-Z_0-9]+\);)/\1\2/'
  find . -name '*.c' | xargs sed -E -i '' -e 's/(OPENSSL_PUT_ERROR\([a-zA-Z_0-9]+, )[a-zA-Z_0-9]+,  ([a-zA-Z_0-9]+\);)/\1\2/'

BUG=468039

Change-Id: I4c75fd95dff85ab1d4a546b05e6aed1aeeb499d8
Reviewed-on: https://boringssl-review.googlesource.com/5276
Reviewed-by: Adam Langley <agl@google.com>

This allows us to merge two of the ssl3_digest_cached_records calls which were
almost, but not completely, redundant. Also catches a missing case: the buffer
may be discarded if doing session resumption but otherwise enabling client
authentication.

BUG=492371

Change-Id: I78e9a4a9cca665e89899ef97b815454c6f5c7e02
Reviewed-on: https://boringssl-review.googlesource.com/4885
Reviewed-by: Adam Langley <agl@google.com>

This is documented as "Only request a client certificate on the initial TLS/SSL
handshake. Do not ask for a client certificate again in case of a
renegotiation." Server-side renegotiation is gone.

I'm not sure this flag has ever worked anyway, dating all the way back to
SSLeay 0.8.1b. ssl_get_new_session overwrites s->session, so the old
session->peer is lost.

Change-Id: Ie173243e189c63272c368a55167b8596494fd59c
Reviewed-on: https://boringssl-review.googlesource.com/4883
Reviewed-by: Adam Langley <agl@google.com>

It's never called and the state is meaningless now.

Change-Id: I5429ec3eb7dc2b789c0584ea88323f0ff18920ae
Reviewed-on: https://boringssl-review.googlesource.com/4826
Reviewed-by: Adam Langley <agl@google.com>

The only case where renego is supported is if we are a client and the
server sends a HelloRequest. That is still needed to support the renego
+ client auth hack in Chrome. Beyond that, no other forms of renego will
work.

The messy logic where the handshake loop is repurposed to send
HelloRequest and the extremely confusing tri-state s->renegotiate (which
makes SSL_renegotiate_pending a lie during the initial handshake as a
server) are now gone. The next change will further simplify things by
removing ssl->s3->renegotiate and the renego deferral logic. There's
also some server-only renegotiation checks that can go now.

Also clean up ssl3_read_bytes' HelloRequest handling. The old logic relied on
the handshake state machine to reject bad HelloRequests which... actually that
code probably lets you initiate renego by sending the first four bytes of a
ServerHello and expecting the peer to read it later.

BUG=429450

Change-Id: Ie0f87d0c2b94e13811fe8e22e810ab2ffc8efa6c
Reviewed-on: https://boringssl-review.googlesource.com/4824
Reviewed-by: Adam Langley <agl@google.com>

BUG=429450

Change-Id: I94846d1fd377bc07044f916d0bb1880e219416df
Reviewed-on: https://boringssl-review.googlesource.com/4821
Reviewed-by: Adam Langley <agl@google.com>

This cuts down on one config knob as well as one case in the renego
combinatorial explosion. Since the only case we care about with renego
is the client auth hack, there's no reason to ever do resumption.
Especially since, no matter what's in the session cache:

- OpenSSL will only ever offer the session it just established,
  whether or not a newer one with client auth was since established.

- Chrome will never cache sessions created on a renegotiation, so
  such a session would never make it to the session cache.

- The new_session + SSL_OP_NO_SESSION_RESUMPTION_ON_RENEGOTIATION
  logic had a bug where it would unconditionally never offer tickets
  (but would advertise support) on renego, so any server doing renego
  resumption against an OpenSSL-derived client must not support
  session tickets.

This also gets rid of s->new_session which is now pointless.

BUG=429450

Change-Id: I884bdcdc80bff45935b2c429b4bbc9c16b2288f8
Reviewed-on: https://boringssl-review.googlesource.com/4732
Reviewed-by: Adam Langley <agl@google.com>

There's multiple different versions of this check, between
s->s3->have_version (only works at some points), s->new_session (really
weird and not actually right), s->renegotiate (fails on the server
because it's always 2 after ClientHello), and s->s3->tmp.finish_md_len
(super confusing). Add an explicit bit with clear meaning. We'll prune
some of the others later; notably s->renegotiate can go away when
initiating renegotiation is removed.

This also tidies up the extensions to be consistent about whether
they're allowed during renego:

- ALPN failed to condition when accepting from the server, so even
  if the client didn't advertise, the server could.

- SCTs now *are* allowed during renego. I think forbidding it was a
  stray copy-paste. It wasn't consistently enforced in both ClientHello
  and ServerHello, so the server could still supply it. Moreover, SCTs
  are part of the certificate, so we should accept it wherever we accept
  certificates, otherwise that session's state becomes incomplete. This
  matches OCSP stapling. (NB: Chrome will never insert a session created
  on renego into the session cache and won't accept a certificate
  change, so this is moot anyway.)

Change-Id: Ic9bd1ebe2a2dbe75930ed0213bf3c8ed8170e251
Reviewed-on: https://boringssl-review.googlesource.com/4730
Reviewed-by: Adam Langley <agl@google.com>

Nothing should call ssl3_setup_read_buffer or ssl3_setup_write_buffer unless it
intends to write into the buffer. This way buffer management can later be an
implementation detail of the record layer.

Change-Id: Idb0effba00e77c6169764843793f40ec37868b61
Reviewed-on: https://boringssl-review.googlesource.com/4687
Reviewed-by: Adam Langley <agl@google.com>

Finally, the ssl stack.

Change-Id: Iea10e302825947da36ad46eaf3e8e2bce060fde2
Reviewed-on: https://boringssl-review.googlesource.com/4518
Reviewed-by: Adam Langley <agl@google.com>

These are all masks of some sort (except id which is a combined version and
cipher), so they should use fixed-size unsigned integers.

Change-Id: I058dd8ad231ee747df4b4fb17d9c1e2cbee21918
Reviewed-on: https://boringssl-review.googlesource.com/4283
Reviewed-by: Adam Langley <agl@google.com>

The rest of ssl/ still includes things everywhere, but this at least fixes the
includes that were implicit from ssl/internal.h.

Change-Id: I7ed22590aca0fe78af84fd99a3e557f4b05f6782
Reviewed-on: https://boringssl-review.googlesource.com/4281
Reviewed-by: Adam Langley <agl@google.com>

Match the other internal headers.

Change-Id: Iff7e2dd06a1a7bf993053d0464cc15638ace3aaa
Reviewed-on: https://boringssl-review.googlesource.com/4280
Reviewed-by: Adam Langley <agl@google.com>

These are the remaining untested cipher suites. Rather than add support in
runner.go, just remove them altogether. Grepping for this is a little tricky,
but nothing enables aNULL (all occurrences disable it), and all occurrences of
["ALL:] seem to be either unused or explicitly disable anonymous ciphers.

Change-Id: I4fd4b8dc6a273d6c04a26e93839641ddf738343f
Reviewed-on: https://boringssl-review.googlesource.com/4258
Reviewed-by: Adam Langley <agl@google.com>

Within the library, only ssl_update_cache read them, so add a dedicated field
to replace that use.

The APIs have a handful of uninteresting callers so I've left them in for now,
but they now always return zero.

Change-Id: Ie4e36fd4ab18f9bff544541d042bf3c098a46933
Reviewed-on: https://boringssl-review.googlesource.com/4101
Reviewed-by: Adam Langley <agl@google.com>

These were added in upstream's 7e159e0133 for
SCTP. As far as I can tell, they were a no-op there too. The corresponding RT
ticket makes no mention of them.

SSL_get_error checks the retry flags of the BIO already. Specifically it checks
BIO_should_read and BIO_should_write, but those two automatically set
BIO_should_retry.

(Minor, but I noticed them idly. One less thing to think about when the state
machines finally unify.)

Change-Id: I17a956a51895fba383063dee574e0fbe3209f9b0
Reviewed-on: https://boringssl-review.googlesource.com/3560
Reviewed-by: Adam Langley <agl@google.com>

I found no users of this. We can restore it if needbe, but I don't expect
anyone to find it useful in its current form. The API is suspect for the same
reasons DTLSv1_listen was. An SSL object is stateful and assumes you already
have the endpoint separated out.

If we ever need it, server-side HelloVerifyRequest and DTLSv1_listen should be
implemented by a separate stateless listener that statelessly handles
cookieless ClientHello + HelloVerifyRequest. Once a ClientHello with a valid
cookie comes in, it sets up a stateful SSL object and passes control along to
that.

Change-Id: I86adc1dfb6a81bebe987784c36ad6634a9a1b120
Reviewed-on: https://boringssl-review.googlesource.com/3480
Reviewed-by: Adam Langley <agl@google.com>

It dates to 2000 from upstream and is only used when serving client auth to
Netscape. It will also get in the way when we get to merging DTLS and TLS
handshake functions because NETSCAPE_HANG_BUG is not valid for DTLS as it is
(the handshake fragmentation code will get confused).

Removing per comment on https://boringssl-review.googlesource.com/#/c/2602/

Change-Id: Ia2d086205bbfed002dc33b2203a47206f373b820
Reviewed-on: https://boringssl-review.googlesource.com/3214
Reviewed-by: Adam Langley <agl@google.com>

This is the DTLS-side equivalent of 6553b379e2.

Change-Id: I5eb7f9d6d5030e375baa1406b00e3166d276dc84
Reviewed-on: https://boringssl-review.googlesource.com/3027
Reviewed-by: Adam Langley <agl@google.com>

This CL removes the last of the EVP_CIPHER codepath in ssl/. The dead code is
intentionally not pruned for ease of review, except in DTLS-only code where
adding new logic to support both, only to remove half, would be cumbersome.

Fixes made:
- dtls1_retransmit_state is taught to retain aead_write_ctx rather than
  enc_write_ctx.
- d1_pkt.c reserves space for the variable-length nonce when echoed into the
  packet.
- dtls1_do_write sizes the MTU based on EVP_AEAD max overhead.
- tls1_change_cipher_state_cipher should not free AEAD write contexts in DTLS.
  This matches the (rather confused) ownership for the EVP_CIPHER contexts.
  I've added a TODO to resolve this craziness.

A follow-up CL will remove all the resultant dead code.

Change-Id: I644557f4db53bbfb182950823ab96d5e4c908866
Reviewed-on: https://boringssl-review.googlesource.com/2699
Reviewed-by: Adam Langley <agl@google.com>

Change-Id: I4dc1463b75b12e15673da32e4945f83aaea123e6

This avoids needing a should_add_to_finished_hash boolean on do_write. The
logic in do_write was a little awkward because do_write would be called
multiple times if the write took several iterations. This also gets complex if
DTLS retransmits are involved. (At a glance, it's not obvious the
BIO_CTRL_DGRAM_MTU_EXCEEDED case actually works.)

Doing it as the handshake message is being prepared avoids this concern. It
also gives a natural point for the extended master secret logic which needs to
do work after the finished hash has been sampled.

As a bonus, we can remove s->d1->retransmitting which was only used to deal
with this issue.

Change-Id: Ifedf23ee4a6c5e08f960d296a6eb1f337a16dc7a
Reviewed-on: https://boringssl-review.googlesource.com/2604
Reviewed-by: Adam Langley <agl@google.com>

Some of the messages did the computation manually which would bite us if we
tried to transplant them between DTLS and TLS. More importantly, it precludes
moving the handshake hash computation from ssl_do_write to
ssl_set_handshake_header.

Change-Id: I9d400deb0720e62cb1ab905242eb0679ad650a46
Reviewed-on: https://boringssl-review.googlesource.com/2600
Reviewed-by: Adam Langley <agl@google.com>

Match the server logic to the client state machine and free if BUF_MEM_grow
fails.

Change-Id: I1a249f7b8c222cd710e969e17a1cba1f469f73e3

SSL3_ENC_METHOD will remain version-specific while SSL_METHOD will become
protocol-specific. This finally removes all the version-specific portions of
SSL_METHOD but the version tag itself.

(SSL3_ENC_METHOD's version-specific bits themselves can probably be handled by
tracking a canonicalized protocol version. It would simplify version
comparisons anyway. The one catch is SSLv3 has a very different table. But
that's a cleanup for future. Then again, perhaps a version-specific method
table swap somewhere will be useful later for TLS 1.3.)

Much of this commit was generated with sed invocation:
    s/method->ssl3_enc/enc_method/g

Change-Id: I2b192507876aadd4f9310240687e562e56e6c0b1
Reviewed-on: https://boringssl-review.googlesource.com/2581
Reviewed-by: Adam Langley <agl@google.com>

These may as well be replaced with assertions. Get them out of the way of the
initialization.

Change-Id: Ie4ab8bdc018e4a1def7d3f6b3b172a77896bfc0a
Reviewed-on: https://boringssl-review.googlesource.com/2563
Reviewed-by: Adam Langley <agl@google.com>

It just inserts extra flushes everywhere and isn't used.

Change-Id: I082e4bada405611f4986ba852dd5575265854036
Reviewed-on: https://boringssl-review.googlesource.com/2456
Reviewed-by: Adam Langley <agl@google.com>

It should be set correctly prior to entering the handshake. Don't mask bugs by
assigning it.

Change-Id: Ib9bca8fad68916b3b242aad8819e3760e59e777a
Reviewed-on: https://boringssl-review.googlesource.com/2443
Reviewed-by: Adam Langley <agl@google.com>

Supporting both schemes seems pointless. Now that s->server and s->state are
set appropriately late and get_ssl_method is gone, the only difference is that
the client/server ones have non-functional ssl_accept or ssl_connect hooks. We
can't lose the generic ones, so let's unify on that.

Note: this means a static linker will no longer drop the client or server
handshake code if unused by a consumer linking statically. However, Chromium
needs the server half anyway for DTLS and WebRTC, so that's probably a lost
cause. Android also exposes server APIs.

Change-Id: I290f5fb4ed558f59fadb5d1f84e9d9c405004c23
Reviewed-on: https://boringssl-review.googlesource.com/2440
Reviewed-by: Adam Langley <agl@google.com>

It's redundant with s->server.

Change-Id: Idb4ca44618477b54f3be5f0630f0295f0708b0f4
Reviewed-on: https://boringssl-review.googlesource.com/2438
Reviewed-by: Adam Langley <agl@google.com>

If the state is SSL_ST_BEFORE, the SSL* was just initialized. Otherwise, we
don't want to call SSL_clear. The one case I found where we do is if a
handshake message is received and someone sets
SSL3_FLAGS_NO_RENEGOTIATE_CIPHERS. This is apparently intended for external
consumers to set, but I see no code in Google that does.

Which is fortunate because it'll trigger SSL_clear. This retains the BIOs but
drops all connection state, including the record. If the client just initiated
renego, that's the ClientHello that's lost. The connection then hangs: the now
reset SSL* wants a ClientHello (under the null cipher because that too's been
dropped) while the peer wants an encrypted ServerHello.

Change-Id: Iddb3e0bb86d39d98155b060f9273a0856f2d1409
Reviewed-on: https://boringssl-review.googlesource.com/2436
Reviewed-by: Adam Langley <agl@google.com>

SSL_clear sets s->state and dtls1_clear sets cookie_len on the server. Setting
cookie_len on the server seems to serve no purpose but to let the callback know
how large the buffer is. This can be done just before calling the callback.

It also avoids a bug where the cookie check can be bypassed, should the server
not specify an app_verify_cookie_cb, by supplying a cookie of all zeros of the
maximum size. (Zero is fine because an empty cookie is rejected.)

The goal here is to avoid needing the SSL_clear calls in the handshake
functions. They are currently needed to fix the cookie_len setting when using
the generic method. (They get set wrong and then flipped back.)

Change-Id: I5095891bc0f7df62d83a9c84312fcf0b84826faa
Reviewed-on: https://boringssl-review.googlesource.com/2435
Reviewed-by: Adam Langley <agl@google.com>

SSL_ST_BEFORE is never standalone. As of upstream's
413c4f45ed0508d2242638696b7665f499d68265, SSL_ST_BEFORE is only ever set paired
with SSL_ST_ACCEPT or SSL_ST_CONNECT.

Conversely, SSL_ST_OK is never paired with SSL_ST_ACCEPT or SSL_ST_CONNECT. As
far as I can tell, this combination has never been possible.

Change-Id: Ifbc8f147be821026cf59f3d5038f0dbad3b0a1d2
Reviewed-on: https://boringssl-review.googlesource.com/2433
Reviewed-by: Adam Langley <agl@google.com>

It should already be assigned, as of upstream's
b31b04d951. I believe these assignments are part
of the reason it used to appear to work. Replace them with assertions. So the
assertions are actually valid, check in SSL_connect / SSL_accept that they are
never called if the socket had been placed in the opposite state. (Or we'd be
in another place where it would have appeared to work with the handshake
functions fixing things afterwards.)

Now the only places handshake_func is set are in SSL_set_{connect,accept}_state
and the method switches.

Change-Id: Ib249212bf4aa889b94c35965a62ca06bdbcf52e1
Reviewed-on: https://boringssl-review.googlesource.com/2432
Reviewed-by: Adam Langley <agl@google.com>

This is a bit of cleanup that probably should have been done at the same time
as 30ddb434bf.

For now, version negotiation is implemented with a method swap. It also
performs this swap on SSL_set_session, but this was neutered in
30ddb434bf. Rather than hackishly neuter it,
remove it outright.  In addition, remove SSL_set_ssl_method. Now all method
swaps are internal: SSLv23_method switch to a version-specific method and
SSL_clear undoing it.

Note that this does change behavior: if an SSL* is created with one
version-specific method and we SSL_set_session to a session from a /different/
version, we would switch to the /other/ version-specific method. This is
extremely confusing, so it's unlikely anyone was actually expecting it.
Version-specific methods in general don't work well.

Change-Id: I72a5c1f321ca9aeb1b52ebe0317072950ba25092
Reviewed-on: https://boringssl-review.googlesource.com/2390
Reviewed-by: Adam Langley <agl@google.com>

This commit fixes a number of crashes caused by malloc failures. They
were found using the -malloc-test=0 option to runner.go which runs tests
many times, causing a different allocation call to fail in each case.

(This test only works on Linux and only looks for crashes caused by
allocation failures, not memory leaks or other errors.)

This is not the complete set of crashes! More can be found by collecting
core dumps from running with -malloc-test=0.

Change-Id: Ia61d19f51e373bccb7bc604642c51e043a74bd83
Reviewed-on: https://boringssl-review.googlesource.com/2320
Reviewed-by: Adam Langley <agl@google.com>

There's not much point in retaining the identity hint in the SSL_SESSION. This
avoids the complexity around setting psk_identity hint on either the SSL or the
SSL_SESSION. Introduce a peer_psk_identity_hint for the client to store the one
received from the server.

This changes the semantics of SSL_get_psk_identity_hint; it now only returns
the value configured for the server. The client learns the hint through the
callback. This is compatible with the one use of this API in conscrypt (it
pulls the hint back out to pass to a callback).

Change-Id: I6d9131636b47f13ac5800b4451436a057021054a
Reviewed-on: https://boringssl-review.googlesource.com/2213
Reviewed-by: Adam Langley <agl@google.com>

This was added in http://rt.openssl.org/Ticket/Display.html?id=2033 to support
a mode where a DTLS socket would statelessly perform the ClientHello /
HelloVerifyRequest portion of the handshake, to be handed off to a socket
specific to this peer address.

This is not used by WebRTC or other current consumers. If we need to support
something like this, it would be cleaner to do the listen portion (cookieless
ClientHello + HelloVerifyRequest) externally and then spin up an SSL instance
on receipt of a cookied ClientHello. This would require a slightly more complex
BIO to replay the second ClientHello but would avoid peppering the DTLS
handshake state with a special short-circuiting mode.

Change-Id: I7a413932edfb62f8b9368912a9a0621d4155f1aa
Reviewed-on: https://boringssl-review.googlesource.com/2220
Reviewed-by: Adam Langley <agl@google.com>

This change implements support for the extended master secret. See
https://tools.ietf.org/html/draft-ietf-tls-session-hash-01
https://secure-resumption.com/

Change-Id: Ifc7327763149ab0894b4f1d48cdc35e0f1093b93
Reviewed-on: https://boringssl-review.googlesource.com/1930
Reviewed-by: David Benjamin <davidben@chromium.org>
Reviewed-by: Adam Langley <agl@google.com>

Remove the old implementation which was excessively general. This mirrors the
SCT support and adds a single boolean flag to request an OCSP response with no
responder IDs, extensions, or frills. The response, if received, is stored on
the SSL_SESSION so that it is available for (re)validation on session
resumption; Chromium revalidates the saved auth parameters on resume.

Server support is unimplemented for now. This API will also need to be adjusted
in the future if we implement RFC 6961.

Change-Id: I533c029b7f7ea622d814d05f934fdace2da85cb1
Reviewed-on: https://boringssl-review.googlesource.com/1671
Reviewed-by: Adam Langley <agl@google.com>

This moves CertificateVerify digest processing to the new
SSL_GET_MESSAGE_DONT_HASH_MESSAGE flag. It also refactors it similarly to
ssl3_send_cert_verify and moves that logic to a common ssl3_cert_verify_hash
function to compute the handshake hash.

This removes a large chunk of duplicate (and divergent!) logic between TLS and
DTLS. It also removes TLS1_FLAGS_KEEP_HANDSHAKE.

Change-Id: Ia63c94f7d76d901bc9c4c33454fbfede411adf63
Reviewed-on: https://boringssl-review.googlesource.com/1633
Reviewed-by: Adam Langley <agl@google.com>

Make the get/send functions match.

ssl3_client_hello -> ssl3_send_client_hello.
ssl3_send_newsession_ticket -> ssl3_send_new_session_ticket.
ssl3_send_client_verify -> ssl3_send_cert_verify

Change-Id: Iea5579479b8a8f392167b8fb3b7e9fe961d0f007
Reviewed-on: https://boringssl-review.googlesource.com/1613
Reviewed-by: Adam Langley <agl@google.com>

It doesn't appear to have ever been implemented on the client. The server code
stopped working anyway because it now skips the ssl_get_message call, so we
never cash in on the reuse_message, attempt to reprocess the repeated
ClientHello, and reject it thinking it's a second MS SGC restart.

Change-Id: Id536846e08460143f6fc0a550bdcc1b26b506b04
Reviewed-on: https://boringssl-review.googlesource.com/1580
Reviewed-by: Adam Langley <agl@google.com>

Some follow-up cleanup to bd30f8e34a.

Change-Id: Id996c6c43f42f4db410e1e612ee1e5e8e0233356
Reviewed-on: https://boringssl-review.googlesource.com/1569
Reviewed-by: Adam Langley <agl@google.com>

In the fixed_ecdh case, it wasn't even implemented, but there was stub code for
it. It complicates the ClientKeyExchange (the client parameters become implicit
in the certificate) and isn't used.

Change-Id: I3627a37042539c90e05e59cd0cb3cd6c56225561
Reviewed-on: https://boringssl-review.googlesource.com/1563
Reviewed-by: Adam Langley <agl@google.com>

I see no internal users and the existence of a THIRD version encoding
complicates all version-checking logic. Also convert another version check to
SSL_IS_DTLS that was missed earlier.

Change-Id: I60d215f57d44880f6e6877889307dc39dbf838f7
Reviewed-on: https://boringssl-review.googlesource.com/1550
Reviewed-by: Adam Langley <agl@google.com>

const-correctness, unsigned long -> size_t.

Change-Id: Ic0c2685a48a0f98396c5753b6077c6c0c3b92326
Reviewed-on: https://boringssl-review.googlesource.com/1540
Reviewed-by: Adam Langley <agl@google.com>

- DTLS server code didn't account for the new ClientHello state. This looks
  like it only matters if a DTLS server uses select_certificate_cb and returns
  asynchronously.

- State A transitions immediately to B and is redundant. No code distinguishes
  A and B.

- The ssl_get_message call transitions to the second state (originally C). This
  makes the explicit transition to C a no-op. More of a problem,
  ssl_get_message may return asynchronously and remain in its second state if the
  handshake body had not completed yet. Fix this by splitting state C in two.
  Combined with the above change, this results in only the top few states getting
  reshuffled.

This fixes the server async tests.

Change-Id: I46703bcd205988b118217b6424ba4f88e731be5a
Reviewed-on: https://boringssl-review.googlesource.com/1412
Reviewed-by: Adam Langley <agl@google.com>