Change-Id: I9ec1a8c87e29ffd4fabef68beb6d094aa7d9a215
Reviewed-on: https://boringssl-review.googlesource.com/8795
Reviewed-by: David Benjamin <davidben@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
CQ-Verified: CQ bot account: commit-bot@chromium.org <commit-bot@chromium.org>
This is already duplicated between client and server and otherwise will
get duplicated yet again for TLS 1.3.
Change-Id: Ia8a352f9bc76fab0f88c1629d08a1da4c13d2510
Reviewed-on: https://boringssl-review.googlesource.com/8778
Reviewed-by: David Benjamin <davidben@google.com>
This will get shared between TLS 1.2 and 1.3.
Change-Id: I9c0d73a087942ac4f8f2075a44bd55647c0dd70b
Reviewed-on: https://boringssl-review.googlesource.com/8777
Reviewed-by: David Benjamin <davidben@google.com>
These will all want to be shared with the TLS 1.3 handshake.
Change-Id: I4e50dc0ed2295d43c7ae800015d71c1406311801
Reviewed-on: https://boringssl-review.googlesource.com/8776
Reviewed-by: David Benjamin <davidben@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
CQ-Verified: CQ bot account: commit-bot@chromium.org <commit-bot@chromium.org>
For now, skip the 1.2 -> 1.1 signal since that will affect shipping
code. We may as well enable it too, but wait until things have settled
down. This implements the version in draft-14 since draft-13's isn't
backwards-compatible.
Change-Id: I46be43e6f4c5203eb4ae006d1c6a2fe7d7a949ec
Reviewed-on: https://boringssl-review.googlesource.com/8724
Reviewed-by: David Benjamin <davidben@google.com>
Implement in both C and Go. To test this, route config into all the
sign.go functions so we can expose bugs to skip the check.
Unfortunately, custom private keys are going to be a little weird since
we can't check their curve type. We may need to muse on what to do here.
Perhaps the key type bit should return an enum that includes the curve?
It's weird because, going forward, hopefully all new key types have
exactly one kind of signature so key type == sig alg == sig alg prefs.
Change-Id: I1f487ec143512ead931e3392e8be2a3172abe3d2
Reviewed-on: https://boringssl-review.googlesource.com/8701
Reviewed-by: David Benjamin <davidben@google.com>
ssl_verify_* already ought to be checking this, so there's only a need
to check against the configured preferences.
Change-Id: I79bc771969c57f953278e622084641e6e20108e3
Reviewed-on: https://boringssl-review.googlesource.com/8698
Reviewed-by: David Benjamin <davidben@google.com>
Rather than blindly select SHA-1 if we can't find a matching one, act as
if the peer advertised rsa_pkcs1_sha1 and ecdsa_sha1. This means that we
will fail the handshake if no common algorithm may be found.
This is done in preparation for removing the SHA-1 default in TLS 1.3.
Change-Id: I3584947909d3d6988b940f9404044cace265b20d
Reviewed-on: https://boringssl-review.googlesource.com/8695
Reviewed-by: David Benjamin <davidben@google.com>
Instead, in SSL_set_private_key_digest_prefs, convert the NID list to a
sigalgs list. We'll need to add a new API later when custom key callers
are ready to start advertising RSA-PSS.
This removes all callers of tls12_get_hash except inside the signing and
verifying functions.
Change-Id: Ie534f3b736c6ac6ebeb0d7770d489f72e3321865
Reviewed-on: https://boringssl-review.googlesource.com/8693
Reviewed-by: David Benjamin <davidben@google.com>
Instead have ssl3_cert_verify_hash output the hash, since it already
knows it. Also add a missing EVP_PKEY_CTX_set_signature_md call on the
client half. (Although, the call isn't actually necessary.)
Also remove now unnecessary static assert. Since EVP_md5_sha1 is an
EVP_MD itself, EVP_MAX_MD_SIZE is required to fit it already.
Change-Id: Ief74fdbdf08e9f124679475bafba2f6f1d8fc687
Reviewed-on: https://boringssl-review.googlesource.com/8692
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
It still places the current message all over the place, but remove the
bizarre init_num/error/ok split. Now callers get the message length out
of init_num, which mirrors init_msg. Also fix some signedness.
Change-Id: Ic2e97b6b99e234926504ff217b8aedae85ba6596
Reviewed-on: https://boringssl-review.googlesource.com/8690
Reviewed-by: David Benjamin <davidben@google.com>
This machinery is so different between TLS and DTLS that there is no
sense in having them share structures. This switches us to maintaining
the full reassembled message in hm_fragment and get_message just lets
the caller read out of that when ready.
This removes the last direct handshake dependency on init_buf,
ssl3_hash_message.
Change-Id: I4eccfb6e6021116255daead5359a0aa3f4d5be7b
Reviewed-on: https://boringssl-review.googlesource.com/8667
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Both DTLS and TLS still use it, but that will change in the following
commit. This also removes the handshake's knowledge of the
dtls_clear_incoming_messages function.
(It's possible we'll want to get rid of begin_handshake in favor of
allocating it lazily depending on how TLS 1.3 post-handshake messages
end up working out. But this should work for now.)
Change-Id: I0f512788bbc330ab2c947890939c73e0a1aca18b
Reviewed-on: https://boringssl-review.googlesource.com/8666
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
In order to delay the digest of the handshake transcript and unify
around message-based signing callbacks, a copy of the transcript is kept
around until we are sure there is no certificate authentication.
This removes support for SSL_PRIVATE_KEY_METHOD as a client in SSL 3.0.
Change-Id: If8999a19ca021b4ff439319ab91e2cd2103caa64
Reviewed-on: https://boringssl-review.googlesource.com/8561
Reviewed-by: David Benjamin <davidben@google.com>
This replaces the old key_exchange_info APIs and does not require the
caller be aware of the mess around SSL_SESSION management. They
currently have the same bugs around renegotiation as before, but later
work to fix up SSL_SESSION tracking will fix their internals.
For consistency with the existing functions, I've kept the public API at
'curve' rather than 'group' for now. I think it's probably better to
have only one name with a single explanation in the section header
rather than half and half. (I also wouldn't be surprised if the IETF
ends up renaming 'group' again to 'key exchange' at some point. We'll
see what happens.)
Change-Id: I8e90a503bc4045d12f30835c86de64ef9f2d07c8
Reviewed-on: https://boringssl-review.googlesource.com/8565
Reviewed-by: Adam Langley <agl@google.com>
We ended up switching this from a curve to a cipher suite, so the group
ID isn't used. This is in preparation for adding an API for the curve
ID, at which point leaving the protocol constants undefined seems
somewhat bad manners.
Change-Id: Icb8bf4594879dbbc24177551868ecfe89bc2f8c3
Reviewed-on: https://boringssl-review.googlesource.com/8563
Reviewed-by: Adam Langley <agl@google.com>
OpenSSL's SSL_OP_NO_* flags allow discontinuous version ranges. This is a
nuisance for two reasons. First it makes it unnecessarily difficult to answer
"are any versions below TLS 1.3 enabled?". Second the protocol does not allow
discontinuous version ranges on the client anyway. OpenSSL instead picks the
first continous range of enabled versions on the client, but not the server.
This is bizarrely inconsistent. It also doesn't quite do this as the
ClientHello sending logic does this, but not the ServerHello processing logic.
So we actually break some invariants slightly. The logic is also cumbersome in
DTLS which kindly inverts the comparison logic.
First, switch min_version/max_version's storage to normalized versions. Next
replace all the ad-hoc version-related functions with a single
ssl_get_version_range function. Client and server now consistently pick a
contiguous range of versions. Note this is a slight behavior change for
servers. Version-range-sensitive logic is rewritten to use this new function.
BUG=66
Change-Id: Iad0d64f2b7a917603fc7da54c9fc6656c5fbdb24
Reviewed-on: https://boringssl-review.googlesource.com/8513
Reviewed-by: David Benjamin <davidben@google.com>
The signing logic itself still depends on pre-hashed messages and will be fixed
in later commits.
Change-Id: I901b0d99917c311653d44efa34a044bbb9f11e57
Reviewed-on: https://boringssl-review.googlesource.com/8545
Reviewed-by: David Benjamin <davidben@google.com>
They're not necessary.
Change-Id: Ifeb3fae73a8b22f88019e6ef9f9ba5e64ed3cfab
Reviewed-on: https://boringssl-review.googlesource.com/8543
Reviewed-by: David Benjamin <davidben@google.com>
As part of the SignatureAlgorithm change in the TLS 1.3 specification,
the existing signature/hash combinations are replaced with a combined
signature algorithm identifier. This change maintains the existing APIs
while fixing the internal representations. The signing code currently
still treats the SignatureAlgorithm as a decomposed value, which will be
fixed as part of a separate CL.
Change-Id: I0cd1660d74ad9bcf55ce5da4449bf2922660be36
Reviewed-on: https://boringssl-review.googlesource.com/8480
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
buffer buffer buffer buffer buffer. At some point, words lose their meaning if
they're used too many times. Notably, the DTLS code can't decide whether a
"buffered message" is an incoming message to be reassembled or an outgoing
message to be (re)transmitted.
Change-Id: Ibdde5c00abb062c603d21be97aff49e1c422c755
Reviewed-on: https://boringssl-review.googlesource.com/8500
Reviewed-by: Adam Langley <agl@google.com>
This allows us to use CBB for all handshake messages. Now, SSL_PROTOCOL_METHOD
is responsible for implementing a trio of CBB-related hooks to assemble
handshake messages.
Change-Id: I144d3cac4f05b6637bf45d3f838673fc5c854405
Reviewed-on: https://boringssl-review.googlesource.com/8440
Reviewed-by: Adam Langley <agl@google.com>
It has size 7. There's no need for a priority queue structure, especially one
that's O(N^2) anyway.
Change-Id: I7609794aac1925c9bbf3015744cae266dcb79bff
Reviewed-on: https://boringssl-review.googlesource.com/8437
Reviewed-by: Adam Langley <agl@google.com>
The pair was a remnant of some weird statefulness and also ChangeCipherSpec
having a "sequence number" to make the pqueue turn into an array.
Change-Id: Iffd82594314df43934073bd141faee0fc167ed5f
Reviewed-on: https://boringssl-review.googlesource.com/8436
Reviewed-by: Adam Langley <agl@google.com>
Now that retransitting is a lot less stateful, a lot of surrounding code can
lose statefulness too. Rather than this overcomplicated pqueue structure,
hardcode that a handshake flight is capped at 7 messages (actually, DTLS can
only get up to 6 because we don't support NPN or Channel ID in DTLS) and used a
fixed size array.
This also resolves several TODOs.
Change-Id: I2b54c3441577a75ad5ca411d872b807d69aa08eb
Reviewed-on: https://boringssl-review.googlesource.com/8435
Reviewed-by: Adam Langley <agl@google.com>
Now dtls1_do_handshake_write takes in a serialized form of the full message and
writes it. It's a little weird to serialize and deserialize the header a bunch,
but msg_callback requires that we keep the full one around in memory anyway.
Between that and the handshake hash definition, DTLS really wants messages to
mean the assembled header, redundancies and all, so we'll just put together
messages that way.
This also fixes a bug where ssl_do_msg_callback would get passed in garbage
where the header was supposed to be. The buffered messages get sampled before
writing the fragment rather than after.
Change-Id: I4e3b8ce4aab4c4ab4502d5428dfb8f3f729c6ef9
Reviewed-on: https://boringssl-review.googlesource.com/8433
Reviewed-by: Adam Langley <agl@google.com>
It doesn't really convey anything useful. Leave ssl_get_message alone for now
since it's called everywhere in the handshake and I'm about to tweak it
further.
Change-Id: I6f3a74c170e818f624be8fbe5cf6b796353406df
Reviewed-on: https://boringssl-review.googlesource.com/8430
Reviewed-by: Adam Langley <agl@google.com>
There's a __pragma expression which allows this. Android builds us Windows with
MinGW for some reason, so we actually do have to tolerate non-MSVC-compatible
Windows compilers. (Clang for Windows is much more sensible than MinGW and
intentionally mimicks MSVC.)
MinGW doesn't understand MSVC's pragmas and warns a lot. #pragma warning is
safe to suppress, so wrap those to shush them. This also lets us do away with a
few ifdefs.
Change-Id: I1f5a8bec4940d4b2d947c4c1cc9341bc15ec4972
Reviewed-on: https://boringssl-review.googlesource.com/8236
Reviewed-by: Adam Langley <agl@google.com>
While most of OpenSSL's assembly allows out < in too, some of it doesn't.
Upstream seems to not consider this a problem (or, at least, they're failing to
make a decision on whether it is a problem, so we should assume they'll stay
their course). Accordingly, require aliased buffers to exactly align so we
don't have to keep chasing this down.
Change-Id: I00eb3df3e195b249116c68f7272442918d7077eb
Reviewed-on: https://boringssl-review.googlesource.com/8231
Reviewed-by: Adam Langley <agl@google.com>
Decrypting is very easy to do in-place, but encrypting in-place is a hassle.
The rules actually were wrong due to record-splitting. The aliasing prefix and
the alignment prefix actually differ by 1. Take it out for now in preparation
for tightening the aliasing rules.
If we decide to do in-place encrypt later, probably it'd be more useful to
return header + in-place ciphertext + trailer. (That, in turn, needs a
scatter/gather thing on the AEAD thanks to TLS 1.3's padding and record type
construction.) We may also wish to rethink how record-splitting works here.
Change-Id: I0187d39c541e76ef933b7c2c193323164fd8a156
Reviewed-on: https://boringssl-review.googlesource.com/8230
Reviewed-by: Adam Langley <agl@google.com>
Reorder states and functions by where they appear in the handshake. Remove
unnecessary hooks on SSL_PROTOCOL_METHOD.
Change-Id: I78dae9cf70792170abed6f38510ce870707e82ff
Reviewed-on: https://boringssl-review.googlesource.com/8184
Reviewed-by: David Benjamin <davidben@google.com>
It can be folded into dtls1_read_app_data. This code, since it still takes an
output pointer, does not yet process records atomically. (Though, being DTLS,
it probably should...)
Change-Id: I57d60785c9c1dd13b5b2ed158a08a8f5a518db4f
Reviewed-on: https://boringssl-review.googlesource.com/8177
Reviewed-by: David Benjamin <davidben@google.com>
This was probably the worst offender of them all as read_bytes is the wrong
abstraction to begin with. Note this is a slight change in how processing a
record works. Rather than reading one fragment at a time, we process all
fragments in a record and return. The intent here is so that all records are
processed atomically since the connection eventually will not be able to retain
a buffer holding the record.
This loses a ton of (though not quite all yet) those a2b macros.
Change-Id: Ibe4bbcc33c496328de08d272457d2282c411b38b
Reviewed-on: https://boringssl-review.googlesource.com/8176
Reviewed-by: David Benjamin <davidben@google.com>
The business with ssl_record_prefix_len is rather a hassle. Instead, have
tls_open_record always decrypt in-place and give back a CBS to where the body
is.
This way the caller doesn't need to do an extra check all to avoid creating an
invalid pointer and underflow in subtraction.
Change-Id: I4e12b25a760870d8f8a503673ab00a2d774fc9ee
Reviewed-on: https://boringssl-review.googlesource.com/8173
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Alert handling is more-or-less identical across all contexts. Push it down from
read_bytes into the low-level record functions. This also deduplicates the code
shared between TLS and DTLS.
Now the only type mismatch managed by read_bytes is if we get handshake data in
read_app_data.
Change-Id: Ia8331897b304566e66d901899cfbf31d2870194e
Reviewed-on: https://boringssl-review.googlesource.com/8124
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
This is getting a little repetitive.
Change-Id: Ib0fa8ab10149557c2d728b88648381b9368221d9
Reviewed-on: https://boringssl-review.googlesource.com/8126
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Move this logic out of dtls1_read_bytes and into dtls1_get_record. Only trigger
it when reading from the buffer fails. The other one shouldn't be necessary.
This exists to handle the blocking BIO case when the
BIO_CTRL_DGRAM_SET_NEXT_TIMEOUT signal triggers, so we only need to do it when
timeouts actually trigger.
There also doesn't seem to be a need for most of the machinery. The
BIO_set_flags call seems to be working around a deficiency in the underlying
BIO. There also shouldn't be a need to check the handshake state as there
wouldn't be a timer to restart otherwise.
Change-Id: Ic901ccfb5b82aeb409d16a9d32c04741410ad6d7
Reviewed-on: https://boringssl-review.googlesource.com/8122
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
This is easier to deploy, and more obvious. This commit reverts a few
pieces of e25775bc, but keeps most of it.
Change-Id: If8d657a4221c665349c06041bb12fffca1527a2c
Reviewed-on: https://boringssl-review.googlesource.com/8061
Reviewed-by: Adam Langley <agl@google.com>
The TLS 1.3 spec has an explicit nonce construction for AEADs that
requires xoring the IV and sequence number.
Change-Id: I77145e12f7946ffb35ebeeb9b2947aa51058cbe9
Reviewed-on: https://boringssl-review.googlesource.com/8042
Reviewed-by: Adam Langley <agl@google.com>
The 'elliptic_curves' extension is being renamed to 'supported_groups'
in the TLS 1.3 draft, and most of the curve-specific methods are
generalized to groups/group IDs.
Change-Id: Icd1a1cf7365c8a4a64ae601993dc4273802610fb
Reviewed-on: https://boringssl-review.googlesource.com/7955
Reviewed-by: David Benjamin <davidben@google.com>
CECPQ1 is a new key exchange that concatenates the results of an X25519
key agreement and a NEWHOPE key agreement.
Change-Id: Ib919bdc2e1f30f28bf80c4c18f6558017ea386bb
Reviewed-on: https://boringssl-review.googlesource.com/7962
Reviewed-by: David Benjamin <davidben@google.com>
They're completely unused now. The handshake message reassembly logic should
not depend on the state machine. This should partially free it up (ugly as it
is) to be shared with a future TLS 1.3 implementation while, in parallel, it
and the layers below, get reworked. This also cuts down on the number of states
significantly.
Partially because I expect we'd want to get ssl_hash_message_t out of there
too. Having it in common code is fine, but it needs to be in the (supposed to
be) protocol-agnostic handshake state machine, not the protocol-specific
handshake message layer.
Change-Id: I12f9dc57bf433ceead0591106ab165d352ef6ee4
Reviewed-on: https://boringssl-review.googlesource.com/7949
Reviewed-by: Adam Langley <agl@google.com>
Previously, SSL_ECDH_METHOD consisted of two methods: one to produce a
public key to be sent to the peer, and another to produce the shared key
upon receipt of the peer's message.
This API does not work for NEWHOPE, because the client-to-server message
cannot be produced until the server's message has been received by the
client.
Solve this by introducing a new method which consumes data from the
server key exchange message and produces data for the client key
exchange message.
Change-Id: I1ed5a2bf198ca2d2ddb6d577888c1fa2008ef99a
Reviewed-on: https://boringssl-review.googlesource.com/7961
Reviewed-by: David Benjamin <davidben@google.com>
OpenSSL used to only forbid it on the server in plain PSK and allow it on the
client. Enforce it properly on both sides. My read of the rule in RFC 5246 ("A
non-anonymous server can optionally request a certificate") and in RFC 4279
("The Certificate and CertificateRequest payloads are omitted from the
response.") is that client auth happens iff we're certificate-based.
The line in RFC 4279 is under the plain PSK section, but that doesn't make a
whole lot of sense and there is only one diagram. PSK already authenticates
both sides. I think the most plausible interpretation is that this is for
certificate-based ciphers.
Change-Id: If195232c83f21e011e25318178bb45186de707e6
Reviewed-on: https://boringssl-review.googlesource.com/7942
Reviewed-by: David Benjamin <davidben@google.com>
A handshake message can go up to 2^24 bytes = 16MB which is a little large for
the peer to force us to buffer. Accordingly, we bound the size of a
handshake message.
Rather than have a global limit, the existing logic uses a different limit at
each state in the handshake state machine and, for certificates, allows
configuring the maximum certificate size. This is nice in that we engage larger
limits iff the relevant state is reachable from the handshake. Servers without
client auth get a tighter limit "for free".
However, this doesn't work for DTLS due to out-of-order messages and we use a
simpler scheme for DTLS. This scheme also is tricky on optional messages and
makes the handshake <-> message layer communication complex.
Apart from an ignored 20,000 byte limit on ServerHello, the largest
non-certificate limit is the common 16k limit on ClientHello. So this
complexity wasn't buying us anything. Unify everything on the DTLS scheme
except, so as not to regress bounds on client-auth-less servers, also correctly
check for whether client auth is configured. The value of 16k was chosen based
on this value.
(The 20,000 byte ServerHello limit makes no sense. We can easily bound the
ServerHello because servers may not send extensions we don't implement. But it
gets overshadowed by the certificate anyway.)
Change-Id: I00309b16d809a3c2a1543f99fd29c4163e3add81
Reviewed-on: https://boringssl-review.googlesource.com/7941
Reviewed-by: David Benjamin <davidben@google.com>
It's only used in one file.
Change-Id: I5d60cbc02799b22317f5f7593faf25eb8eea0a24
Reviewed-on: https://boringssl-review.googlesource.com/7943
Reviewed-by: David Benjamin <davidben@google.com>
This allows an application to override the default of 1 second, which
is what's instructed in RFC 6347 but is not an absolute requirement.
Change-Id: I0bbb16e31990fbcab44a29325b6ec7757d5789e5
Reviewed-on: https://boringssl-review.googlesource.com/7930
Reviewed-by: David Benjamin <davidben@google.com>
Having bbio be tri-state (not allocated, allocated but not active, and
allocated and active) is confusing.
The extra state is only used in the client handshake, where ClientHello is
special-cased to not go through the buffer while everything else is. This dates
to OpenSSL's initial commit and doesn't seem to do much. I do not believe it
can affect renego as the buffer only affects writes; although OpenSSL accepted
interleave on read (though this logic predates it slightly), it never sent
application data while it believed a handshake was active. The handshake would
always be driven to completion first.
My guess is this was to save a copy since the ClientHello is a one-message
flight so it wouldn't need to be buffered? This is probably not worth the extra
variation in the state. (Especially with the DTLS state machine going through
ClientHello twice and pushing the BIO in between the two. Though I suspect that
was a mistake in itself. If the optimization guess is correct, there was no
need to do that.)
Change-Id: I6726f866e16ee7213cab0c3e6abb133981444d47
Reviewed-on: https://boringssl-review.googlesource.com/7873
Reviewed-by: Adam Langley <agl@google.com>
The DTLS bbio logic is rather problematic, but this shouldn't make things
worse. In the in-handshake case, the new code merges the per-message
(unchecked) BIO_flush calls into one call at the end but otherwise the BIO is
treated as is. Otherwise any behavior around non-block writes should be
preserved.
In the post-handshake case, we now install the buffer when we didn't
previously. On write error, the buffer will have garbage in it, but it will be
discarded, so that will preserve any existing retry behavior. (Arguably the
existing retry behavior is a bug, but that's another matter.)
Add a test for all this, otherwise it is sure to regress. Testing for
record-packing is a little fuzzy, but we can assert ChangeCipherSpec always
shares a record with something.
BUG=57
Change-Id: I8603f20811d502c71ded2943b0e72a8bdc4e46f2
Reviewed-on: https://boringssl-review.googlesource.com/7871
Reviewed-by: Adam Langley <agl@google.com>
This was dropped in d27441a9cb due to lack
of use, but node.js now needs it.
Change-Id: I1e207d4b46fc746cfae309a0ea7bbbc04ea785e8
Reviewed-on: https://boringssl-review.googlesource.com/7270
Reviewed-by: David Benjamin <davidben@google.com>
For TLS, this machinery only exists to swallow no_certificate alerts
which only get sent in an SSL 3.0 codepath anyway. It's much less a
no-op for SSL 3.0 which, strictly speaking, has only a subset of TLS's
alerts.
This gets messy around version negotiation because of the complex
relationship between enc_method, have_version, and version which all get
set at different times. Given that SSL 3.0 is nearly dead and all these
alerts are fatal to the connection anyway, this doesn't seem worth
carrying around. (It doesn't work very well anyway. An SSLv3-only server
may still send a record_overflow alert before version negotiation.)
This removes the last place enc_method is accessed prior to version
negotiation.
Change-Id: I79a704259fca69e4df76bd5a6846c9373f46f5a9
Reviewed-on: https://boringssl-review.googlesource.com/6843
Reviewed-by: Adam Langley <alangley@gmail.com>
This removes the various non-PRF checks from SSL3_ENC_METHOD so that can
have a clearer purpose. It also makes TLS 1.0 through 1.2's
SSL3_ENC_METHOD tables identical and gives us an assert to ensure
nothing accesses the version bits before version negotiation.
Accordingly, ssl_needs_record_splitting was reordered slightly so we
don't rely on enc_method being initialized to TLS 1.2
pre-version-negotiation.
This leaves alert_value as the only part of SSL3_ENC_METHOD which may be
accessed before version negotiation.
Change-Id: If9e299e2ef5511b5fa442b2af654eed054c3e675
Reviewed-on: https://boringssl-review.googlesource.com/6842
Reviewed-by: Adam Langley <alangley@gmail.com>
I got that from the TLS 1.3 draft, but it's kind of silly-looking. X25519
already refers to a Diffie-Hellman primitive.
Also hopefully the WG will split NamedGroups and SignatureAlgorithms per the
recent proposal, so it won't be needed anyway. (Most chatter is about what
hashes should be allowed with what NIST curves, so it seems like people like
the split itself? We'll see.)
Change-Id: I7bb713190001199a3ebd30b67df2c00d29132431
Reviewed-on: https://boringssl-review.googlesource.com/6912
Reviewed-by: Adam Langley <agl@google.com>
We have need to normalize other versions during version negotiation, but
almost all will be post-negotiation. Hopefully later this can be
replaced with a value explicitly stored on the object and we do away
with ssl->version.
Change-Id: I595db9163d0af2e7c083b9a09310179aaa9ac812
Reviewed-on: https://boringssl-review.googlesource.com/6841
Reviewed-by: Adam Langley <alangley@gmail.com>
The various SSL3_ENC_METHODs ought to be defined in the same file their
functions are defined in, so they can be static.
Change-Id: I34a1d3437e8e61d4d50f2be70312e4630ea89c19
Reviewed-on: https://boringssl-review.googlesource.com/6840
Reviewed-by: Adam Langley <alangley@gmail.com>
This is a companion to SSL_get_rc4_state and SSL_get_ivs which doesn't
require poking at internal state. Partly since it aligns with the
current code and partly the off chance we ever need to get
wpa_supplicant's EAP-FAST code working, the API allows one to generate
more key material than is actually in the key block.
Change-Id: I58bc3f2b017482dbb8567dcd0cd754947a95397f
Reviewed-on: https://boringssl-review.googlesource.com/6839
Reviewed-by: Adam Langley <alangley@gmail.com>
There's not much point in putting those in the interface as the
final_finished_mac implementation is itself different between SSL 3.0
and TLS.
Change-Id: I76528a88d255c451ae008f1a34e51c3cb57d3073
Reviewed-on: https://boringssl-review.googlesource.com/6838
Reviewed-by: Adam Langley <alangley@gmail.com>
As things stand now, they don't actually do anything.
Change-Id: I9f8b4cbf38a0dffabfc5265805c52bb8d7a8fb0d
Reviewed-on: https://boringssl-review.googlesource.com/6837
Reviewed-by: Adam Langley <alangley@gmail.com>
Both are connection state rather than configuration state. Notably this
cuts down more of SSL_clear that can't just use ssl_free + ssl_new.
Change-Id: I3c05b3ae86d4db8bd75f1cd21656f57fc5b55ca9
Reviewed-on: https://boringssl-review.googlesource.com/6835
Reviewed-by: Adam Langley <alangley@gmail.com>
It's the same between TLS and SSL 3.0. There's also no need for the
do_change_cipher_spec wrapper (it no longer needs checks to ensure it
isn't called at a bad place). Finally fold the setup_key_block call into
change_cipher_spec.
Change-Id: I7917f48e1a322f5fbafcf1dfb8ad53f66565c314
Reviewed-on: https://boringssl-review.googlesource.com/6834
Reviewed-by: Adam Langley <alangley@gmail.com>
Move the actual SSL_AEAD_CTX swap into the record layer. Also revise the
intermediate state we store between setup_key_block and
change_cipher_state. With SSL_AEAD_CTX_new abstracted out, keeping the
EVP_AEAD around doesn't make much sense. Just store enough to partition
the key block.
Change-Id: I773fb46a2cb78fa570f00c0a89339c15bbb1d719
Reviewed-on: https://boringssl-review.googlesource.com/6832
Reviewed-by: Adam Langley <alangley@gmail.com>
This is a minor regression from
https://boringssl-review.googlesource.com/5235.
If the client, for whatever reason, had an ID-based session but also
supports tickets, it will send non-empty ID + empty ticket extension.
If the ticket extension is non-empty, then the ID is not an ID but a
dummy signaling value, so 5235 avoided looking it up. But if it is
present and empty, the ID is still an ID and should be looked up.
This shouldn't have any practical consequences, except if a server
switched from not supporting tickets and then started supporting it,
while keeping the session cache fixed.
Add a test for this case, and tighten up existing ID vs ticket tests so
they fail if we resume with the wrong type.
Change-Id: Id4d08cd809af00af30a2b67fe3a971078e404c75
Reviewed-on: https://boringssl-review.googlesource.com/6554
Reviewed-by: Adam Langley <alangley@gmail.com>
That we're half and half is really confusing.
Change-Id: I1c2632682e8a3e63d01dada8e0eb3b735ff709ce
Reviewed-on: https://boringssl-review.googlesource.com/6785
Reviewed-by: Adam Langley <agl@google.com>
This unifies the ClientKeyExchange code rather nicely. ServerKeyExchange
is still pretty specialized. For simplicity, I've extended the yaSSL bug
workaround for clients as well as servers rather than route in a
boolean.
Chrome's already banished DHE to a fallback with intention to remove
altogether later, and the spec doesn't say anything useful about
ClientDiffieHellmanPublic encoding, so this is unlikely to cause
problems.
Change-Id: I0355cd1fd0fab5729e8812e4427dd689124f53a2
Reviewed-on: https://boringssl-review.googlesource.com/6784
Reviewed-by: Adam Langley <agl@google.com>
The new curve is not enabled by default.
As EC_GROUP/EC_POINT is a bit too complex for X25519, this introduces an
SSL_ECDH_METHOD abstraction which wraps just the raw ECDH operation. It
also tidies up some of the curve code which kept converting back and
force between NIDs and curve IDs. Now everything transits as curve IDs
except for API entry points (SSL_set1_curves) which take NIDs. Those
convert immediately and act on curve IDs from then on.
Note that, like the Go implementation, this slightly tweaks the order of
operations. The client sees the server public key before sending its
own. To keep the abstraction simple, SSL_ECDH_METHOD expects to
generate a keypair before consuming the peer's public key. Instead, the
client handshake stashes the serialized peer public value and defers
parsing it until it comes time to send ClientKeyExchange. (This is
analogous to what it was doing before where it stashed the parsed peer
public value instead.)
It still uses TLS 1.2 terminology everywhere, but this abstraction should also
be compatible with TLS 1.3 which unifies (EC)DH-style key exchanges.
(Accordingly, this abstraction intentionally does not handle parsing the
ClientKeyExchange/ServerKeyExchange framing or attempt to handle asynchronous
plain RSA or the authentication bits.)
BUG=571231
Change-Id: Iba09dddee5bcdfeb2b70185308e8ab0632717932
Reviewed-on: https://boringssl-review.googlesource.com/6780
Reviewed-by: Adam Langley <agl@google.com>
clang-format keeps getting annoyed at it. Also remove some long-dead
constants.
Change-Id: I61e773f5be1e60ca28f1ea085e3afa7cb2c97b9e
Reviewed-on: https://boringssl-review.googlesource.com/6778
Reviewed-by: Adam Langley <agl@google.com>
In doing so, make the asynchronous portion look more like
ssl3_send_server_key_exchange. This is a considerably simpler structure,
so the save/resume doesn't need any state.
Mostly this means writing out the signature algorithm can now go through
CBB rather than a uint8_t* without bounds check.
Change-Id: If99fcffd0d41a84514c3d23034062c582f1bccb2
Reviewed-on: https://boringssl-review.googlesource.com/6771
Reviewed-by: Adam Langley <agl@google.com>
There is some messiness around saving and restoring the CBB, but this is
still significantly clearer.
Note that the BUF_MEM_grow line is gone in favor of a fixed CBB like the
other functions ported thus far. This line was never necessary as
init_buf is initialized to 16k and none of our key exchanges get that
large. (The largest one can get is DHE_RSA. Even so, it'd take a roughly
30k-bit DH group with a 30k-bit RSA key.)
Having such limits and tight assumptions on init_buf's initial size is
poor (but on par for the old code which usually just blindly assumed the
message would not get too large) and the size of the certificate chain
is much less obviously bounded, so those BUF_MEM_grows can't easily go.
My current plan is convert everything but those which legitimately need
BUF_MEM_grow to CBB, then atomically convert the rest, remove init_buf,
and switch everything to non-fixed CBBs. This will hopefully also
simplify async resumption. In the meantime, having a story for
resumption means the future atomic change is smaller and, more
importantly, relieves some complexity budget in the ServerKeyExchange
code for adding Curve25519.
Change-Id: I1de6af9856caaed353453d92a502ba461a938fbd
Reviewed-on: https://boringssl-review.googlesource.com/6770
Reviewed-by: Adam Langley <agl@google.com>
This relieves some complexity budget for adding Curve25519 to this
code.
This also adds a BN_bn2cbb_padded helper function since this seems to be a
fairly common need.
Change-Id: Ied0066fdaec9d02659abd6eb1a13f33502c9e198
Reviewed-on: https://boringssl-review.googlesource.com/6767
Reviewed-by: Adam Langley <agl@google.com>
Only ECDHE-based ciphers are implemented. To ease the transition, the
pre-standard cipher shares a name with the standard one. The cipher rule parser
is hacked up to match the name to both ciphers. From the perspective of the
cipher suite configuration language, there is only one cipher.
This does mean it is impossible to disable the old variant without a code
change, but this situation will be very short-lived, so this is fine.
Also take this opportunity to make the CK and TXT names align with convention.
Change-Id: Ie819819c55bce8ff58e533f1dbc8bef5af955c21
Reviewed-on: https://boringssl-review.googlesource.com/6686
Reviewed-by: Adam Langley <agl@google.com>
This uses ssl3_read_bytes for now. We still need to dismantle that
function and then invert the handshake state machine, but this gets
things closer to the right shape as an intermediate step and is a large
chunk in itself. It simplifies a lot of the CCS/handshake
synchronization as a lot of the invariants much more clearly follow from
the handshake itself.
Tests need to be adjusted since this changes some error codes. Now all
the CCS/Handshake checks fall through to the usual
SSL_R_UNEXPECTED_RECORD codepath. Most of what used to be a special-case
falls out naturally. (If half of Finished was in the same record as the
pre-CCS message, that part of the handshake record would have been left
unconsumed, so read_change_cipher_spec would have noticed, just like
read_app_data would have noticed.)
Change-Id: I15c7501afe523d5062f0e24a3b65f053008d87be
Reviewed-on: https://boringssl-review.googlesource.com/6642
Reviewed-by: Adam Langley <agl@google.com>
Then deprecate the old functions. Thanks to upstream's
6977e8ee4a718a76351ba5275a9f0be4e530eab5 for the idea.
Change-Id: I916abd6fca2a3b2a439ec9902d9779707f7e41eb
Reviewed-on: https://boringssl-review.googlesource.com/6622
Reviewed-by: Adam Langley <agl@google.com>
It has no callers. I prepped for its removal earlier with
c05697c2c5
and then completely forgot.
Thanks to upstream's 6f78b9e824c053d062188578635c575017b587c5 for
the reminder. Quoth them:
> This only gets used to set a specific curve without actually checking
> that the peer supports it or not and can therefor result in handshake
> failures that can be avoided by selecting a different cipher.
It's also a very confusing API since it does NOT pass ownership of the
EC_KEY to the caller.
Change-Id: I6a00643b3a2d6746e9e0e228b47c2bc9694b0084
Reviewed-on: https://boringssl-review.googlesource.com/6621
Reviewed-by: Adam Langley <agl@google.com>
FIPS is the same as HIGH (but for CHACHA20), so those are redundant.
Likewise, MEDIUM vs HIGH was just RC4. Remove those in favor of
redefining those legacy rules to mean this.
One less field to keep track of in each cipher.
Change-Id: I2b2489cffb9e16efb0ac7d7290c173cac061432a
Reviewed-on: https://boringssl-review.googlesource.com/6515
Reviewed-by: Adam Langley <agl@google.com>
It's redundant with other cipher properties. We can express these in code.
Cipher rule matching gets a little bit complicated due to the confusing legacy
protocol version cipher rules, so add some tests for it. (It's really hard to
grep for uses of them, so I've kept them working to be safe.)
Change-Id: Ic6b3fcd55d76d4a51b31bf7ae629a2da50a7450e
Reviewed-on: https://boringssl-review.googlesource.com/6453
Reviewed-by: Adam Langley <agl@google.com>
The keylog BIO is internally synchronized by the SSL_CTX lock, but an
application may wish to log keys from multiple SSL_CTXs. This is in
preparation for switching Chromium to use a separate SSL_CTX per profile
to more naturally split up the session caches.
It will also be useful for routing up SSLKEYLOGFILE in WebRTC. There,
each log line must be converted to an IPC up from the renderer
processes.
This will require changes in Chromium when we roll BoringSSL.
BUG=458365,webrtc:4417
Change-Id: I2945bdb4def0a9c36e751eab3d5b06c330d66b54
Reviewed-on: https://boringssl-review.googlesource.com/6514
Reviewed-by: Adam Langley <agl@google.com>
TLS resets it in t1_enc.c while DTLS has it sprinkled everywhere.
Change-Id: I78f0f0e646b4dc82a1058199c4b00f2e917aa5bc
Reviewed-on: https://boringssl-review.googlesource.com/6511
Reviewed-by: Adam Langley <agl@google.com>
This exposes the ServerKeyExchange signature hash type used in the most recent
handshake, for histogramming on the client.
BUG=549662
Change-Id: I8a4e00ac735b1ecd2c2df824112c3a0bc62332a7
Reviewed-on: https://boringssl-review.googlesource.com/6413
Reviewed-by: Adam Langley <agl@google.com>
Later when TLS 1.3 comes around, we'll need SSL_CIPHER_get_max_version too. In
the meantime, hide the SSL_TLSV1_2 messiness behind a reasonable API.
Change-Id: Ibcc17cccf48dd99e364d6defdfa5a87d031ecf0a
Reviewed-on: https://boringssl-review.googlesource.com/6452
Reviewed-by: Adam Langley <agl@google.com>
They run through completely different logic as only handshake is fragmented.
This'll make it easier to rewrite the handshake logic in a follow-up.
Change-Id: I9515feafc06bf069b261073873966e72fcbe13cb
Reviewed-on: https://boringssl-review.googlesource.com/6420
Reviewed-by: Adam Langley <agl@google.com>
That function doesn't do anything useful for DTLS. It's meant for tracking the
rest of the record we've already committed to by writing half of one. But one
cannot write half a datagram, so DTLS never tracks this. Just call
ssl_write_buffer_flush straight and don't touch wpend_*.
Change-Id: Ibe191907d64c955c7cfeefba26f5c11ad5e4b939
Reviewed-on: https://boringssl-review.googlesource.com/6418
Reviewed-by: Adam Langley <agl@google.com>
This change reduces unnecessary copying and makes the pre-RFC-7539
nonces 96 bits just like the AES-GCM, AES-CCM, and RFC 7539
ChaCha20-Poly1305 cipher suites. Also, all the symbols related to
the pre-RFC-7539 cipher suites now have "_OLD" appended, in
preparation for adding the RFC 7539 variants.
Change-Id: I1f85bd825b383c3134df0b6214266069ded029ae
Reviewed-on: https://boringssl-review.googlesource.com/6103
Reviewed-by: Adam Langley <alangley@gmail.com>
The internal session cache is keyed on session ID, so this is completely
useless for clients (indeed we never look it up internally). Along the way,
tidy up ssl_update_cache to be more readable. The slight behavior change is
that SSL_CTX_add_session's return code no longer controls the external
callback. It's not clear to me what that could have accomplished. (It can only
fail on allocation error. We only call it for new sessions, so the duplicate
case is impossible.)
The one thing of value the internal cache might have provided is managing the
timeout. The SSL_CTX_flush_sessions logic would flip the not_resumable bit and
cause us not to offer expired sessions (modulo SSL_CTX_flush_sessions's delay
and any discrepancies between the two caches). Instead, just check expiration
when deciding whether or not to offer a session.
This way clients that set SSL_SESS_CACHE_CLIENT blindly don't accidentally
consume gobs of memory.
BUG=531194
Change-Id: If97485beab21874f37737edc44df24e61ce23705
Reviewed-on: https://boringssl-review.googlesource.com/6321
Reviewed-by: Adam Langley <alangley@gmail.com>
A random 32-byte (so 256-bit) session ID is never going to collide with
an existing one. (And, if it does, SSL_CTX_add_session does account for
this, so the server won't explode. Just attempting to resume some
session will fail.)
That logic didn't completely work anyway as it didn't account for
external session caches or multiple connections picking the same ID in
parallel (generation and insertion happen at different times) or
multiple servers sharing one cache. In theory one could fix this by
passing in a sufficiently clever generate_session_id, but no one does
that.
I found no callers of these functions, so just remove them altogether.
Change-Id: I8500c592cf4676de6d7194d611b99e9e76f150a7
Reviewed-on: https://boringssl-review.googlesource.com/6318
Reviewed-by: Adam Langley <alangley@gmail.com>
In doing so, simplify the mess around serializing the public key.
Channel ID specifies that you write x and y concatenated. Rather than
using the X9.62 serialization and chopping bits off, get the affine
coordinates and write them out in the same way we write r and s.
Also unify the P-256 sanity check around SSL_set1_tls_channel_id and
actually check the curve NID.
BUG=468889
Change-Id: I228877b736c9722e368d315064ce3ae6893adfc0
Reviewed-on: https://boringssl-review.googlesource.com/6201
Reviewed-by: Adam Langley <alangley@gmail.com>
Start converting the ones we can right now. Some of the messier ones
resize init_buf rather than assume the initial size is sufficient, so
those will probably wait until init_buf is gone and the handshake's
undergone some more invasive surgery. The async ones will also require
some thought. But some can be incrementally converted now.
BUG=468889
Change-Id: I0bc22e4dca37d9d671a488c42eba864c51933638
Reviewed-on: https://boringssl-review.googlesource.com/6190
Reviewed-by: Adam Langley <alangley@gmail.com>
RSA_PSK is really weird in that it takes a Certificate, but you're not
expected to verify it. It's just a funny way to transmit an RSA key.
(They probably should have used the RSA_EXPORT ServerKeyExchange
spelling.) Some code now already doesn't account for it right around
certificate verification.
Given ECDHE_PSK exists, hopefully there will never be any need to add
this.
Change-Id: Ia64dac28099eaa9021f8d915d45ccbfd62872317
Reviewed-on: https://boringssl-review.googlesource.com/5941
Reviewed-by: Adam Langley <agl@google.com>
Allow configuring digest preferences for the private key. Some
smartcards have limited support for signing digests, notably Windows
CAPI keys and old Estonian smartcards. Chromium used the supports_digest
hook in SSL_PRIVATE_KEY_METHOD to limit such keys to SHA1. However,
detecting those keys was a heuristic, so some SHA256-capable keys
authenticating to SHA256-only servers regressed in the switch to
BoringSSL. Replace this mechanism with an API to configure digest
preference order. This way heuristically-detected SHA1-only keys may be
configured by Chromium as SHA1-preferring rather than SHA1-requiring.
In doing so, clean up the shared_sigalgs machinery somewhat.
BUG=468076
Change-Id: I996a2df213ae4d8b4062f0ab85b15262ca26f3c6
Reviewed-on: https://boringssl-review.googlesource.com/5755
Reviewed-by: Adam Langley <agl@google.com>
EVP_MD_CTX_copy_ex was implemented with a memcpy, which doesn't work well when
some of the pointers need to be copied, and ssl_verify_cert_chain didn't
account for set_ex_data failing.
Change-Id: Ieb556aeda6ab2e4c810f27012fefb1e65f860023
Reviewed-on: https://boringssl-review.googlesource.com/5911
Reviewed-by: Adam Langley <agl@google.com>
Applications may require the stapled OCSP response in order to verify
the certificate within the verification callback.
Change-Id: I8002e527f90c3ce7b6a66e3203c0a68371aac5ec
Reviewed-on: https://boringssl-review.googlesource.com/5730
Reviewed-by: David Benjamin <davidben@chromium.org>
Reviewed-by: Adam Langley <agl@google.com>
Move cert_chain to the SSL_SESSION. Now everything on an SSL_SESSION is
properly serialized. The cert_chain field is, unfortunately, messed up
since it means different things between client and server.
There exists code which calls SSL_get_peer_cert_chain as both client and
server and assumes the existing semantics for each. Since that function
doesn't return a newly-allocated STACK_OF(X509), normalizing between the
two formats is a nuisance (we'd either need to store both cert_chain and
cert_chain_full on the SSL_SESSION or create one of the two variants
on-demand and stash it into the SSL).
This CL does not resolve this and retains the client/server difference
in SSL_SESSION. The SSL_SESSION serialization is a little inefficient
(two copies of the leaf certificate) for a client, but clients don't
typically serialize sessions. Should we wish to resolve it in the
future, we can use a different tag number. Because this was historically
unserialized, existing code must already allow for cert_chain not being
preserved across i2d/d2i.
In keeping with the semantics of retain_only_sha256_of_client_certs,
cert_chain is not retained when that flag is set.
Change-Id: Ieb72fc62c3076dd59750219e550902f1ad039651
Reviewed-on: https://boringssl-review.googlesource.com/5759
Reviewed-by: Adam Langley <agl@google.com>
It's completely redundant with the copy in the SSL_SESSION except it
isn't serialized.
Change-Id: I1d95a14cae064c599e4bab576df1dd156da4b81c
Reviewed-on: https://boringssl-review.googlesource.com/5757
Reviewed-by: Adam Langley <agl@google.com>
Gets another field out of the SSL_SESSION.
Change-Id: I9a27255533f8e43e152808427466ec1306cfcc60
Reviewed-on: https://boringssl-review.googlesource.com/5756
Reviewed-by: Adam Langley <agl@google.com>
This begins decoupling the transport from the SSL state machine. The buffering
logic is hidden behind an opaque API. Fields like ssl->packet and
ssl->packet_length are gone.
ssl3_get_record and dtls1_get_record now call low-level tls_open_record and
dtls_open_record functions that unpack a single record independent of who owns
the buffer. Both may be called in-place. This removes ssl->rstate which was
redundant with the buffer length.
Future work will push the buffer up the stack until it is above the handshake.
Then we can expose SSL_open and SSL_seal APIs which act like *_open_record but
return a slightly larger enum due to other events being possible. Likewise the
handshake state machine will be detached from its buffer. The existing
SSL_read, SSL_write, etc., APIs will be implemented on top of SSL_open, etc.,
combined with ssl_read_buffer_* and ssl_write_buffer_*. (Which is why
ssl_read_buffer_extend still tries to abstract between TLS's and DTLS's fairly
different needs.)
The new buffering logic does not support read-ahead (removed previously) since
it lacks a memmove on ssl_read_buffer_discard for TLS, but this could be added
if desired. The old buffering logic wasn't quite right anyway; it tried to
avoid the memmove in some cases and could get stuck too far into the buffer and
not accept records. (The only time the memmove is optional is in DTLS or if
enough of the record header is available to know that the entire next record
would fit in the buffer.)
The new logic also now actually decrypts the ciphertext in-place again, rather
than almost in-place when there's an explicit nonce/IV. (That accidentally
switched in https://boringssl-review.googlesource.com/#/c/4792/; see
3d59e04bce96474099ba76786a2337e99ae14505.)
BUG=468889
Change-Id: I403c1626253c46897f47c7ae93aeab1064b767b2
Reviewed-on: https://boringssl-review.googlesource.com/5715
Reviewed-by: Adam Langley <agl@google.com>
This is a simpler implementation than OpenSSL's, lacking responder IDs
and request extensions support. This mirrors the client implementation
already present.
Change-Id: I54592b60e0a708bfb003d491c9250401403c9e69
Reviewed-on: https://boringssl-review.googlesource.com/5700
Reviewed-by: Adam Langley <agl@google.com>
They're not called (new in 1.0.2). We actually may well need to
configure these later to strike ECDSA from the list on Chrome/XP
depending on what TLS 1.3 does, but for now striking it from the cipher
suite list is both necessary and sufficient. I think we're better off
removing these for now and adding new APIs later if we need them.
(This API is weird. You pass in an array of NIDs that must be even
length and alternating between hash and signature NID. We'd also need a
way to query the configured set of sigalgs to filter away. Those used to
exist but were removed in
https://boringssl-review.googlesource.com/#/c/5347/. SSL_get_sigalgs is
an even uglier API and doesn't act on the SSL_CTX.)
And with that, SSL_ctrl and SSL_CTX_ctrl can *finally* be dropped. Don't
leave no-op wrappers; anything calling SSL_ctrl and SSL_CTX_ctrl should
instead switch to the wrapper macros.
BUG=404754
Change-Id: I5d465cd27eef30d108eeb6de075330c9ef5c05e8
Reviewed-on: https://boringssl-review.googlesource.com/5675
Reviewed-by: Adam Langley <agl@google.com>
This change stores the size of the group/modulus (for RSA/DHE) or curve
ID (for ECDHE) in the |SSL_SESSION|. This makes it available for UIs
where desired.
Change-Id: I354141da432a08f71704c9683f298b361362483d
Reviewed-on: https://boringssl-review.googlesource.com/5280
Reviewed-by: Adam Langley <agl@google.com>
This isn't called and, with the fixed-DH client cert types removed, is
only useful if a server wishes to not accept ECDSA certificates or
something.
BUG=404754
Change-Id: I21d8e1a71aedf446ce974fbeadc62f311ae086db
Reviewed-on: https://boringssl-review.googlesource.com/5673
Reviewed-by: Adam Langley <agl@google.com>
These are unused (new as of 1.0.2). Although being able to separate the
two stores is a reasonable thing to do, we hope to remove the
auto-chaining feature eventually. Given that, SSL_CTX_set_cert_store
should suffice. This gets rid of two more ctrl macros.
BUG=404754,486295
Change-Id: Id84de95d7b2ad5a14fc68a62bb2394f01fa67bb4
Reviewed-on: https://boringssl-review.googlesource.com/5672
Reviewed-by: Adam Langley <agl@google.com>
Rather than support arbitrarily many handshake hashes in the general
case (which the PRF logic assumes is capped at two), special-case the
MD5/SHA1 two-hash combination and otherwise maintain a single rolling
hash.
Change-Id: Ide9475565b158f6839bb10b8b22f324f89399f92
Reviewed-on: https://boringssl-review.googlesource.com/5618
Reviewed-by: Adam Langley <agl@google.com>
Rather than iterate over handshake_dgsts itself, it can just call
tls1_handshake_digest.
Change-Id: Ia518da540e47e65b13367eb1af184c0885908488
Reviewed-on: https://boringssl-review.googlesource.com/5617
Reviewed-by: Adam Langley <agl@google.com>
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>
It's purely the PRF function now, although it's still different from the
rest due to the _DEFAULT field being weird.
Change-Id: Iaea7a99cccdc8be4cd60f6c1503df5be2a63c4c5
Reviewed-on: https://boringssl-review.googlesource.com/5614
Reviewed-by: Adam Langley <agl@google.com>
It's a property of just algorithm_enc and hopefully AES-GCM will
continue to be the only true AEAD that requires this. Simpler to just
keep it in ssl_aead_ctx.c.
Change-Id: Ib7c060a3de2fa8590b2dc36c23a5d5fabff43b07
Reviewed-on: https://boringssl-review.googlesource.com/5613
Reviewed-by: Adam Langley <agl@google.com>
Take the sequence number as a parameter. Also replace satsub64be with
the boring thing: convert to uint64_t and subtract normally.
BUG=468889
Change-Id: Icab75f872b5e55cf4e9d68b66934ec91afeb198b
Reviewed-on: https://boringssl-review.googlesource.com/5558
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>
This change mirrors upstream's custom extension API because we have some
internal users that depend on it.
Change-Id: I408e442de0a55df7b05c872c953ff048cd406513
Reviewed-on: https://boringssl-review.googlesource.com/5471
Reviewed-by: Adam Langley <agl@google.com>
This is unused. It seems to be distinct from the automatic chain
building and was added in 1.0.2. Seems to be an awful lot of machinery
that consumers ought to configure anyway.
BUG=486295
Change-Id: If3d4a2761f61c5b2252b37d4692089112fc0ec21
Reviewed-on: https://boringssl-review.googlesource.com/5353
Reviewed-by: Adam Langley <agl@google.com>
This allows us to remove the confusing EVP_PKEY argument to the
SSL_PRIVATE_KEY_METHOD wrapper functions. It also simplifies some of the
book-keeping around the CERT structure, as well as the API for
configuring certificates themselves. The current one is a little odd as
some functions automatically route to the slot while others affect the
most recently touched slot. Others still (extra_certs) apply to all
slots, making them not terribly useful.
Consumers with complex needs should use cert_cb or the early callback
(select_certificate_cb) to configure whatever they like based on the
ClientHello.
BUG=486295
Change-Id: Ice29ffeb867fa4959898b70dfc50fc00137f01f3
Reviewed-on: https://boringssl-review.googlesource.com/5351
Reviewed-by: Adam Langley <agl@google.com>
The distinction was not well-enforced in the code. In fact, it wasn't
even possible to use the RSA_SIGN slot because ssl_set_pkey and
ssl_set_cert would always use the RSA_ENC slot.
A follow-up will fold away the mechanism altogether, but this is an easy
initial simplfication.
BUG=486295
Change-Id: I66b5bf3e6dc243dac7c75924c1c1983538e49060
Reviewed-on: https://boringssl-review.googlesource.com/5349
Reviewed-by: Adam Langley <agl@google.com>
There's no need to store more than the TLS values.
Change-Id: I1a93c7c6aa3254caf7cc09969da52713e6f8acf4
Reviewed-on: https://boringssl-review.googlesource.com/5348
Reviewed-by: Adam Langley <agl@google.com>
After the custom key method support, the EVP_PKEY parameter is somewhat
confusing (to be resolved with the certificate slots removal) as it must
always refer to a private key. ssl3_cert_verify_hash is sometimes used
with the peer's public key. If custom keys were supported on the server,
this would break.
Fix this by passing a pkey_type parameter and letting the caller decide
whether this uses SSL_PRIVATE_KEY_METHOD or not.
Change-Id: I673b92579a84b4561f28026ec0b1c78a6bfee440
Reviewed-on: https://boringssl-review.googlesource.com/5341
Reviewed-by: Adam Langley <agl@google.com>
Missed that there were two of them.
Change-Id: Ibab169ef1f75be9c5ad1ffa0f232629e76a4512d
Reviewed-on: https://boringssl-review.googlesource.com/5332
Reviewed-by: Adam Langley <agl@google.com>
It (incorrectly) thinks some variables are uninitialized. It also gets confused
about some const parameters.
Change-Id: Ie2b3a5336692e7293cf03d6a4cd5345d30b628b3
Reviewed-on: https://boringssl-review.googlesource.com/5330
Reviewed-by: Adam Langley <agl@google.com>
Use more sensible variable names. Also move some work between the helpers and
s3_srvr.c a little; the session lookup functions now only return a new session.
Whether to send a ticket is now an additional output to avoid the enum
explosion around renewal. The actual SSL state is not modified.
This is somewhat cleaner as s3_srvr.c may still reject a session for other
reasons, so we avoid setting ssl->session and ssl->verify_result to a session
that wouldn't be used. (They get fixed up in ssl_get_new_session, so it didn't
actually matter.)
Change-Id: Ib52fabbe993b5e2b7408395a02cdea3dee66df7b
Reviewed-on: https://boringssl-review.googlesource.com/5235
Reviewed-by: Adam Langley <agl@google.com>
This change also switches the behaviour of the client. Previously the
client would send the SCSV rather than the extension, but now it'll only
do that for SSLv3 connections.
Change-Id: I67a04b8abbef2234747c0dac450458deb6b0cd0a
Reviewed-on: https://boringssl-review.googlesource.com/5143
Reviewed-by: Adam Langley <agl@google.com>
This adds a new API, SSL_set_private_key_method, which allows the consumer to
customize private key operations. For simplicity, it is incompatible with the
multiple slots feature (which will hopefully go away) but does not, for now,
break it.
The new method is only routed up for the client for now. The server will
require a decrypt hook as well for the plain RSA key exchange.
BUG=347404
Change-Id: I35d69095c29134c34c2af88c613ad557d6957614
Reviewed-on: https://boringssl-review.googlesource.com/5049
Reviewed-by: Adam Langley <agl@google.com>
Turns out the safer/simpler method still wasn't quite right. :-)
session->sess_cert isn't serialized and deserialized, which is poor. Duplicate
it manually for now. Leave a TODO to get rid of that field altogether as it's
not especially helpful. The certificate-related fields should be in the
session. The others probably have no reason to be preserved on resumptions at
all.
Test by making bssl_shim.cc assert the peer cert chain is there or not as
expected.
BUG=501220
Change-Id: I44034167629720d6e2b7b0b938d58bcab3ab0abe
Reviewed-on: https://boringssl-review.googlesource.com/5170
Reviewed-by: Adam Langley <agl@google.com>
Most of the logic was redundant with checks already made in
ssl3_get_server_certificate. The DHE check was missing an ECDHE half
(and was impossible). The ECDSA check allowed an ECDSA certificate for
RSA. The only non-redundant check was a key usage check which,
strangely, is only done for ECDSA ciphers.
(Although this function called X509_certificate_type and checked sign
bits, those bits in X509_certificate_type are purely a function of the
key type and don't do anything.)
Change-Id: I8df7eccc0ffff49e4cfd778bd91058eb253b13cb
Reviewed-on: https://boringssl-review.googlesource.com/5047
Reviewed-by: Adam Langley <agl@google.com>
This doesn't even change behavior. Unlike local configuration, the peer
can never have multiple certificates anyway. (Even with a renego, the
SESS_CERT is created anew.)
This does lose the implicit certificate type check, but the certificate
type is already checked in ssl3_get_server_certificate and later checked
post-facto in ssl3_check_cert_and_algorithm (except that one seems to
have some bugs like it accepts ECDSA certificates for RSA cipher suites,
to be cleaned up in a follow-up). Either way, we have the certificate
mismatch tests for this.
BUG=486295
Change-Id: I437bb723bb310ad54ee4150eda67c1cfe43377b3
Reviewed-on: https://boringssl-review.googlesource.com/5044
Reviewed-by: Adam Langley <agl@google.com>
With SSL2 gone, there's no need for this split between the abstract
cipher framework and ciphers. Put the cipher suite table in ssl_cipher.c
and move other SSL_CIPHER logic there. With that gone, prune the
cipher-related hooks in SSL_PROTOCOL_METHOD.
BUG=468889
Change-Id: I48579de8bc4c0ea52781ba1b7b57bc5b4919d21c
Reviewed-on: https://boringssl-review.googlesource.com/4961
Reviewed-by: Adam Langley <agl@google.com>
They're redundant with each other.
Change-Id: I17e7ff8c4e0b1486986dd866fd99673fa2aaa494
Reviewed-on: https://boringssl-review.googlesource.com/4959
Reviewed-by: Adam Langley <agl@google.com>
All ciphers are implemented by an EVP_AEAD.
Change-Id: Ifa754599a34e16bf97e1a4b84a271c6d45462c7c
Reviewed-on: https://boringssl-review.googlesource.com/4958
Reviewed-by: Adam Langley <agl@google.com>
The ctrl hooks are left alone since they should just go away.
Simplifying the cipher story will happen in the next CL.
BUG=468889
Change-Id: I979971c90f59c55cd5d17554f1253158b114f18b
Reviewed-on: https://boringssl-review.googlesource.com/4957
Reviewed-by: Adam Langley <agl@google.com>
This still needs significant work, especially the close_notify half, but
clarify the interface and get *_read_bytes out of SSL_PROTOCOL_METHOD.
read_bytes is an implementation detail of those two and get_message
rather than both an implementation detail of get_message for handshake
and a (wholly inappropriate) exposed interface for the other two.
BUG=468889
Change-Id: I7dd23869e0b7c3532ceb2e9dd31ca25ea31128e7
Reviewed-on: https://boringssl-review.googlesource.com/4956
Reviewed-by: Adam Langley <agl@google.com>
The SSL_PROTOCOL_METHOD table needs work, but this makes it clearer
exactly what the shared interface between the upper later and TLS/DTLS
is.
BUG=468889
Change-Id: I38931c484aa4ab3f77964d708d38bfd349fac293
Reviewed-on: https://boringssl-review.googlesource.com/4955
Reviewed-by: Adam Langley <agl@google.com>
Never send the time as a client. Always send it as a server.
Change-Id: I20c55078cfe199d53dc002f6ee5dd57060b086d5
Reviewed-on: https://boringssl-review.googlesource.com/4829
Reviewed-by: Adam Langley <agl@google.com>
When the peer or caller requests a renegotiation, OpenSSL doesn't
renegotiate immediately. It sets a flag to begin a renegotiation as soon
as record-layer read and write buffers are clear. One reason is that
OpenSSL's record layer cannot write a handshake record while an
application data record is being written. The buffer consistency checks
around partial writes will break.
None of these cases are relevant for the client auth hack. We already
require that renego come in at a quiescent part of the application
protocol by forbidding handshake/app_data interleave.
The new behavior is now: when a HelloRequest comes in, if the record
layer is not idle, the renegotiation is rejected as if
SSL_set_reject_peer_renegotiations were set. Otherwise we immediately
begin the new handshake. The server may not send any application data
between HelloRequest and completing the handshake. The HelloRequest may
not be consumed if an SSL_write is pending.
Note this does require that Chromium's HTTP stack not attempt to read
the HTTP response until the request has been written, but the
renegotiation logic already assumes it. Were Chromium to drive the
SSL_read state machine early and the server, say, sent a HelloRequest
after reading the request headers but before we've sent the whole POST
body, the SSL state machine may racily enter renegotiate early, block
writing the POST body on the new handshake, which would break Chromium's
ERR_SSL_CLIENT_AUTH_CERT_NEEDED plumbing.
BUG=429450
Change-Id: I6278240c3bceb5d2e1a2195bdb62dd9e0f4df718
Reviewed-on: https://boringssl-review.googlesource.com/4825
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>
This is considerably less scary than swapping out connection state. It also
fixes a minor bug where, if dtls1_do_write had an alert to dispatch and we
happened to retry during a rexmit, it would use the wrong epoch.
BUG=468889
Change-Id: I754b0d46bfd02f797f4c3f7cfde28d3e5f30c52b
Reviewed-on: https://boringssl-review.googlesource.com/4793
Reviewed-by: Adam Langley <agl@google.com>
tls1_enc is now SSL_AEAD_CTX_{open,seal}. This starts tidying up a bit
of the record-layer logic. This removes rr->input, as encrypting and
decrypting records no longer refers to various globals. It also removes
wrec altogether. SSL3_RECORD is now only used to maintain state about
the current incoming record. Outgoing records go straight to the write
buffer.
This also removes the outgoing alignment memcpy and simply calls
SSL_AEAD_CTX_seal with the parameters as appropriate. From bssl speed
tests, this seems to be faster on non-ARM and a bit of a wash on ARM.
Later it may be worth recasting these open/seal functions to write into
a CBB (tweaked so it can be malloc-averse), but for now they take an
out/out_len/max_out trio like their EVP_AEAD counterparts.
BUG=468889
Change-Id: Ie9266a818cc053f695d35ef611fd74c5d4def6c3
Reviewed-on: https://boringssl-review.googlesource.com/4792
Reviewed-by: Adam Langley <agl@google.com>
Chromium's session cache has since been rewritten and no longer needs to
muck with those functions in tests.
Change-Id: I2defad81513210dca5e105757e04cbb677583251
Reviewed-on: https://boringssl-review.googlesource.com/4788
Reviewed-by: Adam Langley <agl@google.com>
SSL_get_current_cipher is documented by upstream to return the cipher actually
being used. However, because it reads s->session, it returns information
pertaining to the session to be offered if queried before ServerHello or early
in an abbreviated handshake.
Logic around s->session needs more comprehensive cleanup but for just this
function, defining it to be the current outgoing cipher is close to the current
semantics but for fixing the initial state (s->session->cipher is populated
when sending CCS). Store it in the SSL_AEAD_CTX which seems a natural place to
associate state pertaining to a connection half.
BUG=484744
Change-Id: Ife8db27a16615d0dbb2aec65359537243e08af7c
Reviewed-on: https://boringssl-review.googlesource.com/4733
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>
It's completely redundant with the extend bit. If extend is 0, we're reading a
new record, and rbuf.len is passed. Then it needs to get clamped by ssl3_read_n
post alignment anyway. If extend is 1, we're reading the rest of the current
record and max is always n. (For TLS, we actually could just read more, but not
for DTLS. Basically no one sets it on the TLS side of things, so instead, after
WebRTC's broken DTLS handling is fixed, read_ahead can go away altogether and
DTLS/TLS record layers can be separated.)
This removes ssl3_read_n's callers' dependency on ssl3_setup_read_buffer
setting up rbuf.len.
Change-Id: Iaf11535d01017507a52a33b19240f42984d6cf52
Reviewed-on: https://boringssl-review.googlesource.com/4686
Reviewed-by: Adam Langley <agl@google.com>
With DTLSv1_get_timeout de-ctrl-ified, the type checker complains about
OPENSSL_timeval. Existing callers all use the real timeval.
Now that OPENSSL_timeval is not included in any public structs, simply
forward-declare timeval itself in ssl.h and pull in winsock2.h in internal
headers.
Change-Id: Ieaf110e141578488048c28cdadb14881301a2ce1
Reviewed-on: https://boringssl-review.googlesource.com/4682
Reviewed-by: Adam Langley <agl@google.com>
Nothing ever uses those structs. This to avoid having any structs in the
public header which use struct timeval.
In doing so, move the protocol version constants up to ssl.h so dtls1.h
may be empty. This also removes TLS1_get_version and TLS1_get_client_version
as they're unused and depend on TLS1_VERSION_MAJOR. This still lets tls1.h
be included independently from ssl.h (though I don't think anyone ever includes
it...).
Change-Id: Ieac8b90cf94f7f1e742a88bb75c0ee0aa4b1414c
Reviewed-on: https://boringssl-review.googlesource.com/4681
Reviewed-by: Adam Langley <agl@google.com>
With SSL_get0_raw_cipherlist gone, there's no need to hold onto it.
Change-Id: I258f8bfe21cc354211a777660df680df6c49df2a
Reviewed-on: https://boringssl-review.googlesource.com/4616
Reviewed-by: Adam Langley <agl@google.com>
The only place using it is export keying material which can do the
version check inline.
Change-Id: I1893966c130aa43fa97a6116d91bb8b04f80c6fb
Reviewed-on: https://boringssl-review.googlesource.com/4615
Reviewed-by: Adam Langley <agl@google.com>
clang-format got a little confused there.
Change-Id: I46df523e8a7813a2b4e243da3df22851b3393873
Reviewed-on: https://boringssl-review.googlesource.com/4614
Reviewed-by: Adam Langley <agl@google.com>
Document them while I'm here. This adds a new 'preprocessor
compatibility section' to avoid breaking #ifdefs. The CTRL values
themselves are defined to 'doesnt_exist' to catch anything calling
SSL_ctrl directly until that function can be unexported completely.
BUG=404754
Change-Id: Ia157490ea8efe0215d4079556a0c7643273e7601
Reviewed-on: https://boringssl-review.googlesource.com/4553
Reviewed-by: Adam Langley <agl@google.com>
These are never used and no flags are defined anyway.
Change-Id: I206dc2838c5f68d87559a702dcb299b208cc7e1e
Reviewed-on: https://boringssl-review.googlesource.com/4493
Reviewed-by: Adam Langley <agl@google.com>
This is a really dumb API wart. Now that we have a limited set of curves that
are all reasonable, the automatic logic should just always kick in. This makes
set_ecdh_auto a no-op and, instead of making it the first choice, uses it as
the fallback behavior should none of the older curve selection APIs be used.
Currently, by default, server sockets can only use the plain RSA key exchange.
BUG=481139
Change-Id: Iaabc82de766cd00968844a71aaac29bd59841cd4
Reviewed-on: https://boringssl-review.googlesource.com/4531
Reviewed-by: Adam Langley <agl@google.com>
Just about everything depends on SSL_CIPHER. Move it to the top as the first
section in ssl.h. Match the header order and the source file order and document
everything. Also make a couple of minor style guide tweaks.
Change-Id: I6a810dbe79238278ac480e5ced1447055715a79f
Reviewed-on: https://boringssl-review.googlesource.com/4290
Reviewed-by: Adam Langley <agl@google.com>
Rather than shoehorn real ciphers and cipher aliases into the same type (that's
what cipher->valid is used for), treat them separately. Make
ssl_cipher_apply_rule match ciphers by cipher_id (the parameter was ignored and
we assumed that masks uniquely identify a cipher) and remove the special cases
around zero for all the masks. This requires us to remember which fields
default to 0 and which default to ~0u, but the logic is much clearer.
Finally, now that ciphers and cipher aliases are different, don't process rules
which sum together an actual cipher with cipher aliases. This would AND
together the masks for the alias with the values in the cipher and do something
weird around alg_ssl. (alg_ssl is just weird in general, as everyone trying to
disable SSLv3 in OpenSSL recently discovered.)
With all that, we can finally remove cipher->valid which was always one.
Change-Id: Iefcfe159bd6c22dbaea3a5f1517bd82f756dcfe1
Reviewed-on: https://boringssl-review.googlesource.com/4284
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>
