With the previous DTLS change, the dispatch layer only cares about the
end of the handshake to know when to drop the current message. TLS 1.3
post-handshake messages will need a similar hook, so convert it to this
lower-level one.
BUG=83
Change-Id: I4c8c3ba55ba793afa065bf261a7bccac8816c348
Reviewed-on: https://boringssl-review.googlesource.com/8989
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: Adam Langley <agl@google.com>
CQ-Verified: CQ bot account: commit-bot@chromium.org <commit-bot@chromium.org>
This is in preparation for switching finish_handshake to a
release_current_message hook. finish_handshake in DTLS is also
responsible for releasing any memory associated with extra messages in
the handshake.
Except that's not right and we need to make it an error anyway. Given
that the rest of the DTLS dispatch layer already strongly assumes there
is only one message in epoch one, putting the check in the fragment
processing works fine enough. Add tests for this.
This will certainly need revising when DTLS 1.3 happens (perhaps just a
version check, perhaps bringing finish_handshake back as a function that
can fail... which means we need a state just before SSL_ST_OK), but DTLS
1.3 post-handshake messages haven't really been written down, so let's
do the easy thing for now and add a test for when it gets more
interesting.
This removes the sequence number reset in the DTLS code. That reset
never did anything becase we don't and never will renego. We should make
sure DTLS 1.3 does not bring the reset back for post-handshake stuff.
(It was wrong in 1.2 too. Penultimate-flight retransmits and renego
requests are ambiguous in DTLS.)
BUG=83
Change-Id: I33d645a8550f73e74606030b9815fdac0c9fb682
Reviewed-on: https://boringssl-review.googlesource.com/8988
Reviewed-by: Adam Langley <agl@google.com>
This is the equivalent of FragmentAcrossChangeCipherSuite for DTLS. It
is possible for us to, while receiving pre-CCS handshake messages, to
buffer up a message with sequence number meant for a post-CCS Finished.
When we then get to the new epoch and attempt to read the Finished, we
will process the buffered Finished although it was sent with the wrong
encryption.
Move ssl_set_{read,write}_state to SSL_PROTOCOL_METHOD hooks as this is
a property of the transport. Notably, read_state may fail. In DTLS
check the handshake buffer size. We could place this check in
read_change_cipher_spec, but TLS 1.3 has no ChangeCipherSpec message, so
we will need to implement this at the cipher change point anyway. (For
now, there is only an assert on the TLS side. This will be replaced with
a proper check in TLS 1.3.)
Change-Id: Ia52b0b81e7db53e9ed2d4f6d334a1cce13e93297
Reviewed-on: https://boringssl-review.googlesource.com/8790
Reviewed-by: Steven Valdez <svaldez@google.com>
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>
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>
They're not necessary.
Change-Id: Ifeb3fae73a8b22f88019e6ef9f9ba5e64ed3cfab
Reviewed-on: https://boringssl-review.googlesource.com/8543
Reviewed-by: David Benjamin <davidben@google.com>
This file is still kind of a mess, but put the two halves together at least.
Change-Id: Ib21d9c4a7f4864cf80e521f7d0ebec029e5955a1
Reviewed-on: https://boringssl-review.googlesource.com/8502
Reviewed-by: Adam Langley <agl@google.com>
MSVC 2015 seems to support it just fine.
Change-Id: I9c91c18c260031e6024480d1f57bbb334ed7118c
Reviewed-on: https://boringssl-review.googlesource.com/8501
Reviewed-by: Adam Langley <agl@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>
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>
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>
OpenSSL's bbio logic is kind of crazy. It would be good to eventually do the
buffering in a better way (notably, bbio is fragile, if not outright broken,
for DTLS). In the meantime, this fixes a number of bugs where the existence of
bbio was leaked in the public API and broke things.
- SSL_get_wbio returned the bbio during the handshake. It must always return
the BIO the consumer configured. In doing so, internal accesses of
SSL_get_wbio should be switched to ssl->wbio since those want to see bbio.
For consistency, do the same with rbio.
- The logic in SSL_set_rfd, etc. (which I doubt is quite right since
SSL_set_bio's lifetime is unclear) would get confused once wbio got wrapped.
Those want to compare to SSL_get_wbio.
- If SSL_set_bio was called mid-handshake, bbio would get disconnected and lose
state. It forgets to reattach the bbio afterwards. Unfortunately, Conscrypt
does this a lot. It just never ended up calling it at a point where the bbio
would cause problems.
- Make more explicit the invariant that any bbio's which exist are always
attached. Simplify a few things as part of that.
Change-Id: Ia02d6bdfb9aeb1e3021a8f82dcbd0629f5c7fb8d
Reviewed-on: https://boringssl-review.googlesource.com/8023
Reviewed-by: Kenny Root <kroot@google.com>
Reviewed-by: Adam Langley <agl@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>
Rather than this confusing coordination with the handshake state machine and
init_num changing meaning partway through, use the length field already in
BUF_MEM. Like the new record layer parsing, is no need to keep track of whether
we are reading the header or the body. Simply keep extending the handshake
message until it's far enough along.
ssl3_get_message still needs tons of work, but this allows us to disentangle it
from the handshake state.
Change-Id: Ic2b3e7cfe6152a7e28a04980317d3c7c396d9b08
Reviewed-on: https://boringssl-review.googlesource.com/7948
Reviewed-by: Adam Langley <agl@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>
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>
That we're ignoring the return value is clearly wrong when
dtls1_retransmit_message has other code that doesn't ignore it, by way of
dtls1_do_handshake_write.
Change-Id: Ie3f8c0defdf1f5e709d67af4ca6fa4f0d83c76c9
Reviewed-on: https://boringssl-review.googlesource.com/7872
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>
We reset it to SSL_NOTHING at the start of ever SSL_get_error-using operation.
Then we only set it to a non-NOTHING value in the rest of the stack on error
paths.
Currently, ssl->rwstate is set all over the place. Sometimes the pattern is:
ssl->rwstate = SSL_WRITING;
if (BIO_write(...) <= 0) {
goto err;
}
ssl->rwstate = SSL_NOTHING;
Sometimes we only set it to the non-NOTHING value on error.
if (BIO_write(...) <= 0) {
ssl->rwstate = SSL_WRITING;
}
ssl->rwstate = SSL_NOTHING;
Sometimes we just set it to SSL_NOTHING far from any callback in random places.
The third case is arbitrary and clearly should be removed.
But, in the second case, we sometimes forget to undo it afterwards. This is
largely harmless since an error in the error queue overrides rwstate, but we
don't always put something in the error queue (falling back to
SSL_ERROR_SYSCALL for "I'm not sure why it failed. Perhaps it was one of your
callbacks? Check your errno equivalent."), but in that case a stray rwstate
value will cause it to be wrong.
We could fix the cases where we fail to set SSL_NOTHING on success cases, but
this doesn't account for there being multiple SSL_get_error operations. The
consumer may have an SSL_read and an SSL_write running concurrently. Instead,
it seems the best option is to lift the SSL_NOTHING reset to the operations and
set SSL_WRITING and friends as in the second case.
(Someday hopefully we can fix this to just be an enum that is internally
returned. It can convert to something stateful at the API layer.)
Change-Id: I54665ec066a64eb0e48a06e2fcd0d2681a42df7f
Reviewed-on: https://boringssl-review.googlesource.com/7453
Reviewed-by: David Benjamin <davidben@google.com>
A lot of consumers of obj.h only want the NID values. Others didn't need
it at all. This also removes some OBJ_nid2sn and OBJ_nid2ln calls in EVP
error paths which isn't worth pulling a large table in for.
BUG=chromium:499653
Change-Id: Id6dff578f993012e35b740a13b8e4f9c2edc0744
Reviewed-on: https://boringssl-review.googlesource.com/7563
Reviewed-by: David Benjamin <davidben@google.com>
Change-Id: I6267c9bfb66940d0b6fe5368514210a058ebd3cc
Reviewed-on: https://boringssl-review.googlesource.com/7494
Reviewed-by: Emily Stark (Dunn) <estark@google.com>
Reviewed-by: David Benjamin <davidben@google.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 fixes a number of bugs with the original logic:
- If handshake messages are fragmented and writes need to be retried, frag_off
gets completely confused.
- The BIO_flush call didn't set rwstate, so it wasn't resumable at that point.
- The msg_callback call gets garbage because the fragment header would get
scribbled over the handshake buffer.
The original logic was also extremely confusing with how it handles init_off.
(init_off gets rewound to make room for the fragment header. Depending on
where you pause, resuming may or may not have already been rewound.)
For simplicity, just allocate a new buffer to assemble the fragment in and
avoid clobbering the old one. I don't think it's worth the complexity to
optimize that. If we want to optimize this sort of thing, not clobbering seems
better anyway because the message may need to be retransmitted. We could avoid
doing a copy when buffering the outgoing message for retransmission later.
We do still need to track how far we are in sending the current message via
init_off, so I haven't opted to disconnect this function from
init_{buf,off,num} yet.
Test the fix to the retry + fragment case by having the splitHandshake option
to the state machine tests, in DTLS, also clamp the MTU to force handshake
fragmentation.
Change-Id: I66f634d6c752ea63649db8ed2f898f9cc2b13908
Reviewed-on: https://boringssl-review.googlesource.com/6421
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>
ssl.h should be first. Also two lines after includes and the rest of the
file.
Change-Id: Icb7586e00a3e64170082c96cf3f8bfbb2b7e1611
Reviewed-on: https://boringssl-review.googlesource.com/5892
Reviewed-by: Adam Langley <agl@google.com>
clang scan-build found a memory leak if the overflow codepath in
dtls1_hm_fragment is hit. Along the way, tidy up that function.
Change-Id: I3c4b88916ee56ab3ab63f93d4a967ceae381d187
Reviewed-on: https://boringssl-review.googlesource.com/5870
Reviewed-by: Adam Langley <agl@google.com>
One tedious thing about using CBB is that you can't safely CBB_cleanup
until CBB_init is successful, which breaks the general 'goto err' style
of cleanup. This makes it possible:
CBB_zero ~ EVP_MD_CTX_init
CBB_init ~ EVP_DigestInit
CBB_cleanup ~ EVP_MD_CTX_cleanup
Change-Id: I085ecc4405715368886dc4de02285a47e7fc4c52
Reviewed-on: https://boringssl-review.googlesource.com/5267
Reviewed-by: Adam Langley <agl@google.com>
See also upstream's 9dcab127e14467733523ff7626da8906e67eedd6. The root problem
is dtls1_read_bytes is wrong, but we can get the right behavior now and add a
regression test for it before cleaning it up.
Change-Id: I4e5c39ab254a872d9f64242c9b77b020bdded6e6
Reviewed-on: https://boringssl-review.googlesource.com/5123
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>
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>
Also size them based on the limits in the quantities they control (after
checking bounds at the API boundary).
BUG=404754
Change-Id: Id56ba45465a473a1a793244904310ef747f29b63
Reviewed-on: https://boringssl-review.googlesource.com/4559
Reviewed-by: Adam Langley <agl@google.com>
This is consistent with C's free function and upstream's convention.
Change-Id: I83f6e2f5824e28f69a9916e580dc2d8cb3b94234
Reviewed-on: https://boringssl-review.googlesource.com/4512
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 redundant with the lower level ones in s3_pkt.c just before BIO_read.
Only the operation which actually failed an operation on the BIO should set
the wait state.
Not all failure paths in ssl3_read_bytes and dtls1_read_bytes set SSL_READING,
but those that don't leave the BIO in a retry state, so SSL_READING doesn't
matter.
Change-Id: I2ae064ecc8b2946cc8ae8f724be09dfe49e077b5
Reviewed-on: https://boringssl-review.googlesource.com/4230
Reviewed-by: Adam Langley <agl@google.com>
signed/unsigned comparison. Just add a cast for now as in s3_both.c. Later
we'll properly size_t it alongside other tightening of this interface.
Change-Id: Idc8441d65e8ca65e39ab7172a8ec87d9ad710ed6
Reviewed-on: https://boringssl-review.googlesource.com/3860
Reviewed-by: Adam Langley <agl@google.com>
Turn them into static functions that take in an hm_fragment. It's not
immediately obvious that the frag_off/frag_len bounds checks and the msg_len
consistency check are critical to avoiding an out-of-bounds write. Better to
have dtls1_hm_fragment_mark also check internally.
Also rework the bitmask logic to be clearer and avoid a table.
Change-Id: Ica54e98f66295efb323e033cb6c67ab21e7d6cbc
Reviewed-on: https://boringssl-review.googlesource.com/3765
Reviewed-by: Adam Langley <agl@google.com>
