absl::Span, base::span, and std::span have first() and last() methods
which give prefixes and suffixes. first() just saves 5 characters, but
last() is nicer to write than subspan() for suffixes.
Unlike subspan(), they also do not have clipping behavior, so we're
guaranteed the length is correct. The clipping behavior comes from
absl::Span::subspan() and is not present in std::span or base::span.
I've left it in, in case we switch to absl::Span in the future, but I
imagine absl::Span will need to migrate this at some point.
Change-Id: I042dd6c566b6d753ec6de9d84e8c09ac7c270267
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/48905
Reviewed-by: Adam Langley <agl@google.com>
If a client offers ECH, but the server rejects it, the client completes
the handshake with ClientHelloOuter in order to authenticate retry keys.
Implement this flow. This is largely allowing the existing handshake to
proceed, but with some changes:
- Certificate verification uses the other name. This CL routes this up to
the built-in verifier and adds SSL_get0_ech_name_override for the
callback.
- We need to disable False Start to pick up server Finished in TLS 1.2.
- Client certificates, notably in TLS 1.3 where they're encrypted,
should only be revealed to the true server. Fortunately, not sending
client certs is always an option, so do that.
Channel ID has a similar issue. I've just omitted the extension in
ClientHelloOuter because it's deprecated and is unlikely to be used
with ECH at this point. ALPS may be worth some pondering but, the way
it's currently used, is not sensitive.
(Possibly we should change the draft to terminate the handshake before
even sending that flight...)
- The session is never offered in ClientHelloOuter, but our internal
book-keeping doesn't quite notice.
I had to replace ech_accept with a tri-state ech_status to correctly
handle an edge case in SSL_get0_ech_name_override: when ECH + 0-RTT +
reverify_on_resume are all enabled, the first certificate verification
is for the 0-RTT session and should be against the true name, yet we
have selected_ech_config && !ech_accept. A tri-state tracks when ECH is
actually rejected. I've maintained this on the server as well, though
the server never actually cares.
Bug: 275
Change-Id: Ie55966ca3dc4ffcc8c381479f0fe9bcacd34d0f8
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/48135
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
ssl_update_cache takes the cache lock to add to the session cache,
releases it, and then immediately takes and releases the lock to
increment handshakes_since_cache_flush. Then, in 1/255 connections, does
the same thing again to flush stale sessions.
Merge the first two into one lock. In doing so, move ssl_update_cache to
ssl_session.cc, so it can access a newly-extracted add_session_lock.
Also remove the mode parameter (the SSL knows if it's a client or
server), and move the established_session != session check to the
caller, which more directly knows whether there was a new session.
Also add some TSan coverage for this path in the tests. In an earlier
iteration of this patch, I managed to introduce a double-locking bug
because we weren't testing it at all. Confirmed this test catches both
double-locking and insufficient locking. (It doesn't seem able to catch
using a read lock instead of a write lock in SSL_CTX_flush_sessions,
however. I suspect the hash table is distributing the cells each thread
touches.)
Update-Note: This reshuffles some locks around the session cache.
(Hopefully for the better.)
Change-Id: I78dca53fda74e036b90110cca7fbcc306a5c8ebe
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/48133
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
In renegotiation handshakes and, later, ECH ClientHelloOuter handshakes,
we don't want to add sessions to the session cache. We also don't want
to release a session as resumable until the handshake completes.
Ideally we'd only construct SSL_SESSION at the end of the handshake, but
existing APIs like SSL_get_session must work mid-handshake, so
SSL_SESSION is both a handle to immutable resumption state, and a
container for in-progress connection properties. We manage this with a
not_resumable flag that's only cleared after the handshake is done and
the SSL_SESSION finalized.
However, TLS 1.2 ticket renewal currently clears the flag too early and
breaks the invariant. This won't actually affect renegotiation or
ClientHelloOuter because those handshakes never resume. Still, we can
maintain the invariant storing the copy in hs->new_session. Note this
does sacrifice a different invariant: previously, ssl->session and
hs->new_session were never set at the same time.
This change also means ssl_update_cache does not need to special-case
ticket renewal.
Change-Id: I03230cd9c63e5bee6bd60cd05c0439e16533c6d4
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/48132
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Based on an initial implementation by Dan McArdle at
https://boringssl-review.googlesource.com/c/boringssl/+/46784
This CL contains most of a client implementation for
draft-ietf-tls-esni-10. The pieces missing so far, which will be done in
follow-up CLs are:
1. While the ClientHelloInner is padded, the server Certificate message
is not. I'll add that once we resolve the spec discussions on how to
do that. (We were originally going to use TLS record-level padding,
but that doesn't work well with QUIC.)
2. The client should check the public name is a valid DNS name before
copying it into ClientHelloOuter.server_name.
3. The ClientHelloOuter handshake flow is not yet implemented. This CL
can detect when the server selects ClientHelloOuter, but for now the
handshake immediately fails. A follow-up CL will remove that logic
and instead add the APIs and extra checks needed.
Otherwise, this should be complete, including padding and compression.
The main interesting point design-wise is that we run through
ClientHello construction multiple times. We need to construct
ClientHelloInner and ClientHelloOuter. Then each of those has slight
variants: EncodedClientHelloInner is the compressed form, and
ClientHelloOuterAAD just has the ECH extension erased to avoid a
circular dependency.
I've computed ClientHelloInner and EncodedClientHelloInner concurrently
because the compression scheme requires shifting the extensions around
to be contiguous. However, I've computed ClientHelloOuterAAD and
ClientHelloOuter by running through the logic twice. This probably can
be done better, but the next draft revises the construction anyway, so
I'm thinking I'll rework it then. (In the next draft, we use a
placeholder payload of the same length, so we can construct the
ClientHello once and fill in the payload.)
Additionally, now that we have a client available in ssl_test, this adds
a threading test to confirm that SSL_CTX_set1_ech_keys is properly
synchronized. (Confirmed that, if I drop the lock in
SSL_CTX_set1_ech_keys, TSan notices.)
Change-Id: Icaff68b595035bdcc73c468ff638e67c84239ef4
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/48004
Reviewed-by: Adam Langley <agl@google.com>
The first thing any deployment will want to monitor is whether ECH was
actually used. Also it's useful if the command-line tool can output
this. (The alert is how the client signals it discarded the connection
due to ECH reject.)
This also disables ECH with the handoff mechanism for now. (The
immediate cause being that ech_accept isn't serialized.) We'll probably
need to make some decisions around the ordering here, since ECH affects
where the true ClientHello is available.
Bug: 275
Change-Id: Ie4559733290e653a514fcd94431090bf86bc3172
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47911
Reviewed-by: Adam Langley <agl@google.com>
The channel_id_valid bit is both used for whether channel_id is filled
in (SSL_get_tls_channel_id), and whether this particular handshake will
eventually negotiate Channel ID.
The former means that, if SSL_get_tls_channel_id is called on the
client, we'll return all zeros. Apparently we never fill in channel_id
on the client at all. The latter means the state needs to be reset on
renegotiation because we do not currently forbid renegotiation with
Channel ID (we probably should...), which is the last use of the init
callback for extensions.
Instead, split this into a bit for the handshake and a bit for the
connection. Note this means we actually do not expose or even retain
whether Channel ID was used on the client.
This requires a tweak to the handoff logic, but it should be compatible.
The serialized ssl->s3->channel_id was always a no-op: the handback
happens before the ChannelID message, except in RSA key exchange. But we
forbid Channel ID in RSA key exchange anyway.
Update-Note: SSL_get_tls_channel_id will no longer return all zeros
during the handshake or on the client. I did not find any callers
relying on this.
Change-Id: Icd4b78dd3f311d1c7dfc1cae7d2b86dc7e327a99
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47906
Reviewed-by: Adam Langley <agl@google.com>
When decrypting a ticket we would copy the client's session ID into the
session and then copy the session's ID into the ServerHello (if
resuming). That seems icky. Instead install the same placeholder on the
server as we do on the client.
Change-Id: Icb50a3be2f05e6428f1b286c8c09015f7bb4af16
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47784
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
It's kind of weird that we assign a session ID, based on whether we
detect the handshake wants stateful resumption, and then erase it
afterwards.
Also remove the is_server parameter, which we can get from hs.
Change-Id: I94ac817c63abb08a457e0e0c29f5c2d2b60aa498
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47444
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This introduces an EVP_HPKE_KEM, to capture the KEM choice, and
EVP_HPKE_KEY, to capture the key import (and thus avoids asking
receivers to pass in the full keypair). It is a bit more wordy now, but
we'll be in a better place when some non-TLS user inevitably asks for a
P-256 version.
Bug: 410
Change-Id: Icb9cc8b028e6d1f86e6d8adb31ebf1f975181675
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47329
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Callers using private key callbacks may retain non-trivial state with a
private key. In many cases, the private key is no longer necessary
immediately after the first round-trip (e.g. non-HRR TLS 1.3
connections). Add a function that callers can query to drop the state a
hair earlier.
This is tested in two ways. First, the asserts in front of using the
key, combined with existing tests, ensure we don't start reporting it
too early. Second, I've added tests in ssl_test.cc to assert we report
it as early as we expect to.
In doing so, the number of parameters on ConnectClientAndServer()
started getting tedious, so I've split that into a
CreateClientAndServer() and CompleteHandshakes(). Callers that need to
configure weird things or drive the handshake manually can call
CreateClientAndServer() (which takes care of the BIO pair business) and
continue from there.
Bug: b/183734559
Change-Id: I05e1edb6d269c8468ba7cde7dc90e0856694a0ca
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47344
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This replaces the ID-based API with one that is more static linker
friendly. For ECH, it doesn't make a difference because we currently
pull in all the options we've implemented. But this means other HPKE
uses need not pull in everything ECH needs and vice versa.
Along the way, fix an inconsistency: we prefixed all the AEAD constants
with "AEAD", but not the others. Since the rest of the name already
determines everything, go with the shorter version.
Bug: 410
Change-Id: I56e46c13b43c97e15eeb45204cde7019dd21e250
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47327
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Bug: 275
Change-Id: I8096070386af7d2b5020875ea09bcc0c04ebc8cd
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/47245
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Commit-Queue: Steven Valdez <svaldez@google.com>
See go/handshake-hints (internal).
CL originally by Bin Wu <wub@google.com>. I just reworked the tests and
tidied it up a bit. This is the start of a replacement for the split
handshakes API. For now, only TLS 1.3 is supported. It starts with an
initial set of hints, but we can add more later. (In particular, we
should probably apply the remote handshaker's extension order to avoid
needing to capability protect such changes.)
Change-Id: I7b6a6dfaa84c6c6e3436d2a4026c3652b8a79f0f
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/46535
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This CL adds an initial implementation of the ECH server, with pieces of
the client in BoGo as necessary for testing. In particular, the server
supports ClientHelloInner compression with ech_outer_extensions. When
ECH decryption fails, it can send retry_configs back to the client.
This server passes the "ech-accept" and "ech-reject" test cases in
tls-interop-runner[0] when tested against both the cloudflare-go and nss
clients. For reproducibility, I started with the main branch at commit
707604c262d8bcf3e944ed1d5a675077304732ce and updated the endpoint's
script to pass the server's ECHConfig and private key to the boringssl
tool.
Follow-up CLs will update HPKE to the latest draft and catch us up to
draft-10.
[0]: https://github.com/xvzcf/tls-interop-runner
Bug: 275
Change-Id: I49be35af46d1fd5dd9c62252f07d0bae179381ab
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/45285
Reviewed-by: David Benjamin <davidben@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
This CL implements the backend server behavior described in Section 7.2
of draft-ietf-tls-esni-09.
Bug: 275
Change-Id: I2e162673ce564db0cb75fc9b71ef11ed15037f4b
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/43924
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
It's not even accurate. The term "master key" dates to SSL 2, which we
do not implement. (Starting SSL 3, "key" was replaced with "secret".)
The field stores, at various points, the TLS 1.2 master secret, the TLS
1.3 resumption master secret, and the TLS 1.3 resumption PSK. Simply
rename the field to 'secret', which is as descriptive of a name as we
can get at this point.
I've left SSL_SESSION_get_master_key alone for now, as it's there for
OpenSSL compatibility, as well as references to the various TLS secrets
since those refer to concepts in the spec. (When the dust settles a bit
on rfc8446bis, we can fix those.)
Change-Id: I3c1007eb7982788789cc5db851de8724c7f35baf
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/44144
Reviewed-by: Adam Langley <agl@google.com>
This was introduced in OpenSSL 1.1.1, and wpa_supplicant expects us to
have it. We had this same function as SSL_CIPHER_get_value (to match
SSL_get_cipher_by_value). Align with upstream's name.
It seems we also had a ssl_cipher_get_value lying around, so fold them
together. (I've retained the assert in ssl_cipher_get_value as it seems
reasonable enough; casting a hypothetical SSLv2 cipher ID to uint16_t
would not behave correctly.)
Change-Id: Ifbec460435bbc483f2c3de988522e321f2708172
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/42966
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <agl@google.com>
David Benjamin
Adam Langley
Steven Valdez
Daniel McArdle