We shouldn't have protocol constraints that are sensitive to whether
data is returned synchronously or not.
Per https://boringssl-review.googlesource.com/#/c/4112/, the original
limitation was to avoid OpenSSL ABI changes. This is no longer a
concern.
Add tests for the sync and async case. Send the empty records in two
batches to ensure the count is reset correctly.
Change-Id: I3fee839438527e71adb83d437879bb0d49ca5c07
Reviewed-on: https://boringssl-review.googlesource.com/5040
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>
9a41d1b946 broke handling of multiple records in
a single packet. If |extend| is true, not all of the previous packet should be
consumed, only up to the record length.
Add a test which stresses the DTLS stack's handling of multiple handshake
fragments in a handshake record and multiple handshake records in a packet.
Change-Id: I96571098ad9001e96440501c4730325227b155b8
Reviewed-on: https://boringssl-review.googlesource.com/4950
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>
Now that WebRTC honors packet boundaries (https://crbug.com/447431), we
can start enforcing them correctly. Configuring read-ahead now does
nothing. Instead DTLS will always set "read-ahead" and also correctly
enforce packet boundaries when reading records. Add tests to ensure that
badly fragmented packets are ignored. Because such packets don't fail
the handshake, the tests work by injecting an alert in the front of the
handshake stream and ensuring the DTLS implementation ignores them.
ssl3_read_n can be be considerably unraveled now, but leave that for
future cleanup. For now, make it correct.
BUG=468889
Change-Id: I800cfabe06615af31c2ccece436ca52aed9fe899
Reviewed-on: https://boringssl-review.googlesource.com/4820
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>
This cuts down on one config knob as well as one case in the renego
combinatorial explosion. Since the only case we care about with renego
is the client auth hack, there's no reason to ever do resumption.
Especially since, no matter what's in the session cache:
- OpenSSL will only ever offer the session it just established,
whether or not a newer one with client auth was since established.
- Chrome will never cache sessions created on a renegotiation, so
such a session would never make it to the session cache.
- The new_session + SSL_OP_NO_SESSION_RESUMPTION_ON_RENEGOTIATION
logic had a bug where it would unconditionally never offer tickets
(but would advertise support) on renego, so any server doing renego
resumption against an OpenSSL-derived client must not support
session tickets.
This also gets rid of s->new_session which is now pointless.
BUG=429450
Change-Id: I884bdcdc80bff45935b2c429b4bbc9c16b2288f8
Reviewed-on: https://boringssl-review.googlesource.com/4732
Reviewed-by: Adam Langley <agl@google.com>
As of crbug.com/484543, Chromium's SSLClientSocket is not sensitive to whether
renegotiation is enabled or not. Disable it by default and require consumers to
opt into enabling this protocol mistake.
BUG=429450
Change-Id: I2329068284dbb851da010ff1fd398df3d663bcc3
Reviewed-on: https://boringssl-review.googlesource.com/4723
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>
They date to https://rt.openssl.org/Ticket/Display.html?id=2533, but no
particularly good justification was given for them. It seems it was just a
bandaid because d1_pkt.c forgot to initialize the buffer. I went through
codesearch for all accesses to SSL3_BUFFER::buf and SSL::packet. They seem
appropriately guarded but for this one.
Change-Id: Ife4e7afdb7a7c137d6be4791542eb5de6dd5b1b6
Reviewed-on: https://boringssl-review.googlesource.com/4685
Reviewed-by: Adam Langley <agl@google.com>
There's no real need to ever disable it, so this is one fewer configuration to
test. It's still disabled for DTLS, but a follow-up will resolve that.
Change-Id: Ia95ad8c17ae8236ada516b3968a81c684bf37fd9
Reviewed-on: https://boringssl-review.googlesource.com/4683
Reviewed-by: Adam Langley <agl@google.com>
This causes any unexpected handshake records to be met with a fatal
no_renegotiation alert.
In addition, restore the redundant version sanity-checks in the handshake state
machines. Some code would zero the version field as a hacky way to break the
handshake on renego. Those will be removed when switching to this API.
The spec allows for a non-fatal no_renegotiation alert, but ssl3_read_bytes
makes it difficult to find the end of a ClientHello and skip it entirely. Given
that OpenSSL goes out of its way to map non-fatal no_renegotiation alerts to
fatal ones, this seems probably fine. This avoids needing to account for
another source of the library consuming an unbounded number of bytes without
returning data up.
Change-Id: Ie5050d9c9350c29cfe32d03a3c991bdc1da9e0e4
Reviewed-on: https://boringssl-review.googlesource.com/4300
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>
MSVC doesn't like unary - on unsigned numbers. Also switch ssl3_read_n's
version to uintptr_t to match the write half. This gets us closer to clearing
through C4311 violations. (The remaining one is in asn1_add_error which can go
after verifying that most of asn1_mac.h is safe to drop.)
Change-Id: Idb33dda8863bf1a3408b14d5513a667338311b6b
Reviewed-on: https://boringssl-review.googlesource.com/4255
Reviewed-by: Adam Langley <agl@google.com>
At this point, has_version has been set and we may even have a non-null cipher.
Trying to assign meaning to the record-layer version number is not worth making
s->version's semantics even more complicated.
Change-Id: Ia1cf341cf7306eb48d2d11241316dc2116306968
Reviewed-on: https://boringssl-review.googlesource.com/4237
Reviewed-by: Adam Langley <agl@google.com>
Compression is gone, so don't allow for compression overhead. With that fixed,
the second rr->length check in ssl3_get_record matches the length computation
which sizes the read buffer. The first is wrong and doesn't account for the
alignment padding. Move the second to the first.
Change-Id: I3f4f05de9fdf5c645ff24493bbfdf303dcc1aa90
Reviewed-on: https://boringssl-review.googlesource.com/4236
Reviewed-by: Adam Langley <agl@google.com>
Separate actually writing the fragment to the network from assembling it so
there is no need for is_fragment. record_split_done also needn't be a global;
as of 7fdeaf11017b82368e0a97547fc491b90ad40f67, it is always reset to 0 whether
or not SSL3_WANT_WRITE occurred, despite the comment.
I believe this is sound, but the pre-7fdeaf1 logic wasn't quiiite right;
ssl3_write_pending allows a retry to supply *additional* data, so not all
plaintext had been commited to before the IV was randomized. We could fix this
by tracking how many bytes were committed to the last time we fragmented, but
this is purely an optimization and doesn't seem worth the complexity.
This also fixes the alignment computation in the record-splitting case. The
extra byte was wrong, as demonstrated by the assert.
Change-Id: Ia087a45a6622f4faad32e501942cc910eca1237b
Reviewed-on: https://boringssl-review.googlesource.com/4234
Reviewed-by: Adam Langley <agl@google.com>
It's still rather a mess, but this is at least somewhat clearer. The old one
had a lot of remnants of compression, etc.
Change-Id: Iffcb4dd4e8c4ab14f60abf917d22b7af960c93ba
Reviewed-on: https://boringssl-review.googlesource.com/4233
Reviewed-by: Adam Langley <agl@google.com>
This mostly[*] doesn't matter for TLS since the message would have been
rejected anyway, but, in DTLS, if the peer rejects our Finished, it will send
an encrypted alert. This will then cause it to hang, which isn't very helpful.
I've made the change on both TLS and DTLS so the two protocols don't diverge on
this point. It is true that we're accepting nominally encrypted and
authenticated alerts before Finished, but, prior to ChangeCipherSpec, the
alerts are sent in the clear anyway so an attacker could already inject alerts.
A consumer could only be sensitive to it being post-CCS if it was watching
msg_callback. The only non-debug consumer of msg_callback I've found anywhere
is some hostapd code to detect Heartbeat.
See https://code.google.com/p/webrtc/issues/detail?id=4403 for an instance
where the equivalent behavior in OpenSSL masks an alert.
[*] This does change behavior slightly if the peer sends a warning alert
between CCS and Finished. I believe this is benign as warning alerts are
usually ignored apart from info_callback and msg_callback. The one exception is
a close_notify which is a slightly new state (accepting close_notify during a
handshake seems questionable...), but they're processed pre-CCS too.
Change-Id: Idd0d49b9f9aa9d35374a9f5e2f815cdb931f5254
Reviewed-on: https://boringssl-review.googlesource.com/3883
Reviewed-by: Adam Langley <agl@google.com>
The assert was supposed to be *added* in fcf25833 but instead replaced
the check.
BUG=465557
Change-Id: I0d3db5038515021e5bdd1ccb9ff08d4f78552621
Reviewed-on: https://boringssl-review.googlesource.com/3850
Reviewed-by: David Benjamin <davidben@chromium.org>
Reviewed-by: Adam Langley <agl@google.com>
Align the DTLS and TLS implementations more. s3_pkt.c's version still has
remnants of fragmentable alerts and only one side marks some variables as
const. Also use warning/fatal constants rather than the numbers with comments.
Change-Id: Ie62d3af1747b6fe4336496c047dfccc9d71fde3f
Reviewed-on: https://boringssl-review.googlesource.com/3562
Reviewed-by: Adam Langley <agl@google.com>
False Start is the name it's known by now. Deprecate the old API and expose new
ones with the new name.
Change-Id: I32d307027e178fd7d9c0069686cc046f75fdbf6f
Reviewed-on: https://boringssl-review.googlesource.com/3481
Reviewed-by: Adam Langley <agl@google.com>
This makes the following changes:
- SSL_cutthrough_complete no longer rederives whether cutthrough happened and
just maintains a handshake bit.
- SSL_in_init no longer returns true if we are False Starting but haven't
completed the handshake. That logic was awkward as it depended on querying
in_read_app_data to force SSL_read to flush the entire handshake. Defaulting
SSL_in_init to continue querying the full handshake and special-casing
SSL_write is better. E.g. the check in bidirectional SSL_shutdown wants to know
if we're in a handshake. No internal consumer of
SSL_MODE_HANDSHAKE_CUTTHROUGH ever queries SSL_in_init directly.
- in_read_app_data is gone now that the final use is dead.
Change-Id: I05211a116d684054dfef53075cd277b1b30623b5
Reviewed-on: https://boringssl-review.googlesource.com/3336
Reviewed-by: Adam Langley <agl@google.com>
It may take up to two iterations of s->handshake_func before it is safe to
continue. Fortunately, even if anything was using False Start this way
(Chromium doesn't), we don't inherit NSS's security bug. The "redundant" check
in the type match case later on in this function saves us.
Amusingly, the success case still worked before this fix. Even though we fall
through to the post-handshake codepath and get a handshake record while
"expecting" app data, the handshake state machine is still pumped thanks to a
codepath meant for renego!
Change-Id: Ie129d83ac1451ad4947c4f86380879db8a3fd924
Reviewed-on: https://boringssl-review.googlesource.com/3335
Reviewed-by: Adam Langley <agl@google.com>
The new V2ClientHello sniff asserts, for safety, that nothing else has
initialized the record layer before it runs. However, OpenSSL allows you to
avoid explicitly calling SSL_connect/SSL_accept and instead let
SSL_read/SSL_write implicitly handshake for you. This check happens at a fairly
low-level in the ssl3_read_bytes function, at which point the record layer has
already been initialized.
Add some tests to ensure this mode works.
(Later we'll lift the handshake check to a higher-level which is probably
simpler.)
Change-Id: Ibeb7fb78e5eb75af5411ba15799248d94f12820b
Reviewed-on: https://boringssl-review.googlesource.com/3334
Reviewed-by: Adam Langley <agl@google.com>
This is the source of much of renegotiation's complexity, and of OpenSSL's
implementation of it. In practice, we only care about renegotiation because of
the client auth hack. There, we can safely assume that no server will send
application data between sending the HelloRequest and completing the handshake.
BUG=429450
Change-Id: I37f5abea5fdedb1d53e24ceb11f71287c74bb777
Reviewed-on: https://boringssl-review.googlesource.com/3332
Reviewed-by: Adam Langley <agl@google.com>
The check on the DTLS side was broken anyway. On the TLS side, the spec does
say to ignore them, but there should be no need for this in future-proofing and
NSS doesn't appear to be lenient here. See also
https://boringssl-review.googlesource.com/#/c/3233/
Change-Id: I0846222936c5e08acdcfd9d6f854a99df767e468
Reviewed-on: https://boringssl-review.googlesource.com/3290
Reviewed-by: Adam Langley <agl@google.com>
The distinction between publicly and non-publicly invalid is barely acted upon
and slightly silly now that the CBC padding check has been folded into
EVP_AEAD.
Change-Id: Idce4b9b8d29d624e3c95243a147265d071612127
Reviewed-on: https://boringssl-review.googlesource.com/2980
Reviewed-by: Adam Langley <agl@google.com>
RAND_pseudo_bytes just calls RAND_bytes now and only returns 0 or 1. Switch all
callers within the library call the new one and use the simpler failure check.
This fixes a few error checks that no longer work (< 0) and some missing ones.
Change-Id: Id51c79deec80075949f73fa1fbd7b76aac5570c6
Reviewed-on: https://boringssl-review.googlesource.com/2621
Reviewed-by: Adam Langley <agl@google.com>
Match the DTLS code. Rather than sniffing the handshake state, use the
have_version bit.
Change-Id: I40e92f187647417c34b4cfdc3ad258f5562e781b
Reviewed-on: https://boringssl-review.googlesource.com/2588
Reviewed-by: Adam Langley <agl@google.com>
SSL3_ENC_METHOD will remain version-specific while SSL_METHOD will become
protocol-specific. This finally removes all the version-specific portions of
SSL_METHOD but the version tag itself.
(SSL3_ENC_METHOD's version-specific bits themselves can probably be handled by
tracking a canonicalized protocol version. It would simplify version
comparisons anyway. The one catch is SSLv3 has a very different table. But
that's a cleanup for future. Then again, perhaps a version-specific method
table swap somewhere will be useful later for TLS 1.3.)
Much of this commit was generated with sed invocation:
s/method->ssl3_enc/enc_method/g
Change-Id: I2b192507876aadd4f9310240687e562e56e6c0b1
Reviewed-on: https://boringssl-review.googlesource.com/2581
Reviewed-by: Adam Langley <agl@google.com>
Bruce Dawson pointed out that the shadowing of |ret| in |s3_srvr.c|
looked dodgy. It was actually deliberate (we don't want to reset the
default value of the function's |ret| variable with a successful return
from the callback) but it does look dodgy.
This change adds -Wshadow to ban variable shadowing and fixes all
current instances.
Change-Id: I1268f88b9f26245c7d16d6ead5bb9014ea471c01
Reviewed-on: https://boringssl-review.googlesource.com/2520
Reviewed-by: Adam Langley <agl@google.com>
PR#1767
(Imported from upstream's fe78f08d15)
Not sure this is reachable (upstream's PR references custom engines), but
better be tidy. Note this is slightly different from upstream's: EVP_Cipher is
documented to return -1 on failure, not 0.
Change-Id: I836f12b73c6912a8ae8cbd37cfd3d33466acbc9e
Reviewed-on: https://boringssl-review.googlesource.com/2478
Reviewed-by: Adam Langley <agl@google.com>
first_packet is a temporary connection-global flag set for the duration of some
call and then queried from other code. This kind of logic is too difficult to
reason through. It also incorrectly treats renegotiate ClientHellos as
pre-version-negotiation records. This eliminates the need to query
enc_write_ctx (which wasn't EVP_AEAD-aware anyway).
Instead, take a leaf from Go TLS's book and add a have_version bit. This is
placed on s->s3 as it is connection state; s->s3 automatically gets reset on
SSL_clear while s doesn't.
This new flag will also be used to determine whether to do the V2ClientHello
sniff when the version-locked methods merge into SSLv23_method. It will also
replace needing to condition s->method against a dummy DTLS_ANY_VERSION value
to determine whether DTLS version negotiation has happened yet.
Change-Id: I5c8bc6258b182ba4ab175a48a84eab6d3a001333
Reviewed-on: https://boringssl-review.googlesource.com/2442
Reviewed-by: Adam Langley <agl@google.com>
It's unused. Also per the previous commit message, it historically had a bug
anyway.
Change-Id: I5868641e7938ddebbc0ffd72d218c81cd17c7739
Reviewed-on: https://boringssl-review.googlesource.com/2437
Reviewed-by: Adam Langley <agl@google.com>
Prior to this change, BoringSSL maintained a 2-byte buffer for alerts,
and would support reassembly of fragmented alerts.
NSS does not support fragmented alerts, nor would any reasonable
implementation produce them. Remove fragmented alert handling and
produce an error if a fragmented alert has ever been encountered.
Change-Id: I31530ac372e8a90b47cf89404630c1c207cfb048
Reviewed-on: https://boringssl-review.googlesource.com/2125
Reviewed-by: Adam Langley <agl@google.com>
This was added in upstream's 82e610e2cf. The
commit message cites draft-ietf-tls-renegotiation which was on
draft-ietf-tls-renegotiation-01 at the time. The text in question (6.2 Server
Considerations) is no longer in RFC 5746. The RFC now recommends terminating
the connection which is much simpler.
It also was wrong anyway as it checked s->ctx->options instead of s->options
for SSL_OP_ALLOW_UNSAFE_LEGACY_RENEGOTIATION.
Removing that block will result in the connection being terminated in
ssl_scan_clienthello_tlsext.
Change-Id: Ie222c78babd3654c5023ad07ac0d8e0adde68698
Reviewed-on: https://boringssl-review.googlesource.com/2235
Reviewed-by: Adam Langley <agl@google.com>
Without SSL_MODE_AUTO_RETRY, even blocking mode will return
SSL_ERROR_WANT_{READ|WRITE} in the event of a renegotiation.
The comments in the code speak only of "nasty problems" unless this is
done. The original commit that added SSL_MODE_AUTO_RETRY
(54f10e6adc) gives a little more detail:
The [...] behaviour is needed by applications such as s_client and
s_server that use select() to determine when to use SSL_read.
Without the -nbio flag, s_client will use select() to find when the
socket is readable and then call SSL_read with a blocking socket.
However, this will still block in the event of an incomplete record, so
the delay is already unbounded. This it's very unclear what the point of
this behaviour ever was.
Perhaps if the read and write paths were different sockets where the
read socket was non-blocking but the write socket was blocking. But that
seems like an implausible situation to worry too much about.
Change-Id: I9d9f2526afc2e0fd0e5440e9a047f419a2d61afa
Reviewed-on: https://boringssl-review.googlesource.com/2140
Reviewed-by: Adam Langley <agl@google.com>
This resolves a pile of MSVC warnings in Chromium.
Change-Id: Ib9a29cb88d8ed8ec4118d153260f775be059a803
Reviewed-on: https://boringssl-review.googlesource.com/1865
Reviewed-by: Adam Langley <agl@google.com>
David Benjamin
Adam Langley
Alex Chernyakhovsky