This involves more synchronization with child exits as the kernel no longer
closes the pre-created pipes for free, but it works on Windows. As long as
TCP_NODELAY is set, the performance seems comparable. Though it does involve
dealing with graceful socket shutdown. I couldn't get that to work on Windows
without draining the socket; not even SO_LINGER worked. Current (untested)
theory is that Windows refuses to gracefully shutdown a socket if the peer
sends data after we've stopped reading.
cmd.ExtraFiles doesn't work on Windows; it doesn't use fds natively, so you
can't pass fds 4 and 5. (stdin/stdout/stderr are special slots in
CreateProcess.) We can instead use the syscall module directly and mark handles
as inheritable (and then pass the numerical values out-of-band), but that
requires synchronizing all of our shim.Start() calls and assuming no other
thread is spawning a process.
PROC_THREAD_ATTRIBUTE_HANDLE_LIST fixes threading problems, but requires
wrapping more syscalls. exec.Cmd also doesn't let us launch the process
ourselves. Plus it still requires every handle in the list be marked
inheritable, so it doesn't help if some other thread is launching a process
with bInheritHandles TRUE but NOT using PROC_THREAD_ATTRIBUTE_HANDLE_LIST.
(Like Go, though we can take syscall.ForkLock there.)
http://blogs.msdn.com/b/oldnewthing/archive/2011/12/16/10248328.aspx
The more natively Windows option seems to be named pipes, but that too requires
wrapping more system calls. (To be fair, that isn't too painful.) They also
involve a listening server, so we'd still have to synchronize with shim.Wait()
a la net.TCPListener.
Then there's DuplicateHandle, but then we need an out-of-band signal.
All in all, one cross-platform implementation with a TCP sockets seems
simplest.
Change-Id: I38233e309a0fa6814baf61e806732138902347c0
Reviewed-on: https://boringssl-review.googlesource.com/3563
Reviewed-by: Adam Langley <agl@google.com>
Tighten up the requirements for False Start. At this point, neither
AES-CBC or RC4 are something that we want to use unless we're sure that
the server wants to speak them.
Rebase of original CL at: https://boringssl-review.googlesource.com/#/c/1980/
BUG=427721
Change-Id: I9ef7a596edeb8df1ed070aac67c315b94f3cc77f
Reviewed-on: https://boringssl-review.googlesource.com/3501
Reviewed-by: Adam Langley <agl@google.com>
Also adds a flag to runner.go to make it more suitable for printing to a pipe.
Change-Id: I26fae21f3e4910028f6b8bfc4821c8c595525504
Reviewed-on: https://boringssl-review.googlesource.com/3490
Reviewed-by: Adam Langley <agl@google.com>
I found no users of this. We can restore it if needbe, but I don't expect
anyone to find it useful in its current form. The API is suspect for the same
reasons DTLSv1_listen was. An SSL object is stateful and assumes you already
have the endpoint separated out.
If we ever need it, server-side HelloVerifyRequest and DTLSv1_listen should be
implemented by a separate stateless listener that statelessly handles
cookieless ClientHello + HelloVerifyRequest. Once a ClientHello with a valid
cookie comes in, it sets up a stateful SSL object and passes control along to
that.
Change-Id: I86adc1dfb6a81bebe987784c36ad6634a9a1b120
Reviewed-on: https://boringssl-review.googlesource.com/3480
Reviewed-by: Adam Langley <agl@google.com>
Found by running malloc tests with -valgrind. Unfortunately, the next one is
deep in crypto/asn1 itself, so I'm going to stop here for now.
Change-Id: I7a33971ee07c6b7b7a98715f2f18e0f29380c0a1
Reviewed-on: https://boringssl-review.googlesource.com/3350
Reviewed-by: Adam Langley <agl@google.com>
This is consistent with ignoring writeRecord failures. Without doing this, the
DTLS MinimumVersion test now flakily fails with:
FAILED (MinimumVersion-Client-TLS12-TLS1-DTLS)
bad error (wanted ':UNSUPPORTED_PROTOCOL:' / 'remote error: protocol version not supported'): local error 'write unix @: broken pipe', child error 'exit status 2', stdout:
2092242157:error:1007b1a7:SSL routines:ssl3_get_server_hello:UNSUPPORTED_PROTOCOL:../ssl/s3_clnt.c:783:
This is because the MinimumVersion tests assert on /both/ expectedError and
expectedLocalError. The latter is valuable as it asserts on the alert the peer
returned. (I would like us to add more such assertions to our tests where
appropriate.) However, after we send ServerHello, we also send a few messages
following it. This races with the peer shutdown and we sometimes get EPIPE
before reading the alert.
Change-Id: I3fe37940a6a531379673a00976035f8e76e0f825
Reviewed-on: https://boringssl-review.googlesource.com/3337
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 works fine, but I believe NSS had a bug here a couple years ago. Also move
all the Skip* bug options next to each other in order.
Change-Id: I72dcb3babeee7ba73b3d7dc5ebef2e2298e37438
Reviewed-on: https://boringssl-review.googlesource.com/3333
Reviewed-by: Adam Langley <agl@google.com>
We have test coverage for invalid alerts, but not for normal ones on the DTLS
side.
Change-Id: I359dce8d4dc80dfa99b5d8bacd73f48a8e4ac310
Reviewed-on: https://boringssl-review.googlesource.com/3291
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>
SSL_AEAD_CTX ownership is currently too confusing. Instead, rely on the lack of
renego, so the previous epoch always uses the NULL cipher. (Were we to support
DTLS renego, we could keep track of s->d1->last_aead_write_ctx like
s->d1->last_write_sequence, but it isn't worth it.)
Buffered messages also tracked an old s->session, but this is unnecessary. The
s->session NULL check in tls1_enc dates to the OpenSSL initial commit and is
redundant with the aead NULL check.
Change-Id: I9a510468d95934c65bca4979094551c7536980ae
Reviewed-on: https://boringssl-review.googlesource.com/3234
Reviewed-by: Adam Langley <agl@google.com>
Nothing recognized through those codepaths is fragmentable in DTLS. Also remove
an unnecessary epoch check. It's not possible to process a record from the
wrong epoch.
Change-Id: I9d0f592860bb096563e2bdcd2c8e50a0d2b65f59
Reviewed-on: https://boringssl-review.googlesource.com/3232
Reviewed-by: Adam Langley <agl@google.com>
For now, only test reorderings when we always or never fragment messages.
There's a third untested case: when full messages and fragments are mixed. That
will be tested later after making it actually work.
Change-Id: Ic4efb3f5e87b1319baf2d4af31eafa40f6a50fa6
Reviewed-on: https://boringssl-review.googlesource.com/3216
Reviewed-by: Adam Langley <agl@google.com>
No behavior change. This is in preparation for buffering a flight of handshake
messages to reorder vigorously on flush.
Change-Id: Ic348829b340bf58d28f332027646559cb11046ac
Reviewed-on: https://boringssl-review.googlesource.com/3215
Reviewed-by: Adam Langley <agl@google.com>
This extends the packet adaptor protocol to send three commands:
type command =
| Packet of []byte
| Timeout of time.Duration
| TimeoutAck
When the shim processes a Timeout in BIO_read, it sends TimeoutAck, fails the
BIO_read, returns out of the SSL stack, advances the clock, calls
DTLSv1_handle_timeout, and continues.
If the Go side sends Timeout right between sending handshake flight N and
reading flight N+1, the shim won't read the Timeout until it has sent flight
N+1 (it only processes packet commands in BIO_read), so the TimeoutAck comes
after N+1. Go then drops all packets before the TimeoutAck, thus dropping one
transmit of flight N+1 without having to actually process the packets to
determine the end of the flight. The shim then sees the updated clock, calls
DTLSv1_handle_timeout, and re-sends flight N+1 for Go to process for real.
When dropping packets, Go checks the epoch and increments sequence numbers so
that we can continue to be strict here. This requires tracking the initial
sequence number of the next epoch.
The final Finished message takes an additional special-case to test. DTLS
triggers retransmits on either a timeout or seeing a stale flight. OpenSSL only
implements the former which should be sufficient (and is necessary) EXCEPT for
the final Finished message. If the peer's final Finished message is lost, it
won't be waiting for a message from us, so it won't time out anything. That
retransmit must be triggered on stale message, so we retransmit the Finished
message in Go.
Change-Id: I3ffbdb1de525beb2ee831b304670a3387877634c
Reviewed-on: https://boringssl-review.googlesource.com/3212
Reviewed-by: Adam Langley <agl@google.com>
This reverts commit c67a3ae6ba. With a
deterministic clock, we can now go back to being strict about retransmits. Our
tests will now require that the shim only retransmit when we expect it to.
Change-Id: Iab1deb9665dcd294790c8253d920089e83a9140c
Reviewed-on: https://boringssl-review.googlesource.com/3211
Reviewed-by: Adam Langley <agl@google.com>
This is fatal for TLS but buffered in DTLS. The buffering isn't strictly
necessary (it would be just as valid to drop the record on the floor), but so
long as we want this behavior it should have a test.
Change-Id: I5846bb2fe80d78e25b6dfad51bcfcff2dc427c3f
Reviewed-on: https://boringssl-review.googlesource.com/3029
Reviewed-by: Adam Langley <agl@google.com>
This regressed in e95d20dcb8. EVP_AEAD will push
errors on the error queue (unlike the EVP_CIPHER codepath which checked
everything internally to ssl/ and didn't bother pushing anything). This meant
that a dropped packet would leave junk in the error queue.
Later, when SSL_read returns <= 0 (EOF or EWOULDBLOCK), the non-empty error
queue check in SSL_get_error kicks in and SSL_read looks to have failed.
BUG=https://code.google.com/p/webrtc/issues/detail?id=4214
Change-Id: I1e5e41c77a3e5b71e9eb0c72294abf0da677f840
Reviewed-on: https://boringssl-review.googlesource.com/2982
Reviewed-by: Adam Langley <agl@google.com>
BoringSSL currently retransmits non-deterministically on an internal timer
(rather than one supplied externally), so the tests currently fail flakily
depending on timing. Valgrind is a common source for this. We still assume an
in-order and reliable channel, but drop retransmits silently:
- Handshake messages may arrive with old sequence numbers.
- Retransmitted CCS records arrive from the previous epoch.
- We may receive a retransmitted Finished after we believe the handshake has
completed. (Aside: even in a real implementation, only Finished is possible
here. Even with out-of-order delivery, retransmitted or reordered messages
earlier in the handshake come in under a different epoch.)
Note that because DTLS renego and a Finished retransmit are ambiguous at the
record layer[*], this precludes us writing tests for DTLS renego. But DTLS
renego should get removed anyway. As BoringSSL currently implements renego,
this ambiguity is also a source of complexity in the real implementation. (See
the SSL3_MT_FINISHED check in dtls1_read_bytes.)
[*] As a further fun aside, it's also complex if dispatching renego vs Finished
after handshake message reassembly. The spec doesn't directly say the sequence
number is reset across renegos, but it says "The first message each side
transmits in /each/ handshake always has message_seq = 0". This means that such
an implementation needs the handshake message reassembly logic be aware that a
Finished fragment with high sequence number is NOT an out-of-order fragment for
the next handshake.
Change-Id: I35d13560f82bcb5eeda62f4de1571d28c818cc36
Reviewed-on: https://boringssl-review.googlesource.com/2770
Reviewed-by: Adam Langley <agl@google.com>
This CL removes the last of the EVP_CIPHER codepath in ssl/. The dead code is
intentionally not pruned for ease of review, except in DTLS-only code where
adding new logic to support both, only to remove half, would be cumbersome.
Fixes made:
- dtls1_retransmit_state is taught to retain aead_write_ctx rather than
enc_write_ctx.
- d1_pkt.c reserves space for the variable-length nonce when echoed into the
packet.
- dtls1_do_write sizes the MTU based on EVP_AEAD max overhead.
- tls1_change_cipher_state_cipher should not free AEAD write contexts in DTLS.
This matches the (rather confused) ownership for the EVP_CIPHER contexts.
I've added a TODO to resolve this craziness.
A follow-up CL will remove all the resultant dead code.
Change-Id: I644557f4db53bbfb182950823ab96d5e4c908866
Reviewed-on: https://boringssl-review.googlesource.com/2699
Reviewed-by: Adam Langley <agl@google.com>
The minimum MTU (not consistently enforced) is just under 256, so it's
difficult to test everything, but this is a basic test. (E.g., without renego,
the only handshake message with encryption is Finished which fits in the MTU.)
It tests the server side because the Certificate message is large enough to
require fragmentation.
Change-Id: Ida11f1057cebae2b800ad13696f98bb3a7fbbc5e
Reviewed-on: https://boringssl-review.googlesource.com/2824
Reviewed-by: Adam Langley <agl@google.com>
Add a dedicated error code to the queue for a handshake_failure alert in
response to ClientHello. This matches NSS's client behavior and gives a better
error on a (probable) failure to negotiate initial parameters.
BUG=https://crbug.com/446505
Change-Id: I34368712085a6cbf0031902daf2c00393783d96d
Reviewed-on: https://boringssl-review.googlesource.com/2751
Reviewed-by: Adam Langley <agl@google.com>
Ensure that both the client and the server emit a protocol_version alert
(except in SSLv3 where it doesn't exist) with a record-layer version which the
peer will recognize.
Change-Id: I31650a64fe9b027ff3d51e303711910a00b43d6f
This makes SSLv23_method go through DTLS_ANY_VERSION's version negotiation
logic. This allows us to get rid of duplicate ClientHello logic. For
compatibility, SSL_METHOD is now split into SSL_PROTOCOL_METHOD and a version.
The legacy version-locked methods set min_version and max_version based this
version field to emulate the original semantics.
As a bonus, we can now handle fragmented ClientHello versions now.
Because SSLv23_method is a silly name, deprecate that too and introduce
TLS_method.
Change-Id: I8b3df2b427ae34c44ecf972f466ad64dc3dbb171
These tests use both APIs. This also modifies the inline version negotiation's
error codes (currently only used for DTLS) to align with SSLv23's error codes.
Note: the peer should send a protocol_version alert which is currently untested
because it's broken.
Upstream would send such an alert if TLS 1.0 was supported but not otherwise,
which is somewhat bizarre. We've actually regressed and never send the alert in
SSLv23. When version negotiation is unified, we'll get the alerts back.
Change-Id: I4c77bcef3a3cd54a039a642f189785cd34387410
Reviewed-on: https://boringssl-review.googlesource.com/2584
Reviewed-by: Adam Langley <agl@google.com>
Amend the version negotiation tests to test this new spelling of max_version.
min_version will be tested in a follow-up.
Change-Id: Ic4bfcd43bc4e5f951140966f64bb5fd3e2472b01
Reviewed-on: https://boringssl-review.googlesource.com/2583
Reviewed-by: Adam Langley <agl@google.com>
else block got lost in a rewrite of this code.
Change-Id: I51f1655474ec8bbd4eccb4297124e8584329444e
Reviewed-on: https://boringssl-review.googlesource.com/2560
Reviewed-by: Adam Langley <agl@google.com>
The client_version needs to be preserved, both for the RSA key exchange and
(when this codepath is used for TLS) for the SChannel renego workaround. Fix
the tests to enforce this so the cipher suite version tests catch this.
Change-Id: I0c42dc3ec4830f3724026b400e5066e7a7f1ee97
Reviewed-on: https://boringssl-review.googlesource.com/2551
Reviewed-by: Adam Langley <agl@google.com>
The record-layer version of the ServerHello should match the final version. The
record-layer version of the ClientHello should be the advertised version, but
clamped at TLS 1.0. This is to ensure future rewrites do not regress this.
Change-Id: I96f1f0674944997ff38b562453a322ce61652635
Reviewed-on: https://boringssl-review.googlesource.com/2540
Reviewed-by: Adam Langley <agl@google.com>
The ClientHello record is padded to 1024 bytes when
fastradio_padding is enabled. As a result, the 3G cellular radio
is fast forwarded to DCH (high data rate) state. This mechanism
leads to a substantial redunction in terms of TLS handshake
latency, and benefits mobile apps that are running on top of TLS.
Change-Id: I3d55197b6d601761c94c0f22871774b5a3dad614
The comment has it right, but the rewritten code was wrong.
Change-Id: I450193c39fb62eae32aae090a3834dd83db53421
Reviewed-on: https://boringssl-review.googlesource.com/2444
Reviewed-by: Adam Langley <agl@google.com>
Replace the comment with a clearer one and reimplement it much more tidily. The
mask thing was more complicated than was needed.
This slightly changes behavior on the DTLS_ANY_VERSION side in that, if only
one method is enabled, we no longer short-circuit to the version-locked method
early. This "optimization" seems unnecessary.
Change-Id: I571c8b60ed16bd4357c67d65df0dd1ef9cc5eb57
Reviewed-on: https://boringssl-review.googlesource.com/2451
Reviewed-by: Adam Langley <agl@google.com>
This fixes bugs that kept the tests from working:
- Resolve DTLS version and cookie before the session.
- In DTLS_ANY_VERSION, ServerHello should be read with first_packet = 1. This
is a regression from f2fedefdca. We'll want to
do the same for TLS, but first let's change this to a boolean has_version in a
follow-up.
Things not yet fixed:
- DTLS code is not EVP_AEAD-aware. Those ciphers are disabled for now.
- On the client, DTLS_ANY_VERSION creates SSL_SESSIONs with the wrong
ssl_version. The tests pass because we no longer enforce the match as of
e37216f56009fbf48c3a1e733b7a546ca6dfc2af. (In fact, we've gone from the server
ignoring ssl_version and client enforcing to the client mostly ignoring
ssl_version and the server enforcing.)
- ssl3_send_client_hello's ssl_version check checks for equality against
s->version rather than >.
Change-Id: I5a0dde221b2009413df9b9443882b9bf3b29519c
Reviewed-on: https://boringssl-review.googlesource.com/2403
Reviewed-by: Adam Langley <agl@google.com>
We forgot to add those when we implemented the features. (Also relevant because
they will provide test coverage later for configuring features when using the
generic method tables rather than *_client_method.)
Change-Id: Ie08b27de893095e01a05a7084775676616459807
Reviewed-on: https://boringssl-review.googlesource.com/2410
Reviewed-by: Adam Langley <agl@google.com>
This implements session IDs in client and server in runner.go.
Change-Id: I26655f996b7b44c7eb56340ef6a415d3f2ac3503
Reviewed-on: https://boringssl-review.googlesource.com/2350
Reviewed-by: Adam Langley <agl@google.com>
This commit fixes a number of crashes caused by malloc failures. They
were found using the -malloc-test=0 option to runner.go which runs tests
many times, causing a different allocation call to fail in each case.
(This test only works on Linux and only looks for crashes caused by
allocation failures, not memory leaks or other errors.)
This is not the complete set of crashes! More can be found by collecting
core dumps from running with -malloc-test=0.
Change-Id: Ia61d19f51e373bccb7bc604642c51e043a74bd83
Reviewed-on: https://boringssl-review.googlesource.com/2320
Reviewed-by: Adam Langley <agl@google.com>
Just the negotiation portion as everything else is external. This feature is
used in WebRTC.
Change-Id: Iccc3983ea99e7d054b59010182f9a56a8099e116
Reviewed-on: https://boringssl-review.googlesource.com/2310
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>
All of NSS, upstream OpenSSL, SChannel, and Secure Transport require, on the
client, that the ServerHello version match the session's version on resumption.
OpenSSL's current behavior is incompatible with all of these. Fall back to a
full handshake on the server instead of mismatch.
Add a comment on the client for why we are, as of
30ddb434bf, not currently enforcing the same in
the client.
Change-Id: I60aec972d81368c4ec30e2fd515dabd69401d175
Reviewed-on: https://boringssl-review.googlesource.com/2244
Reviewed-by: Adam Langley <agl@google.com>
Clients all consistently reject mismatches. If a different version was
negotiated, a server should ignore the resumption. This doesn't actually affect
current tests. We really want to be making this change in BoringSSL (and then
upstream), but get the Go half into shape first.
Change-Id: Ieee7e141331d9e08573592e661889bd756dccfa9
Reviewed-on: https://boringssl-review.googlesource.com/2243
Reviewed-by: Adam Langley <agl@google.com>
These'll get removed once most of renego support is gone, but this is to prove
removing the warning alert from the previous commit still prevents legacy
renegotiations.
Change-Id: I7d9d95e1d4c5d23d3b6d170938a5499a65f2d5ea
Reviewed-on: https://boringssl-review.googlesource.com/2236
Reviewed-by: Adam Langley <agl@google.com>
Ensure that the client rejects it with UNEXPECTED_MESSAGE, not by attempting to
decode it.
Change-Id: Ifc5613cf1152e0f7dcbee73e05df1ef367dfbfd5
Reviewed-on: https://boringssl-review.googlesource.com/2232
Reviewed-by: Adam Langley <agl@google.com>
At the record layer, DTLS maintains a window of seen sequence numbers to detect
replays. Add tests to cover that case. Test both repeated sequence numbers
within the window and sequence numbers past the window's left edge. Also test
receiving sequence numbers far past the window's right edge.
Change-Id: If6a7a24869db37fdd8fb3c4b3521b730e31f8f86
Reviewed-on: https://boringssl-review.googlesource.com/2221
Reviewed-by: Adam Langley <agl@google.com>
