we will support then dtls1_do_write can go into an infinite loop. This commit
fixes that.
Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit d3d9eef316)
at least the minimum or it will fail.
There were some instances in dtls1_query_mtu where the final mtu can end up
being less than the minimum, i.e. where the user has set an mtu manually. This
shouldn't be allowed. Also remove dtls1_guess_mtu that, despite having
logic for guessing an mtu, was actually only ever used to work out the minimum
mtu to use.
Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit 1620a2e49c)
and instead use the value provided by the underlying BIO. Also provide some
new DTLS_CTRLs so that the library user can set the mtu without needing to
know this constant. These new DTLS_CTRLs provide the capability to set the
link level mtu to be used (i.e. including this IP/UDP overhead). The previous
DTLS_CTRLs required the library user to subtract this overhead first.
Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit 59669b6abf)
Conflicts:
ssl/d1_both.c
ssl/ssl_lib.c
mtu that we have received is not less than the minimum. If its less it uses the
minimum instead. The second call to query the mtu does not do that, but
instead uses whatever comes back. We have seen an instance in RT#3592 where we
have got an unreasonably small mtu come back. This commit makes both query
checks consistent.
Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit 6abb0d1f8e)
automatically updated, and we should use the one provided instead.
Unfortunately there are a couple of locations where this is not respected.
Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit 001235778a)
RT#3592 provides an instance where the OPENSSL_assert that this commit
replaces can be hit. I was able to recreate this issue by forcing the
underlying BIO to misbehave and come back with very small mtu values. This
happens the second time around the while loop after we have detected that the
MTU has been exceeded following the call to dtls1_write_bytes.
Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit cf75017bfd)
Previously, state variant was not advanced, which resulted in state
being stuck in the st1 variant (usually "_A").
This broke certificate callback retry logic when accepting connections
that were using SSLv2 ClientHello (hence reusing the message), because
their state never advanced to SSL3_ST_SR_CLNT_HELLO_C variant required
for the retry code path.
Reported by Yichun Zhang (agentzh).
Signed-off-by: Piotr Sikora <piotr@cloudflare.com>
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
the session's version (server).
See also BoringSSL's commit bdf5e72f50e25f0e45e825c156168766d8442dde.
Reviewed-by: Dr. Stephen Henson <steve@openssl.org>
(cherry picked from commit 9e189b9dc1)
once the ChangeCipherSpec message is received. Previously, the server would
set the flag once at SSL3_ST_SR_CERT_VRFY and again at SSL3_ST_SR_FINISHED.
This would allow a second CCS to arrive and would corrupt the server state.
(Because the first CCS would latch the correct keys and subsequent CCS
messages would have to be encrypted, a MitM attacker cannot exploit this,
though.)
Thanks to Joeri de Ruiter for reporting this issue.
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit e94a6c0ede)
Conflicts:
CHANGES
ssl/s3_srvr.c
The server must send a NewSessionTicket message if it advertised one
in the ServerHello, so make a missing ticket message an alert
in the client.
An equivalent change was independently made in BoringSSL, see commit
6444287806d801b9a45baf1f6f02a0e3a16e144c.
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit de2c7504eb)
Conflicts:
CHANGES
The client sends a session ID with the session ticket, and uses
the returned ID to detect resumption, so we do not need to peek
at handshake messages: s->hit tells us explicitly if we're resuming.
An equivalent change was independently made in BoringSSL, see commit
407886f589cf2dbaed82db0a44173036c3bc3317.
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 980bc1ec61)
Conflicts:
ssl/d1_clnt.c
ssl/s3_clnt.c
The same change was independently made in BoringSSL, see commit
9eaeef81fa2d4fd6246dc02b6203fa936a5eaf67
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 7b3ba508af)
This ensures that it's zeroed even if the SSL object is reused
(as in ssltest.c). It also ensures that it applies to DTLS, too.
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit a06cd5d056)
When no-ssl3 is set only make SSLv3 disabled by default. Retain -ssl3
options for s_client/s_server/ssltest.
When no-ssl3-method is set SSLv3_*method() is removed and all -ssl3
options.
We should document this somewhere, e.g. wiki, FAQ or manual page.
Reviewed-by: Emilia Käsper <emilia@openssl.org>
(cherry picked from commit 3881d8106d)
Conflicts:
util/mkdef.pl
Tighten client-side session ticket handling during renegotiation:
ensure that the client only accepts a session ticket if the server sends
the extension anew in the ServerHello. Previously, a TLS client would
reuse the old extension state and thus accept a session ticket if one was
announced in the initial ServerHello.
Reviewed-by: Bodo Moeller <bodo@openssl.org>
(cherry picked from commit d663df2399)
Conflicts:
CHANGES
When we're configured with no-ssl3 and we receive an SSL v3 Client Hello, we set
the method to NULL. We didn't used to do that, and it breaks things. This is a
regression introduced in 62f45cc27d. Keep the old
method since the code is not able to deal with a NULL method at this time.
CVE-2014-3569, PR#3571
Reviewed-by: Emilia Käsper <emilia@openssl.org>
(cherry picked from commit 392fa7a952)
CVE-2014-3513
This issue was reported to OpenSSL on 26th September 2014, based on an origi
issue and patch developed by the LibreSSL project. Further analysis of the i
was performed by the OpenSSL team.
The fix was developed by the OpenSSL team.
Reviewed-by: Tim Hudson <tjh@openssl.org>
algorithms MD2 and RC5 don't get built.
Also, disable building the test apps in crypto/des and crypto/pkcs7, as
they have no support at all.
Reviewed-by: Tim Hudson <tjh@openssl.org>
SSL_set_SSL_CTX is normally called for SNI after ClientHello has
received and the digest to use for each certificate has been decided.
The original ssl->cert contains the negotiated digests and is now
copied to the new ssl->cert.
PR: 3560
Reviewed-by: Tim Hudson <tjh@openssl.org>
(Original commit adb46dbc6d)
Use the new constant-time methods consistently in s3_srvr.c
Reviewed-by: Kurt Roeckx <kurt@openssl.org>
(cherry picked from commit 455b65dfab)
that bad encryptions are treated like random session keys in constant
time.
(cherry picked from commit adb46dbc6d)
Reviewed-by: Rich Salz <rsalz@openssl.org>
Also tweak s3_cbc.c to use new constant-time methods.
Also fix memory leaks from internal errors in RSA_padding_check_PKCS1_OAEP_mgf1
This patch is based on the original RT submission by Adam Langley <agl@chromium.org>,
as well as code from BoringSSL and OpenSSL.
Reviewed-by: Kurt Roeckx <kurt@openssl.org>
Conflicts:
crypto/rsa/rsa_oaep.c
that fixed PR#3450 where an existing cast masked an issue when i was changed
from int to long in that commit
Picked up on z/linux (s390) where sizeof(int)!=sizeof(long)
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit b5ff559ff9)
Fix a bug in handling of 128 byte long PSK identity in
psk_client_callback.
OpenSSL supports PSK identities of up to (and including) 128 bytes in
length. PSK identity is obtained via the psk_client_callback,
implementors of which are expected to provide a NULL-terminated
identity. However, the callback is invoked with only 128 bytes of
storage thus making it impossible to return a 128 byte long identity and
the required additional NULL byte.
This CL fixes the issue by passing in a 129 byte long buffer into the
psk_client_callback. As a safety precaution, this CL also zeroes out the
buffer before passing it into the callback, uses strnlen for obtaining
the length of the identity returned by the callback, and aborts the
handshake if the identity (without the NULL terminator) is longer than
128 bytes.
(Original patch amended to achieve strnlen in a different way.)
Reviewed-by: Rich Salz <rsalz@openssl.org>
(cherry picked from commit be0d851732)
Pull constant-time methods out to a separate header, add tests.
Reviewed-by: Bodo Moeller <bodo@openssl.org>
(cherry picked from commit 9a9b0c0401)
Conflicts:
test/Makefile
Limit the number of empty records that will be processed consecutively
in order to prevent ssl3_get_record from never returning.
Reported by "oftc_must_be_destroyed" and George Kadianakis.
Reviewed-by: Bodo Moeller <bodo@openssl.org>
(cherry picked from commit 3aac17a82f)
The addition of SRP authentication needs to be checked in various places
to work properly. Specifically:
A certificate is not sent.
A certificate request must not be sent.
Server key exchange message must not contain a signature.
If appropriate SRP authentication ciphersuites should be chosen.
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 8f5a8805b82d1ae81168b11b7f1506db9e047dec)
Conflicts:
ssl/s3_clnt.c
ssl/s3_lib.c
If a client attempted to use an SRP ciphersuite and it had not been
set up correctly it would crash with a null pointer read. A malicious
server could exploit this in a DoS attack.
Thanks to Joonas Kuorilehto and Riku Hietamäki from Codenomicon
for reporting this issue.
CVE-2014-5139
Reviewed-by: Tim Hudson <tjh@openssl.org>
In a couple of functions, a sequence number would be calculated twice.
Additionally, in |dtls1_process_out_of_seq_message|, we know that
|frag_len| <= |msg_hdr->msg_len| so the later tests for |frag_len <
msg_hdr->msg_len| can be more clearly written as |frag_len !=
msg_hdr->msg_len|, since that's the only remaining case.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Emilia Käsper <emilia@openssl.org>
Previously, a truncated DTLS fragment in
|dtls1_process_out_of_seq_message| would cause *ok to be cleared, but
the return value would still be the number of bytes read. This would
cause |dtls1_get_message| not to consider it an error and it would
continue processing as normal until the calling function noticed that
*ok was zero.
I can't see an exploit here because |dtls1_get_message| uses
|s->init_num| as the length, which will always be zero from what I can
see.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Emilia Käsper <emilia@openssl.org>
The |pqueue_insert| function can fail if one attempts to insert a
duplicate sequence number. When handling a fragment of an out of
sequence message, |dtls1_process_out_of_seq_message| would not call
|dtls1_reassemble_fragment| if the fragment's length was zero. It would
then allocate a fresh fragment and attempt to insert it, but ignore the
return value, leaking the fragment.
This allows an attacker to exhaust the memory of a DTLS peer.
Fixes CVE-2014-3507
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Emilia Käsper <emilia@openssl.org>
In |dtls1_reassemble_fragment|, the value of
|msg_hdr->frag_off+frag_len| was being checked against the maximum
handshake message size, but then |msg_len| bytes were allocated for the
fragment buffer. This means that so long as the fragment was within the
allowed size, the pending handshake message could consume 16MB + 2MB
(for the reassembly bitmap). Approx 10 outstanding handshake messages
are allowed, meaning that an attacker could consume ~180MB per DTLS
connection.
In the non-fragmented path (in |dtls1_process_out_of_seq_message|), no
check was applied.
Fixes CVE-2014-3506
Wholly based on patch by Adam Langley with one minor amendment.
Reviewed-by: Emilia Käsper <emilia@openssl.org>
The |item| variable, in both of these cases, may contain a pointer to a
|pitem| structure within |s->d1->buffered_messages|. It was being freed
in the error case while still being in |buffered_messages|. When the
error later caused the |SSL*| to be destroyed, the item would be double
freed.
Thanks to Wah-Teh Chang for spotting that the fix in 1632ef74 was
inconsistent with the other error paths (but correct).
Fixes CVE-2014-3505
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Emilia Käsper <emilia@openssl.org>
Don't call internal functions directly call them through
SSL_test_functions(). This also makes unit testing work on
Windows and platforms that don't export internal functions
from shared libraries.
By default unit testing is not enabled: it requires the compile
time option "enable-unit-test".
Reviewed-by: Geoff Thorpe <geoff@openssl.org>
(cherry picked from commit e0fc7961c4)
Conflicts:
ssl/Makefile
util/mkdef.pl
In the ssl_cipher_get_evp() function, fix off-by-one errors in index validation before accessing arrays.
Bug discovered and fixed by Miod Vallat from the OpenBSD team.
PR#3375
Allow CCS after finished has been sent by client: at this point
keys have been correctly set up so it is OK to accept CCS from
server. Without this renegotiation can sometimes fail.
PR#3400
(cherry picked from commit 99cd6a91fcb0931feaebbb4832681d40a66fad41)
Defines SETUP_TEST_FIXTURE and EXECUTE_TEST, and updates ssl/heartbeat_test.c
using these macros. SETUP_TEST_FIXTURE makes use of the new TEST_CASE_NAME
macro, defined to use __func__ or __FUNCTION__ on platforms that support those
symbols, or to use the file name and line number otherwise. This should fix
several reported build problems related to lack of C99 support.
SRP ciphersuites do not have no authentication. They have authentication
based on SRP. Add new SRP authentication flag and cipher string.
(cherry picked from commit a86b88acc373ac1fb0ca709a5fb8a8fa74683f67)
If application uses tls_session_secret_cb for session resumption
set the CCS_OK flag.
(cherry picked from commit 953c592572e8811b7956cc09fbd8e98037068b58)
Unnecessary recursion when receiving a DTLS hello request can be used to
crash a DTLS client. Fixed by handling DTLS hello request without recursion.
Thanks to Imre Rad (Search-Lab Ltd.) for discovering this issue.
Only accept change cipher spec when it is expected instead of at any
time. This prevents premature setting of session keys before the master
secret is determined which an attacker could use as a MITM attack.
Thanks to KIKUCHI Masashi (Lepidum Co. Ltd.) for reporting this issue
and providing the initial fix this patch is based on.
A buffer overrun attack can be triggered by sending invalid DTLS fragments
to an OpenSSL DTLS client or server. This is potentially exploitable to
run arbitrary code on a vulnerable client or server.
Fixed by adding consistency check for DTLS fragments.
Thanks to Jüri Aedla for reporting this issue.
Add TLS padding extension to SSL_OP_ALL so it is used with other
"bugs" options and can be turned off.
This replaces SSL_OP_SSLREF2_REUSE_CERT_TYPE_BUG which is an ancient
option referring to SSLv2 and SSLREF.
PR#3336
Replace manual ASN.1 decoder with ASN1_get object. This
will decode the tag and length properly and check against
it does not exceed the supplied buffer length.
PR#3335
(cherry picked from commit b0308dddd1cc6a8e1de803ef29ba6da25ee072c2)
A missing bounds check in the handling of the TLS heartbeat extension
can be used to reveal up to 64k of memory to a connected client or
server.
Thanks for Neel Mehta of Google Security for discovering this bug and to
Adam Langley <agl@chromium.org> and Bodo Moeller <bmoeller@acm.org> for
preparing the fix (CVE-2014-0160)
The flag SSL_OP_MSIE_SSLV2_RSA_PADDING hasn't done anything since OpenSSL
0.9.7h but deleting it will break source compatibility with any software
that references it. Restore it but #define to zero.
(cherry picked from commit b17d6b8d1d)
When sending an invalid version number alert don't change the
version number to the client version if a session is already
established.
Thanks to Marek Majkowski for additional analysis of this issue.
PR#3191
For DTLS we might need to retransmit messages from the previous session
so keep a copy of write context in DTLS retransmission buffers instead
of replacing it after sending CCS. CVE-2013-6450.
When deciding whether to use TLS 1.2 PRF and record hash algorithms
use the version number in the corresponding SSL_METHOD structure
instead of the SSL structure. The SSL structure version is sometimes
inaccurate. Note: OpenSSL 1.0.2 and later effectively do this already.
(CVE-2013-6449)
PR: 2808
With DTLS/SCTP the SCTP extension SCTP-AUTH is used to protect DATA and
FORWARD-TSN chunks. The key for this extension is derived from the
master secret and changed with the next ChangeCipherSpec, whenever a new
key has been negotiated. The following Finished then already uses the
new key. Unfortunately, the ChangeCipherSpec and Finished are part of
the same flight as the ClientKeyExchange, which is necessary for the
computation of the new secret. Hence, these messages are sent
immediately following each other, leaving the server very little time to
compute the new secret and pass it to SCTP before the finished arrives.
So the Finished is likely to be discarded by SCTP and a retransmission
becomes necessary. To prevent this issue, the Finished of the client is
still sent with the old key.
(cherry picked from commit 9fb523adce)
(cherry picked from commit b9ef52b078)
Since the TLS 1.2 supported signature algorithms extension is less
sophisticaed in OpenSSL 1.0.1 this has to be done in two stages.
RSA+MD5 is removed from supported signature algorithms extension:
any compliant implementation should never use RSA+MD5 as a result.
To cover the case of a broken implementation using RSA+MD5 anyway
disable lookup of MD5 algorithm in TLS 1.2.
(I'd rather use an option, but it appears that the options field is
full.)
Now, we send the time in the gmt_unix_time field if the appropriate
one of these mode options is set, but randomize the field if the flag
is not set.
Instead, send random bytes.
While the gmt_unix_time record was added in an ostensible attempt to
mitigate the dangers of a bad RNG, its presence leaks the host's view
of the current time in the clear. This minor leak can help
fingerprint TLS instances across networks and protocols... and what's
worse, it's doubtful thet the gmt_unix_time record does any good at
all for its intended purpose, since:
* It's quite possible to open two TLS connections in one second.
* If the PRNG output is prone to repeat itself, ephemeral
* handshakes (and who knows what else besides) are broken.
This fix ensures that
* A HelloRequest is retransmitted if not responded by a ClientHello
* The HelloRequest "consumes" the sequence number 0. The subsequent
ServerHello uses the sequence number 1.
* The client also expects the sequence number of the ServerHello to
be 1 if a HelloRequest was received earlier.
This patch fixes the RFC violation.
(cherry picked from commit b62f4daac0)
The version check for DTLS1_VERSION was redundant as
DTLS1_VERSION > TLS1_1_VERSION, however we do need to
check for DTLS1_BAD_VER for compatibility.
PR:2984
(cherry picked from commit d980abb22e)
Now we set the current certificate to the one used by a server
there is no need to call ssl_get_server_send_cert which will
fail if we haven't sent a certificate yet.
Fix the calculation that checks there is enough room in a record
after removing padding and optional explicit IV. (by Steve)
For AEAD remove the correct number of padding bytes (by Andy)
MD5 should use little endian order. Fortunately the only ciphersuite
affected is EXP-RC2-CBC-MD5 (TLS_RSA_EXPORT_WITH_RC2_CBC_40_MD5) which
is a rarely used export grade ciphersuite.
We have to use EVP in FIPS mode so we can only partially mitigate
timing differences.
Make an extra call to EVP_DigestSignUpdate to hash additonal blocks
to cover any timing differences caused by removal of padding.
The previous CBC patch was bugged in that there was a path through enc()
in s3_pkt.c/d1_pkt.c which didn't set orig_len. orig_len would be left
at the previous value which could suggest that the packet was a
sufficient length when it wasn't.
This patch makes the decoding of SSLv3 and TLS CBC records constant
time. Without this, a timing side-channel can be used to build a padding
oracle and mount Vaudenay's attack.
This patch also disables the stitched AESNI+SHA mode pending a similar
fix to that code.
In order to be easy to backport, this change is implemented in ssl/,
rather than as a generic AEAD mode. In the future this should be changed
around so that HMAC isn't in ssl/, but crypto/ as FIPS expects.
This change adds CRYPTO_memcmp, which compares two vectors of bytes in
an amount of time that's independent of their contents. It also changes
several MAC compares in the code to use this over the standard memcmp,
which may leak information about the size of a matching prefix.
change the current certificate (in s->cert->key) to the one used and then
SSL_get_certificate and SSL_get_privatekey will automatically work.
Note for 1.0.1 and earlier also includes backport of the function
ssl_get_server_send_pkey.
Reported by: Phil Pennock <openssl-dev@spodhuis.org>
Make renegotiation work for TLS 1.2, 1.1 by not using a lower record
version client hello workaround if renegotiating.