This fixes the following vulnerability scenario:
* Victim logs into SAFE-HOST on SAFE-IP-ADDRESS regularly.
* Victim is cajoled into connecting to attacker controlled
ATTACK-HOST, on ATTACK-IP-ADDRESS. ATTACK-HOST uses a different host
key type (e.g. Ed25519 vs RSA). The new key is added at the end of
known_hosts.
* Attacker makes DNS system return ATTACK-IP-ADDRESS for SAFE-HOST.
* Victim logs into SAFE-HOST, but is not warned because the host key
matches ATTACK-IP-ADDRESS.
For this attack to work, the key type has to be different, because
knownhosts gives precedence to the first key found for each type. Add
a test that asserts this behavior.
The new semantics simplify the code, but callers that modify
.ssh/known_host interactviely must now take an extra step to remain
OpenSSH compatible: on successful login, the IP address must be
checked without hostname, and if it is not known, added separately to
the known_hosts file, so future logins that use an IP address only
will be protected too.
Thanks to Daniel Parks <security@demonhorse.net> for finding this
vulnerability.
Change-Id: I62b1b60ceb02e2f583a4657213feac1a8885dd42
Reviewed-on: https://go-review.googlesource.com/104939
Reviewed-by: Adam Langley <agl@golang.org>
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
The extra Close call runs in a separate goroutine, and races with the
deferred Close call, leading to test flakes if the extra call fails
Change-Id: Ia1a2fb3932906da7cd0a8854b55f82785bb33716
Reviewed-on: https://go-review.googlesource.com/104335
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Incomplete reads leave (part of) the verification tag zeroed, leading
to a failing MAC, and this is more likely to happen with larger
packets. The test added in the previous commit should stop this from
regressing.
Reinstate chacha20-poly1305 as a default cipher and prefer it over AES
CTR flavors.
Fixesgolang/go#23510
Change-Id: I7599897e59448edb7b814eebcc8226ea15b365d6
Reviewed-on: https://go-review.googlesource.com/89075
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This effectively disables the cipher.
Add a test against OpenSSH which sends larger packets through the
tested ciphers. This reproduces the problem reported in golang/go#23510
Change-Id: I4b124c690c409c6a0af2621e58a964ff55815f57
Reviewed-on: https://go-review.googlesource.com/88995
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Adam Langley <agl@golang.org>
This exposes the chacha20 stream cipher to the entire x/crypto
package, and in particular to the SSH package, which uses separate,
unauthenticated chacha20 encryption for packet lengths.
Change-Id: I0b705482128f0657c09292370f03d08b588f7fec
Reviewed-on: https://go-review.googlesource.com/87075
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
The writing side would generate a maximum of 19 bytes of padding, so
the reading side erroneously checked this. However, RFC 5647 specifies
255 as the maximum amount of padding for AES-GCM.
Fixesgolang/go#18953.
Change-Id: I416b0023c6e4cbd91a6a1b4214a03f1663b77248
Reviewed-on: https://go-review.googlesource.com/47590
Reviewed-by: Adam Langley <agl@golang.org>
Run-TryBot: Adam Langley <agl@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Fixes a nil pointer dereference that slipped through buildbots because
it was introduced by the last two commits.
Change-Id: Ib269e910956cd8b3b46e217b03fde1b61572260a
Reviewed-on: https://go-review.googlesource.com/40530
Reviewed-by: Han-Wen Nienhuys <hanwen@google.com>
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
The normal handshake kicks off with a waitSession(), which guarantees
that we never attempt to send data before the first kex is completed,
but ensuring readPacketsLeft > 0 and writePacketsLeft > 0 helps
understand that thresholds can never cause spurious rekeying at the
start of a connection.
Change-Id: If5bcafcda0c7d16fd21f22c664101ac5f5b487d7
Reviewed-on: https://go-review.googlesource.com/38696
Reviewed-by: Adam Langley <agl@golang.org>
Run-TryBot: Adam Langley <agl@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This change breaks existing behavior.
Before, a missing ClientConfig.HostKeyCallback would cause host key
checking to be disabled. In this configuration, establishing a
connection to any host just works, so today, most SSH client code in
the wild does not perform any host key checks.
This makes it easy to perform a MITM attack:
* SSH installations that use keyboard-interactive or password
authentication can be attacked with MITM, thereby stealing
passwords.
* Clients that use public-key authentication with agent forwarding are
also vulnerable: the MITM server could allow the login to succeed, and
then immediately ask the agent to authenticate the login to the real
server.
* Clients that use public-key authentication without agent forwarding
are harder to attack unnoticedly: an attacker cannot authenticate the
login to the real server, so it cannot in general present a convincing
server to the victim.
Now, a missing HostKeyCallback will cause the handshake to fail. This
change also provides InsecureIgnoreHostKey() and FixedHostKey(key) as
ready made host checkers.
A simplistic parser for OpenSSH's known_hosts file is given as an
example. This change does not provide a full-fledged parser, as it
has complexity (wildcards, revocation, hashed addresses) that will
need further consideration.
When introduced, the host checking feature maintained backward
compatibility at the expense of security. We have decided this is not
the right tradeoff for the SSH library.
Fixesgolang/go#19767
Change-Id: I45fc7ba9bd1ea29c31ec23f115cdbab99913e814
Reviewed-on: https://go-review.googlesource.com/38701
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
The previous attempt would fail in the following scenario:
* select picks "first" kex from requestKex
* read loop receives a remote kex, posts on requestKex (which is now
empty) [*] for sending out a response, and sends pendingKex on startKex.
* select picks pendingKex from startKex, and proceeds to run the key
exchange.
* the posting on requestKex in [*] now triggers a second key exchange.
Fixes#18861.
Change-Id: I443e82f1d04c7f17d1485fdb87072b9feec26aa8
Reviewed-on: https://go-review.googlesource.com/36055
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Han-Wen Nienhuys <hanwen@google.com>
Since encryption messes up the packets, the wrongly retained packets
look like noise and cause application protocol errors or panics in the
SSH library.
This normally triggers very rarely: the mandatory key exchange doesn't
have parallel writes, so this failure condition would be setup on the
first key exchange, take effect only after the second key exchange.
Fortunately, the tests against openssh exercise this. This change adds
also adds a unittest.
Fixes#18850.
Change-Id: I656c8b94bfb265831daa118f4d614a2f0c65d2af
Reviewed-on: https://go-review.googlesource.com/36056
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
1) Always force a key exchange if we exchange 2^31 packets. In the past
this might not happen if RekeyThreshold was set to a very large
interval.
2) Follow recommendations from RFC 4344 for block ciphers. For AES, we
can encrypt 2^(blocksize/4) blocks under the same keys.
On modern hardware, the previous default of 1Gb could force a key
exchange within ~10 seconds. Since the key exchange takes 3 roundtrips
(send kex init, send DH init, send NEW_KEYS), this is relatively
expensive on high-latency links.
Change-Id: I1297124a307c541b7bf22d814d136ec0c6d8ed97
Reviewed-on: https://go-review.googlesource.com/35410
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Adam Langley <agl@golang.org>
The initial kex is started from both sides simultaneously, and before,
we could consume the the incoming kex request before we consumed from
our internal channel. This would result in initiating a key exchange
just after completing the initial one, which is not only an extra
delay, but also an error when using OpenSSH (OpenSSH does not support
key exchanges during user authentication).
Change-Id: Ia7e0748ea2bca80ae97d187bcf2931ab6422276b
Reviewed-on: https://go-review.googlesource.com/35851
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Adam Langley <agl@golang.org>
Use channels and a dedicated write loop for managing the rekeying
process. This lets us collect packets to be written while a key
exchange is in progress.
Previously, the read loop ran the key exchange, and writers would
block if a key exchange was going on. If a reader wrote back a packet
while processing a read packet, it could block, stopping the read
loop, thus causing a deadlock. Such coupled read/writes are inherent
with handling requests that want a response (eg. keepalive,
opening/closing channels etc.). The buffered channels (most channels
have capacity 16) papered over these problems, but under load SSH
connections would occasionally deadlock.
Fixes#18439.
Change-Id: I7c14ff4991fa3100a5d36025125d0cf1119c471d
Reviewed-on: https://go-review.googlesource.com/35012
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Han-Wen Nienhuys <hanwen@google.com>
According to RFC 4254 section 6.10, SSH server implementations may
omit the exit-status and exit-signal messages. If this happens, we
now return &ExitMissingError{}, so clients can handle this case
specifically.
This came up in the discussion of issue #16194.
Change-Id: Iae5e916b18aa5bd8e95618e9fcfcab8b19e147d9
Reviewed-on: https://go-review.googlesource.com/24727
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
If one of both sides is slow, the first kex completes implicitly. The
first kex also produces msgNewKeys, and this must be read to ensure
that the authentication code is not confused by it.
Before this fix, the problem could be reproduced by inserting a sleep
just before the requestInitialKeyChange call.
Fixes#15198
Change-Id: I602db5dd37b2d8556c88ab4cdb693ccf90147a3d
Reviewed-on: https://go-review.googlesource.com/23137
Run-TryBot: Han-Wen Nienhuys <hanwen@google.com>
Reviewed-by: Adam Langley <agl@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This is done by running the key exchange and setting the session ID
under mutex. If the first exchange encounters an already set session
ID, then do nothing.
This fixes a race condition:
On setting up the connection, both sides sent a kexInit to initiate
the first (mandatory) key exchange. If one side was faster, the
faster side might have completed the key exchange, before the slow
side had a chance to send a kexInit. The slow side would send a
kexInit which would trigger a second key exchange. The resulting
confirmation message (msgNewKeys) would confuse the authentication
loop.
This fix removes sessionID from the transport struct.
This fix also deletes the unused interface rekeyingTransport.
Fixes#15066
Change-Id: I7f303bce5d3214c9bdd58f52d21178a185871d90
Reviewed-on: https://go-review.googlesource.com/21606
Reviewed-by: Adam Langley <agl@golang.org>
Reviewed-by: Han-Wen Nienhuys <hanwen@google.com>
Unblock writers if a read error occurs while writers are blocked on a
pending key change.
Add test to check for deadlocks in error paths in handshake.go
Fixesgolang/go#11992.
Change-Id: Id098bd9fec3d4fe83daeb2b7f935e5647c19afd3
Reviewed-on: https://go-review.googlesource.com/13594
Reviewed-by: Adam Langley <agl@golang.org>
This deprives an attacker of feedback for guesses against the packet
length given by the connection dropping.
Change-Id: I14939a82e5243a86d192bb18be93d45589227147
Reviewed-on: https://go-review.googlesource.com/9908
Reviewed-by: Adam Langley <agl@golang.org>
Test that different extended data streams within a channel are
thread-safe.
benchmark old MB/s new MB/s speedup
BenchmarkEndToEnd 79.26 87.98 1.11x
benchmark old allocs new allocs delta
BenchmarkEndToEnd 110 73 -33.64%
benchmark old bytes new bytes delta
BenchmarkEndToEnd 2605720 1299768 -50.12%
LGTM=dave, jpsugar
R=agl, dave, jpsugar
CC=golang-codereviews
https://golang.org/cl/136420043
set.
Otherwise, the package leaks data about total number of
connections established through its channel IDs.
R=agl, jpsugar
CC=golang-codereviews
https://golang.org/cl/87280043
transport.
Sending the msgNewKeys packet and setting up the key material
now happen under a lock, preventing races with concurrent
writers.
R=kardianos, agl, jpsugar, hanwenn
CC=golang-dev
https://golang.org/cl/14476043
Adds readPacket() to conn, and renames conn to packetConn.
Key exchanges operate on packetConn, so they can be
unittested.
R=agl, jpsugar, dave
CC=golang-dev
https://golang.org/cl/13352055
This obviates custom code to emulate a thread-safe connection.
Use this for testing that listeners close if the connection breaks.
R=dave, agl, fullung
CC=golang-dev
https://golang.org/cl/11781043
* Remove special handling for dynamically allocated
ports. This was a bug in OpenSSH 5.x sshd.
* Run the test with a preselected port number.
* Run TestPortForward only on unix platforms.
R=dave, agl
CC=golang-dev
https://golang.org/cl/10049045
Set maxPacket in forwarded connection, and use the requested port
number as key in forwardList.
R=golang-dev, agl, dave
CC=golang-dev
https://golang.org/cl/9753044