Darren Tucker [Fri, 25 Oct 2024 08:04:30 +0000 (19:04 +1100)]
Retire the minix3 test config.
It got broken by the sshd-auth change, it's not obvious why, and the
platform lacks the debugging tools (eg gdb, strace) to figure it out.
The upstream project seems effectively dead (6 years since the last
commit, 10 since the last release). It was useful while it lasted
(we found a real bug because of it) but its time seems to have passed.
Darren Tucker [Fri, 25 Oct 2024 08:01:02 +0000 (19:01 +1100)]
Simplify pselect shim and remove side effects.
Instead of maintaing state (pipe descriptors, signal handlers) across
pselect-on-select invocations, set up and restore them each call.
This prevents outside factors (eg a closefrom or signal handler
installation) from potentially causing problems. This does result in a
drop in throughput of a couple of percent on geriatric platforms without
a native pselect due to the extra overhead. Tweaks & ok djm@
... and ssh and sshd log wrappers before recreating them. Prevents "can't
create" errors during tests when running tests without SUDO after having
run them with SUDO.
... to run all of the subprograms from the build directory while
developing and debugging. Should help prevent accidentally testing
against unchanged installed sshd-auth and sshd-session binaries. ok djm@
Darren Tucker [Thu, 17 Oct 2024 09:50:29 +0000 (20:50 +1100)]
Seed RNG when starting up sshd-auth.
Makes builds configured --without-openssl work again since otherwise
the first use of the RNG comes after the sandbox init and it can't
open /dev/random.
This splits the user authentication code from the sshd-session
binary into a separate sshd-auth binary. This will be executed by
sshd-session to complete the user authentication phase of the
protocol only.
Splitting this code into a separate binary ensures that the crucial
pre-authentication attack surface has an entirely disjoint address
space from the code used for the rest of the connection. It also
yields a small runtime memory saving as the authentication code will
be unloaded after thhe authentication phase completes.
djm@openbsd.org [Sun, 6 Oct 2024 23:37:17 +0000 (23:37 +0000)]
upstream: Turn off finite field (a.k.a modp) Diffie-Hellman key
exchange in sshd by default. Specifically, this removes the
diffie-hellman-group* and diffie-hellman-group-exchange-* methods. The client
is unchanged and continues to support these methods by default.
Finite field Diffie Hellman is slow and computationally expensive for
the same security level as Elliptic Curve DH or PQ key agreement while
offering no redeeming advantages.
ECDH has been specified for the SSH protocol for 15 years and some
form of ECDH has been the default key exchange in OpenSSH for the last
14 years.
upstream: fix regression introduced when I switched the "Match"
criteria tokeniser to a more shell-like one. Apparently the old tokeniser
(accidentally?) allowed "Match criteria=argument" as well as the "Match
criteria argument" syntax that we tested for.
People were using this syntax so this adds back support for
"Match criteria=argument"
upstream: update the Streamlined NTRU Prime code from the "ref"
implementation in SUPERCOP 20201130 to the "compact" implementation in
SUPERCOP 20240808. The new version is substantially faster. Thanks to Daniel
J Bernstein for pointing out the new implementation (and of course for
writing it).
Damien Miller [Mon, 9 Sep 2024 07:30:38 +0000 (17:30 +1000)]
declare defeat trying to detect C89 compilers
I can't find a reliable way to detect the features the ML-KEM code
requires in configure. Give up for now and use VLA support (that we
can detect) as a proxy for "old compiler" and turn off ML-KEM if
it isn't supported.
Damien Miller [Mon, 9 Sep 2024 06:06:21 +0000 (16:06 +1000)]
test for compiler feature needed for ML-KEM
The ML-KEM implementation we uses need the compiler to support
C99-style named struct initialisers (e.g foo = {.bar = 1}). We
still support (barely) building OpenSSH with older compilers, so
add a configure test for this.
upstream: make parsing user@host consistently look for the last '@' in
the string rather than the first. This makes it possible to use usernames
that contain '@' characters.
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
upstream: Add experimental support for hybrid post-quantum key exchange
ML-KEM768 with ECDH/X25519 from the Internet-draft:
https://datatracker.ietf.org/doc/html/draft-kampanakis-curdle-ssh-pq-ke-03
This is based on previous patches from markus@ but adapted to use the
final FIPS203 standard ML-KEM using a formally-verified implementation
from libcrux.
Note this key exchange method is still a draft and thus subject to
change. It is therefore disabled by default; set MLKEM=yes to build it.
We're making it available now to make it easy for other SSH
implementations to test against it.
Antonio Larrosa [Fri, 23 Aug 2024 10:21:06 +0000 (12:21 +0200)]
Don't skip audit before exitting cleanup_exit
This fixes an issue where the SSH_CONNECTION_ABANDON event is not
audited because cleanup_exit overrides the regular _exit too soon and
as a result, failed auth attempts are not logged correctly.
The problem was introduced in 81c1099d22b81ebfd20a334ce986c4f753b0db29
where the code from upstream was merged before the audit_event call when
it should have been merged right before the _exit call in order to honor
the comment that just mentions an override of the exit value.
upstream: As defined in the RFC, the SSH protocol has negotiable
compression support (which is requested as the name "zlib"). Compression
starts very early in the session. Relative early in OpenSSH lifetime, privsep
was added to sshd, and this required a shared-memory hack so the two
processes could see what was going on in the dataflow. This shared-memory
hack was soon recognized as a tremendous complexity risk, because it put libz
(which very much trusts it's memory) in a dangerous place, and a new option
("zlib@openssh.com") was added begins compression after authentication (aka
delayed-compression). That change also permitted removal of the
shared-memory hack. Despite removal from the server, the old "zlib" support
remained in the client, to allow negotiation with non-OpenSSH daemons which
lack the delayed-compression option. This commit deletes support for the
older "zlib" option in the client. It reduces our featureset in a small way,
and encourages other servers to move to a better design. The SSH protocol is
different enough that compressed-key-material attacks like BEAST are
unlikely, but who wants to take the chance? We encourage other ssh servers
who care about optional compression support to add delayed-zlib support.
(Some already do "zlib@openssh.com") ok djm markus
Darren Tucker [Thu, 22 Aug 2024 10:36:12 +0000 (20:36 +1000)]
Move rekey test into valgrind-2.
Now that the rekey test has been optimized it's fast enough to not be in
its own valgrind test, so move it into valgrind-2, which is currently
the quickest of the others, bringing all of them to roughly the same
runtime of ~1.1 hours.
projects / thirdparty / openssh-portable.git / log
