The previous CI attemtped to use Homebrew for builds, but
unfortunately Homebrew has dropped support for universal packages aka
multiarch (fat) binaries. This means that in order to build an arm64 +
x86_64 package, macports has to be used instead of homebrew.
Unlike Homebrew, MacPorts is not installed by default on the GitHub
Actions runners, so we need to install it ourselves. This means
managing our own instance of the cache, which itself produces
challenges as the `gtar` binary run by the action doesn't have enough
permissions to restore the MacPorts checkout. So we have to shim gtar
with a sudo wrapper.
With this commit, we produce a Mosh package that works on macOS 11.0
and newer, on both arm64 and x86_64 architectures. The protobuf
library is statically linked, but all other libraries are provided by
the system.
This change ports the Travis CI release workflow for macOS to Github
Actions. Note that while this is functionally identical to the
previous Travis CI flow, no work has been done to update the macOS
build scripts to build for arm64.
This Github Actions workflow uses a Linux-based running to create the
release tarball for mosh. This is necessary since mosh does not check
in the autoconf/automake generated files, so the default release
action source download is missing files that are needed for
distributions that use the upstream-provided ./configure script.
Previously, ocb_internal.cc supported different key sizes, by way of
the deprecated aes_* function family. However, in practice, mosh
always uses AES-128. In Nettle, the explicit key-size APIs are not
deprecated, so switch to AES-128 directly.
Fixes: 1202
Explicitly define the primitive AES API used by the internal OCB
implementation, and move it into its own namespace (ocb_aes). This will
ease future implementation changes.
Also make some style fixes to affected lines: Replace C-style casts
with C++-style casts, add some missing spaces in argument lists, and
remove some `inline` that the compiler will ignore.
Bug: https://github.com/mobile-shell/mosh/issues/1174
bacc024083 inadvertently stopped
propagating `pkg-config --libs` output into the link line. This didn’t
affect OpenSSL (since configure.ac puts -lcrypto there manually) or
Apple Common Crypto (since it’s not a separate dylib), but it broke
Nettle builds. Fix Nettle builds by ensuring that `pkg-config --libs`
output actually makes it to the linker.
After further discussion, the Mosh maintainers have decided to stick
with the internal OCB implementation for this release. Restore support
for using OpenSSL’s AES but internal OCB. To make this commit easy to
audit, restore the code exactly, including calls to AES functions that
are deprecated in OpenSSL 3; a future commit will update ocb_internal.cc
to use EVP instead of directly calling the AES primitives.
In anticipation of future changes, preserve support for OpenSSL’s
AES-OCB, but don’t compile it in. Add
--with-crypto-library=openssl-with-openssl-ocb and
--with-crypto-library=openssl-with-internal-ocb options to configure so
that developers can easily test Mosh using OpenSSL’s AES-OCB. These
options are intended only for testing, are undocumented, and are not
subject to any API stability guarantees.
Rework configure to look for all possible cryptography libraries first
and then dispatch on --with-crypto-library as appropriate.
OpenSSL 3.0 deprecated many of the functions that ocb.cc used to
implement OCB-AES, causing a build failure when -Wdeprecated collided
with -Werror. Debian temporarily fixed this by suppressing the error
in #1191.
Since mosh 1.4 will be the next stable release of mosh, it should not
depend on deprecated functions in OpenSSL. Since version 1.1.0,
OpenSSL natively supports OCB-AES through the EVP_CIPHER API. @cgull
started early support for this in #924.
This change extends upon the previous work by @cgull in a few ways
* EVP_CipherInit_ex is called in ae_init to set up the
EVP_CIPHER_CTX. It is later called in ae_encrypt and ae_decrypt
just to load nonce (IV in OpenSSL EVP parlance), which reduces the
amount of initialization done per-packet. However, due to OpenSSL
API limitations, two copies of the EVP_CIPHER_CTX are kept: one for
encryption, and one for decryption.
* Adds missing support for an external tag, rather than just one
appended to the ciphertext
* Support for non-default-sized tags
as well as some improved error handling.
Note that this change raises the minimum OpenSSL version for Mosh to
1.1.0. OpenSSL does not provide security support for versions prior to
1.1 at this time, so this is in principle reasonable dependency. If we
want to continue to support distributions (such as RHEL7) which
continue to be supported by their vendor but use an unsupported
OpenSSL, then some future work will have to restore the ocb.cc
implementation that uses the deprecated functions.
Bugs: #1174
Split src/crypto/ocb.cc into two files – one containing the AES-OCB
implementation backed by OpenSSL, and the other containing
implementations backed by Apple Common Crypto and Nettle. This paves the
way for a new OpenSSL implementation that uses OpenSSL 1.1’s OCB support
directly, rather than one that merely uses OpenSSL to provide the
underlying block cipher.
Remove support for rijndael-alg-fst.c and compiler-provided AES
intrinsics, since they’re not in use anymore. (Mosh can still use
hardware-accelerated AES if it’s available; it just now relies
exclusively on the underlying cryptography library to accelerate AES if
possible.)
Update the build system to conditionally compile in either
ocb_openssl.cc or ocb_internal.cc, depending on which cryptography
library you pass to ./configure.
To make this commit easy to audit, ocb_openssl.cc and ocb_internal.cc
are trivially diffable against ocb.cc (now deleted). Expected diffs
consist of a copyright notice update, a preprocessor check to ensure the
appropriate cryptography implementation has been selected, and deletions
to remove code that’s no longer in use. This does mean a substantial
amount of code is duplicated between ocb_openssl.cc and ocb_internal.cc;
however, ocb_openssl.cc should be completely replaced soon, so it won’t
be an issue in the long term.
Bug: https://github.com/mobile-shell/mosh/issues/1174
This reverts commit 6321b1d9c5.
The original commit 6321b1d9c5 switched
from a malloc call of a 22400 byte buffer to a stack-allocated 22400
byte buffer, in addition to the fairly large buffers already allocated
in the functions. Some systems have fairly small stack frames, making
this 22K allocation potentially dangerous. On my stock Debian bullseye
system, I have 200809 bytes (from `getconf
_POSIX_THREAD_ATTR_STACKSIZE`); a 22400 byte buffer already represents
about 10% of the available stacksize.
Other systems, such as those with musl libc, may have either 80KiB or
128KiB [1], making this allocation represent between 18% to 28% of the
available stack space.
[1] https://wiki.musl-libc.org/functional-differences-from-glibc.html#Thread-stack-size
This commit adds the --enable-fuzzing (and --enable-asan, to make
fuzzing more useful) options and a sample fuzzer for the terminal
parser. At this time only libfuzzer is supported. Future changes to
add AFL to get more fuzzing capability should be possible with the
addition of the afl_driver.cc from Chromium.
This change adds autoconf/automake support for building all of mosh
with gcov, and generates an lcov html report. This allows seeing which
parts ofthe source tree have good test coverage, and which can be
shored up. Eventually, it would be good to hook this up to Github
Actions to be generated automatically.
sshd has a bug in which the sometimes it may get stuck trying to read from the client
even though the child process has already exited. This is visible at
https://cvsweb.openbsd.org/cgi-bin/cvsweb/src/usr.bin/ssh/serverloop.c?annotate=1.226 line 274:
once the child is waited on, `child_terminated` is reset to 0, which causes it to use an infinite timeout
in the select there.
This workaround causes mosh to disconnect from the server, thereby allowing sshd finish.
Fixes#1051, mosh fails to build on case-insensitive filesystems.
XXX This isn't perfect because autoconf/automake drop several extra
files in the directory anyway.
The uninitialized variable warning from T x; broke the test if the
user passed CXXFLAGS='-O2 -Wall -Werror'. (Users shouldn’t do that;
our own --enable-compile-warnings=error option was unaffected.)
Signed-off-by: Anders Kaseorg <andersk@mit.edu>
Change the openssl-devel package to libssl-devel. Cygwin still has an
"obsolete" (essentially virtual) package for openssl-devel, but it
doesn't seem to work properly on Appveyor's Cygwin install.
Protobuf >= 3.6.0 requires C++11, which I added support for last year.
But when I did that, I requested strict ANSI C++ compatibility, which
causes Cygwin/newlib's libc feature test macros to be set to disable
(at least) POSIX.1 extensions. Let the Autoconf macro use its default
instead (prefer GNU/extended C++, accept anything).
Statically linking mosh-server with glibc isn’t a great idea for
various reasons (nsswitch modules, locale format incompatibilities).
But we can provide most of the benefits of static linking by allowing
specific library dependencies to be linked statically using -Bstatic
and -Bdynamic. The full set is enabled by
./configure --enable-static-libraries
which is equivalent to
./configure --enable-static-libstdc++ --enable-static-libgcc \
--enable-static-utempter --enable-static-zlib --enable-static-curses \
--enable-static-crypto --enable-static-protobuf
and results in binaries whose only runtime library dependencies are
provided with libc:
$ ldd src/frontend/mosh-server
linux-vdso.so.1 (0x00007ffe0b377000)
libutil.so.1 => /lib/x86_64-linux-gnu/libutil.so.1 (0x00007fa0d9970000)
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fa0d97e3000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fa0d97c2000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fa0d95d8000)
/lib64/ld-linux-x86-64.so.2 (0x00007fa0d9f6a000)
Signed-off-by: Anders Kaseorg <andersk@mit.edu>
Alex Chernyakhovsky