Alex Chernyakhovsky
325098ac07
Previously, mosh used extensive -I flags and all of the mosh-local makes it really hard to tell what the proper dependency graph is, so instead remove the -I arguments in favvor of $(top_srcdir) and qualify the paths wherever they are used.
Mosh Tests
ocb-aes
This is a unit test for the OCB-AES encryption used in mosh, including Rogaway's OCB implementation and some of mosh's surrounding C++ support code.
encrypt-decrypt
This is a simple functional test of mosh's implementation of encrypted messages.
base64
This tests Mosh's homegrown base64 functionality. The associated
genbase64.pl script is used to independently generate validated test
vectors.
e2e-test
This is a test framework for end-to-end testing of mosh. It uses tmux
to invoke mosh in a nicely stable interactive pty, and also uses
tmux's capture-pane command to get a dump of the terminal screen
that mosh-client has drawn, neatly getting around Mosh's somewhat
non-deterministic display redraw.
There are four essential parts to the framework:
- your test script
e2e-teste2e-test-servere2e-test-subrs
The test script has two roles: when invoked without arguments, it is a wrapper script for the overall test, and when invoked with an argument, it performs a testing-related action. In wrapper mode, it invokes e2e-test with action arguments, which are used to invoke the test script for actions at appropriate points by e2e-test. These provide a suite of behaviors that you can use to test various mosh behaviors.
e2e-test is the heart of the framework. It runs actions as
requested, logs their output, compares and/or validates their results,
and generates the final result (exitstatus, mostly) for the Automake
testing framework used by the mosh build. For test execution, it runs
an action in an interactive session, in a tmux screen, to exercise
some behavior. The action can optionally be run in a mosh session, or
directly in tmux (doing both and comparing the result is a useful way
to test complex terminal emulation behaviors). The action generally
writes some output to the terminal that can later be verified by
another action. Optionally, a client action can generate tty input or
otherwise exercise mosh in some fashion (this capability is untested,
but it's a useful place to use expect or other interactive
simulations). The action is run by e2e-test-server, which is a
relatively small wrapper script to capture errors, and capture the
tmux screen.
There are several different categories of actions:
Execution
baseline is an action that almost all tests will use. This invokes
the test script inside mosh, where it can generate some output, and
then captures the client-side tmux display with tmux capture-pane.
direct is the same as the above, except that mosh is not used--
e2e-wrapper-script and the test script are invoked directly inside
tmux.
variant can be used to provide a slightly different action from
baseline.
Verification
verify compares captures from the baseline and direct test
actions, which are expected to be identical.
same compares captures from the baseline and variant test
actions, which are expected to be identical.
different compares captures from the baseline and variant test
actions, which are expected to be different.
post is a catchall script hook which allows custom verification
acions to be coded.
Client wrappers
tmux injects a wrapper command into the test command before tmux.
If this is not run, a default command called hold-stdin is run
instead. These commands are expected to hold tmux's stdin open,
possibly injecting tmux commands, while the test runs. See
window-resize.test for an example of this that manipulates tmux
state. Alternately, this could use expect or something similar.
client simply injects a wrapper command into the (long) test command
between tmux and mosh. It's expected to interact with its wrapped
command line as expect might do. This is not actually tested yet.
Flags
mosh-args, client-args and server-args inject extra arguments
into the invocations of the respective commands.
Logging and error reporting
Each execution action is run, and recorded in
<testname>.test.d/<action>.*. <action>.exitstatus is the
exitstatus from the server wrapper. <action>.tmux.log is the output
of tmux for the entire test run for that action; <action>.capture is
a capture of the Mosh client screen after the test action is complete,
generated with tmux capture-pane.
In accordance with GNU Automake's test framework, the test should return these exit status values:
- 0 test success
- 1 test failure
- 77 test skipped (tmux or ssh is unavailable if needed)
- 99 hard error
These values are also used internally between the various scripts; errors are conveyed out to the build test framework.
Sample tests
A few tests have been implemented so far to test the framework itself, and to provide examples for further development.
e2e-success is a simple test that executes baseline and direct
with the same stimulus (simply clearing the screen), and expects to
see identical results.
e2e-failure is similar to e2e-success, but expects to see
different results from baseline and variant. Since it uses the
same stimulus for the two execution action, it fails. A more
realistic test might be to have variant execute some escape sequence
that is absent from baseline; this would verify that the escape
sequence actually does something.
emulation-back-tab tests an escape sequence that mosh does not
support. It expects the test to produce the output that would be
generated if the escape sequence were implemented. If it gets output
as expected when the escape sequence is not implemented, the test
fails. But if the output does not match one of these two cases, the
test returns an error. This is an example of error handling within
the test framework.
unicode-later-combining demonstrates mosh's handling of a Unicode
edge case, a combining character drawn without a printing character in
the same cell. It verifies the output in the post action; since
there are a couple of different Unicode renderings that are reasonable
in this case, a regex that covers both is used. It also implements an
unused variant action that draws blank-space+combiner in a correct
fashion.
Notes
The shell command printf is generally used in place of
echo in this framework, because of its more precisely-specified and
portable behavior. But beware, even printf varies between systems--
GNU printf, for example, implements \e, which is a non-POSIX
extension unavailable in BSD implementations
It's fairly simple to test each of these scripts independently, but
the entire chain is a bit prone to behaving oddly in hard-to-debug
ways. set -x is your friend here.
The test scripts are a bit fragile about timeouts. They will
generally run correctly on an unloaded machine without the make -j
flag. Using make -j is obviously very convenient for development,
and it works fine on faster machines, but I don't recommend it for
automated testing.