platform_system_core/cli-test
Elliott Hughes f276140d0f cli-test: a tool for testing command-line programs.
Not looking for other users right now, this is just enough to test
unzip/zip/zipinfo.

This includes tests for unzip and ziptool, along with a change to
unzip's behavior to fix AOSP `make dist` when using ziptool unzip.

Also add the boilerplate to run these tests on the device, in presubmit.

Fix command name in --help output.

Test: atest ziptool-tests
Change-Id: I5c0215a3ab8cb2cd5fc517ed9c188f81a7bf4520
2019-12-13 12:23:51 -08:00
..
.clang-format cli-test: a tool for testing command-line programs. 2019-12-13 12:23:51 -08:00
Android.bp cli-test: a tool for testing command-line programs. 2019-12-13 12:23:51 -08:00
cli-test.cpp cli-test: a tool for testing command-line programs. 2019-12-13 12:23:51 -08:00
README.md cli-test: a tool for testing command-line programs. 2019-12-13 12:23:51 -08:00

cli-test

What?

cli-test makes integration testing of command-line tools easier.

Goals

  • Readable syntax. Common cases should be concise, and pretty much anyone should be able to read tests even if they've never seen this tool before.

  • Minimal issues with quoting. The toybox tests -- being shell scripts -- quickly become a nightmare of quoting. Using a non ad hoc format (such as JSON) would have introduced similar but different quoting issues. A custom format, while annoying, side-steps this.

  • Sensible defaults. We expect your exit status to be 0 unless you say otherwise. We expect nothing on stderr unless you say otherwise. And so on.

  • Convention over configuration. Related to sensible defaults, we don't let you configure things that aren't absolutely necessary. So you can't keep your test data anywhere except in the files/ subdirectory of the directory containing your test, for example.

Non Goals

  • Portability. Just being able to run on Linux (host and device) is sufficient for our needs. macOS is probably easy enough if we ever need it, but Windows probably doesn't make sense.

Syntax

Any all-whitespace line, or line starting with # is ignored.

A test looks like this:

name: unzip -l
command: unzip -l $FILES/example.zip d1/d2/x.txt
after: [ ! -f d1/d2/x.txt ]
expected-stdout:
	Archive:  $FILES/example.zip
	  Length      Date    Time    Name
	---------  ---------- -----   ----
	     1024  2017-06-04 08:45   d1/d2/x.txt
	---------                     -------
	     1024                     1 file
---

The name: line names the test, and is only for human consumption.

The command: line is the command to be run. Additional commands can be supplied as zero or more before: lines (run before command:) and zero or more after: lines (run after command:). These are useful for both setup/teardown but also for testing post conditions (as in the example above).

Any command:, before:, or after: line is expected to exit with status 0. Anything else is considered a test failure.

The expected-stdout: line is followed by zero or more tab-prefixed lines that are otherwise the exact output expected from the command. (There's magic behind the scenes to rewrite the test files directory to $FILES because otherwise any path in the output would depend on the temporary directory used to run the test.)

There is currently no expected-stderr: line. Standard error is implicitly expected to be empty, and any output will cause a test failure. (The support is there, but not wired up because we haven't needed it yet.)

The fields can appear in any order, but every test must contain at least a name: line and a command: line.

Output

The output is intended to resemble gtest.

Future Directions

  • It's often useful to be able to match against stdout/stderr/a file rather than give exact expected output. We might want to add explicit support for this. In the meantime, it's possible to use an after: with grep -q if you redirect in your command:.

  • In addition to using a before: (which will fail a test), it can be useful to be able to specify tests that would cause us to skip a test. An example would be "am I running as root?".

  • It might be useful to be able to make exit status assertions other than 0?

  • There's currently no way (other than the files/ directory) to share repeated setup between tests.