tree: 14fb2453a3ac4b3611f249ad32c76f0273ef0769 [path history] [tgz]
  1. include/
  2. all-tests.cpp
  3. BUILD.gn
  4. crash-handler.cpp
  5. crash-handler.h
  6. crash-list.cpp
  7. crash-list.h
  8. README.md
  9. unittest-main.cpp
  10. unittest.cpp
  11. watchdog.cpp
  12. watchdog.h
zircon/system/ulib/unittest/README.md

ulib/unittest

This directory contains a harness, named Unittest, for writing tests used by system/utest.

N.B. This library cannot use fdio since system/utest/core uses it and system/utest/core cannot use fdio. See system/utest/core/README.md.

If you want your unit tests to print to standard out, you must link your test executable to system/ulib/fdio!

Rules for use of printf vs unittest_printf

Tests are expected to not call printf(). By default we want tests to be less verbose, and use of printfs about means the output cannot not be disabled. This test harness can call printf(), but tests should not. Instead tests are expected to either call unittest_printf() or unitest_printf_critical(), the difference being that the former is controlled by the verbosity level, and the latter is not. Obviously unittest_printf_critical() is intended to be used sparingly.

Test verbosity

Verbosity is controlled by passing v=N when invoking the test, where N has the range 0-9. The default is zero, which means unittest_printf() output does not appear. Any value above zero enables unittest_printf() output.

Rules for parsing argv

Unittest has a set of options that it recognizes. All tests are expected to call unittest_run_all_tests(), which will ensure all tests get these options. The library supplies a default main() function that does this. So if a test has no interest in the arguments itself and needs no other special global initialization, it need not define its own main at all.

However, tests can also have their own options. Since Unittest does not use any kind of general argv parsing library, and each test as well as Unittest do their own parsing, one issue is how to support both Unittest options and test-specific options without either having to know about the other. This becomes important when parsing an option that takes a value and the value might begin with “-”. E.g.,

$ foo-test --foo -f -f bar

where the first -f is the value for option --foo and the second -f is an option specific to foo-test.

Argv processing is first done in the main() of the testcase, and then again in Unittest when the testcase calls unittest_run_all_tests(). If --foo is a Unittest option, how does the testcase know to ignore the first -f? The solution we employ is very simple:

Parse argv one element at a time, and ignore anything that is not recognized.

This simple rule makes writing tests easy, but it does have some consequences one needs to be aware of. For example, this means that option values cannot begin with “-”, which makes the above example invalid. A second consequence is that there are no positional parameters. E.g.,

$ foo-test --foo ./-f a b c -f bar

is equivalent to

$ foo-test --foo ./-f -f bar a b c

While not entirely clean, this allows for a simple implementation, and preserves the status quo.