Catch and report unhandled exceptions before they propagate to C land
The C++ interface to run test cases goes like this:
1) C++ run function -> 2) C run function -> 3) C++ wrapper for
test case -> 4) test case head/body/cleanup
The previous code caught and reported unhandled exceptions in 1).
However, such approach does not seem to work everywhere. It fails,
for example, in NetBSD/i386 but works in NetBSD/amd64. I am not sure
which platform implementation is correct nor if there even _is_ a
defined behavior. No matter what, it feels wrong and clunky.
Move the last-resort exception catching to happen in 3) so that
exceptions don't propagate back to C.
Fixes the test-programs/result_test:result_exception test case in
NetBSD/i386 5.99.34.
Miscellaneous features
* Added expected failures support to test cases and atf-run. These
include, for example, expected clean exits, expected reception of fatal
signals, expected timeouts and expected errors in condition checks.
These statuses can be used to denote test cases that are known to fail
due to a bug in the code they are testing. atf-report reports these
tests separately but they do not count towards the failed test cases
amount.
* Added the ATF_CHECK_ERRNO and ATF_REQUIRE_ERRNO to the C library to
allow easy checking of call failures that update errno.
* Added the has.cleanup meta-data property to test caes that specifies
whether the test case has a cleanup routine or not; its value is
automatically set. This property is read by atf-run to know if it has to
run the cleanup routine; skipping this run for every test case
significantly speeds up the run time of test suites.
* Reversed the order of the ATF_CHECK_THROW macro in the C++ binding to
take the expected exception as the first argument and the statement to
execute as the second argument.
Changes in atf-check
* Changed atf-check to support negating the status and output checks by
prefixing them with not- and added support to specify multiple checkers
for stdout and stderr, not only one.
* Added the match output checker to atf-check to look for regular
expressions in the stdout and stderr of commands.
* Modified the exit checks in atf-check to support checking for the
reception of signals.
Code simplifications and cleanups
* Removed usage messages from test programs to simplify the
implementation of every binding by a significant amount. They just now
refer the user to the appropriate manual page and do not attempt to wrap
lines on terminal boundaries. Test programs are not supposed to be run
by users directly so this minor interface regression is not important.
* Removed the atf-format internal utility, which is unused after the
change documented above.
* Removed the atf-cleanup internal utility. It has been unused since the
test case isolation was moved to atf-run in 0.8
* Splitted the Makefile.am into smaller files for easier maintenance and
dropped the use of M4. Only affects users building from the repository
sources.
* Intermixed tests with the source files in the source tree to provide
them more visibility and easier access. The tests directory is gone from
the source tree and tests are now suffixed by _test, not prefixed by t_.
* Simplifications to the atf-c library: removed the io, tcr and ui
modules as they had become unnecessary after all simplifications
introduced since the 0.8 release.
* Removed the application/X-atf-tcr format introduced in 0.8 release.
Tests now print a much simplified format that is easy to parse and nicer
to read by end users. As a side effect, the default for test cases is
now to print their results to stdout unless otherwise stated by providing
the -r flag.
* Removed XML distribution documents and replaced them with plain-text
documents. They provided little value and introduced a lot of complexity
to the build system.
* Simplified the output of atf-version by not attempting to print a
revision number when building form a distfile. Makes the build system
easier to maintain.
which are known to fail, e.g.:
atf_tc_set_md_var(tc, "xfail", "PR kern/43456");
Expected failures do not count towards the ultimate pass/fail result
from the test run:
pain-rustique:39:~/<2>src/tests/fs/ptyfs> atf-run t_nullpts | atf-report
Tests root: /home/pooka/src/wholesrc2/src/tests/fs/ptyfs
t_nullpts (1/1): 1 test cases
nullrevoke: Expected failure: PR kern/43456
Summary for 1 test programs:
0 passed test cases.
0 failed test cases.
1 expected failures.
0 skipped test cases.
pain-rustique:40:~/<2>src/tests/fs/ptyfs> echo $?
0
However, an xfail test which passes will count as a failure, i.e.
xfail inverts test case success/fail. This way we can get a better
sense from the ultimate verdict of the NetBSD atf run by seeing if
there were any unexpected failures, i.e. new regressions.
This feature will be present in the upcoming atf 0.10 release,
possibly with finer grained control.
patch reviewed by jmmv
----------
Fix detection of crashed test cases
Prevent cross-test-case contamination that occured when a first test case
passed and the second crashed. The second could pick up the result of the
first test case and not be reported as failed.
Similarly, change the way timed out test cases are reported back to the
caller. The creation of a temporary results file was just a really stupid
way of passing information around and introduced false positives if the
test case creates a results file before timing out.
Fixes ticket #35.
----------
Problem originally reported by pooka@.
* Added atf-sh, an interpreter to process test programs written using
the shell API. This is not really a shell interpreter by itself
though: it is just a wrapper around the system shell that eases the
loading of the necessary ATF libraries.
* Removed atf-compile in favour of atf-sh.
* Added the use.fs metadata property to test case, which is used to
specify which test cases require file system access. This is to
highlight dependencies on external resources more clearly and to speed
up the execution of test suites by skipping the creation of many
unnecessary work directories.
* Fixed test programs to get a sane default value for their source
directory. This means that it should not be necessary any more to pass
-s when running test programs that do not live in the current
directory.
* Defining test case headers became optional. This is trivial to achieve
in shell-based tests but a bit ugly in C and C++. In C, use the new
ATF_TC_WITHOUT_HEAD macro to define the test case, and in C++ use
ATF_TEST_CASE_WITHOUT_HEAD.
* Test programs no longer run several test cases in a row. The execution
of a test program now requires a test case name, and that single test
case is executed. To execute several test cases, use the atf-run
utility as usual.
* Test programs no longer fork a subprocess to isolate the execution of
test cases. They run the test case code in-process, and a crash of the
test case will result in a crash of the test program. This is to ease
debugging of faulty test cases.
* Test programs no longer isolate their test cases. This means that they
will not create temporary directories nor sanitize the environment any
more. Yes: running a test case that depends on system state by hand
will most likely yield different results depending on where (machine,
directory, user environment, etc.) it is run. Isolation has been moved
to atf-run.
* Test programs no longer print a cryptic format (application/X-atf-tcs)
on a special file channel. They can now print whatever they want on
the screen. Because test programs can now only run one test case every
time, providing controlled output is not necessary any more.
* Test programs no longer write their status into a special file
descriptor. Instead, they create a file with the results, which is
later parsed by atf-run. This changes the semantics of the -r flag.
* atf-run has been adjusted to perform the test case isolation. As a
result, there is now a single canonical place that implements the
isolation of test caes. In previous releases, the three language
bindings (C, C++ and shell) had to be kept in sync with each other
(read: not a nice thing to do at all). As a side effect of this
change, writing bindings for other languages will be much, much easier
from now on.
* atf-run forks test programs on a test case basis, instead of on a test
program basis as it did before. This is to provide the test case
isolation that was before implemented by the test programs themselves.
* Removed the atf-exec tool. This was used to implement test case
isolation in atf-sh, but it is now unnecessary.
* It is now optional to define the descr meta-data property. It has been
proven to be mostly useless, because test cases often carry a
descriptive name of their own.
* Added build-time checks to atf-c and atf-c++. A binding for atf-sh
will come later.
* Migrated all build-time checks for header files to proper ATF tests.
This demonstrates the use of the new feature described above.
* Added an internal API for child process management.
* Converted all plain-text distribution documents to a Docbook canonical
version, and include pre-generated plain text and HTML copies in the
distribution file.
* Simplified the contents of the Makefile.am by regenerating it from a
canonical Makefile.am.m4 source. As a side-effect, some dependency
specifications were fixed.
* Migrated all checks from the check target to installcheck, as these
require ATF to be installed.
* Fixed sign comparison mismatches triggered by the now-enabled
-Wsign-compare.
* Fixed many memory and object leaks.
Changes in this release:
* Make atf-exec be able to kill its child process after a certain period of
time; this is controlled through the new -t option.
* Change atf-sh to use atf-exec's -t option to control the test case's
timeouts, instead of doing it internally. Same behavior as before, but
noticeably faster.
* atf-exec's -g option and atf-killpg are gone due to the previous change.
* Added the atf-check(1) tool, a program that executes a given command and
checks its exit code against a known value and allows the management of
stdout and stderr in multiple ways. This replaces the previous atf_check
function in the atf-sh library and exposes this functionality to both
atf-c and atf-c++.
* Added the ATF_REQUIRE family of macros to the C interface. These help
in checking for fatal test conditions. The old ATF_CHECK macros now
perform non-fatal checks only. I.e. by using ATF_CHECK, the test case
can now continue its execution and the failures will not be reported
until the end of the whole run.
* Extended the amount of ATF_CHECK_* C macros with new ones to provide more
features to the developer. These also have their corresponding
counterparts in the ATF_REQUIRE_* family. The new macros (listing the
suffixes only) are: _EQ (replaces _EQUAL), _EQ_MSG, _STREQ and
_STREQ_MSG.
