[LTP] [PATCH 2/2] lib: Add test library design document

Cyril Hrubis chrubis@suse.cz
Fri Nov 27 17:31:50 CET 2020


Which tries to explain high level overview and design choices for the
test library.

Signed-off-by: Cyril Hrubis <chrubis@suse.cz>
---
 lib/README.md | 130 ++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 130 insertions(+)
 create mode 100644 lib/README.md

diff --git a/lib/README.md b/lib/README.md
new file mode 100644
index 000000000..6efd3cf33
--- /dev/null
+++ b/lib/README.md
@@ -0,0 +1,130 @@
+# Test library design document
+
+## Test lifetime overview
+
+When a test is executed the very first thing to happen is that the we check for
+various test pre-requisities. These are described in the tst\_test structure
+and range from simple '.require\_root' to a more complicated kernel .config
+boolean expressions such as:
+"CONFIG\_X86\_INTEL\_UMIP=y | CONFIG\_X86\_UMIP=y".
+
+If all checks are passed the process carries on with setting up the test
+environment as requested in the tst\_test structure. There are many different
+setup steps that have been put into the test library again ranging from rather
+simple creation of a unique test temporary directory to a bit more complicated
+ones such as preparing, formatting, and mounting a block device.
+
+The test library also intializes shrared memory used for IPC at this step.
+
+Once all the prerequisities are checked and test environment has been prepared
+we can move on executing the testcase itself. The actual test is executed in a
+forked process, however there are a few hops before we get there.
+
+First of all there are test variants, which means that the test is re-executed
+several times with a slightly different settings. This is usually used to test
+a family of similar syscalls, where we test each of these syscalls exactly the
+same, but without re-executing the test binary itself. Test varianst are
+implemented as a simple global variable counter that gets increased on each
+iteration. In a case of syscall tests we switch between which syscall to call
+based on the global counter.
+
+Then there is all\_filesystems flag which is mostly the same as test variants
+but executes the test for each filesystem supported by the system. Note that we
+can get cartesian product between test variants and all filesystems as well.
+
+In a pseoudo code it could be expressed as:
+
+```
+for test_variants:
+	for all_filesystems:
+		fork_testrun()
+```
+
+Before we fork() the test process the test library sets up a timeout alarm and
+also a heartbeat signal handlers and also sets up an alarm(2) accordingly to
+the test timeout. When a test timeouts the test library gets SIGALRM and the
+alarm handler mercilesly kills all forked children by sending SIGKILL to the
+whole process group. The heartbeat handler is used by the test process to reset
+this timer for example when the test functions runs in a loop.
+
+With that done we finally fork() the test process. The test process firstly
+resets signal handlers and sets its pid to be a process group leader so that we
+can slaughter all children if needed. The test library proceeds with suspending
+itself in waitpid() syscall and waits for the child to finish at this point.
+
+The test process goes ahead and call the test setup() function if present in
+the tst\_test structure. It's important that we execute all test callbacks
+after we have forked the process, that way we cannot crash the test library
+process. The setup can also cause the the test to exit prematurely by either
+direct or indirect (SAFE\_MACROS()) call to tst\_brk().  In this case the
+fork\_testrun() function exits, but the loops for test variants or filesystems
+carries on.
+
+All that is left to be done is to actually execute the tests, what happnes now
+depends on the -i and -I command line parameters that can request that the
+run() or run\_all() callbacks are executed N times or for a N seconds. Again
+the test can exit at any time by direct or indirect call to tst\_brk().
+
+Once the test is finished all that is left for the test process is the test
+cleanup(). So if a there is a cleanup() callback in the tst\_test strucuture
+it's executed. Callback runs in a special context where the tst\_brk(TBROK,
+...) calls are converted into tst\_res(TWARN, ...) calls. This is because we
+found out that carrying up with partially broken cleanup is usually better
+option than exitting it in the middle.
+
+The test cleanup() is also called by the tst\_brk() handler in order to cleanup
+before exitting the test process, hence it must be able to cope even with
+partiall test setup. Usually it suffices to make sure to clean up only
+resources that already have been set up and to do that in an inverse order that
+we did in setup().
+
+Once the test process exits or leaves the run() or run\_all() function the test
+library wakes up from the waitpid() call, and checks if the test process
+exitted normally.
+
+Once the testrun is finished the test library does a cleanup() as well to clean
+up resources set up in the test library setup(), reports test results and
+finally exits the process.
+
+### Test library and fork()-ing
+
+Things are a bit more complicated when fork()-ing is involved, however the
+tests results are stored in a page of a shared memory and incremented by atomic
+operations, hence the results are stored rigth after the test reporting
+fucntion returns from the test library and the access is, by definition,
+race-free as well.
+
+On the other hand the test library, apart from sending a SIGKILL to the whole
+process group on timeout, does not track granchildren.
+
+This especially means that:
+
+- The test exits once the main test process exits.
+
+- While the test results are, by the desing, propagated to the test library
+  we may still miss a child that gets killed by a signal or exits unexpectedly.
+
+The test writer should, because of these, take care for mourning these proceses
+properly, in most cases this could be simply done by calling
+tst\_reap\_children() to collect and dissect deceased.
+
+Also note that tst\_brk() does exit only the current process, so if child
+process calls tst\_brk() the counters are incremented and the process exits.
+
+### Test library and exec()
+
+The piece of mapped memory to store the results to is not preserved over
+exec(2), hence to use the test library from a binary started by an exec() it
+has to be remaped. In this case the process must to call tst\_reinit() before
+calling any other library functions. In order to make this happen the program
+environment carries LTP\_IPC\_PATH variable with a path to the backing file on
+tmpfs. This also allows us to use the test library from a shell testcases.
+
+### Test library and process synchronization
+
+The piece of mapped memory is also used as a base for a futex-based
+synchronization primitives called checkpoints. And as said previously the
+memory can be mapped to any process by calling the tst\_reinit() function. As a
+matter of a fact there is even a tst\_checkpoint binary that allows use to use
+the checkpoints from shell code as well.
+
-- 
2.26.2



More information about the ltp mailing list