On 10/7/21 4:22 AM, David Ahern wrote:
On 10/6/21 3:35 PM, Leonard Crestez wrote:
I counted the [FAIL] or [ OK ] markers but not the output of nettest
itself. I don't know what to look for, I guess I could diff the outputs?
Shouldn't it be sufficient to compare the exit codes of the nettest client?
mistakes happen. The 700+ tests that exist were verified by me when I
submitted the script - that each test passes when it should and fails
when it should. "FAIL" has many reasons. I tried to have separate exit
codes for nettest.c to capture the timeouts vs ECONNREFUSED, etc., but I
could easily have made a mistake. scanning the output is the best way.
Most of the 'supposed to fail' tests have a HINT saying why it should fail.
It is not good to have a test for which correctness is ambiguous to such
an extent, it makes reliable future changes difficult. In theory an
uniform TAP format is supposed to solve this but it is not applied
inside selftests/net.
I attempted to write a script to compare two logs in their current
format: https://gitlab.com/cdleonard/kselftest-parse-nettest-fcnal
It does a bunch of nasty scrubbing of irrelevant behavior and got it to
the point where no diffs are found between repeated runs on the same
machine.
One nasty issue was that many tests kill processes inside log_test so
relevant output may be shown either before or after the "TEST: " result
line. This was solved by associating output until the next ### with
previous test.
The output is also modified by a previous change to not capture server
output separately and instead let it be combined with that of the
client. That change is required for this one, doing out=$(nettest -k)
does not return on fork unless the pipe is also closed.
I did not look at your change, mine is relatively minimal because it
only changes who decide when the server goes into the background: the
shell script or the server itself. This makes it work very easily even
for tests with multiple server instances.
The logging issue is why I went with 1 binary do both server and client
after nettest.c got support for changing namespaces.
It's possible to just compare the "client" and "server" logs separately
by sorting them on their prefix.
I think a decent approach would be to do a bulk replace for all
"run_cmd{,_nsb,_nsc} nettest" with a new "run_nettest" function that
passes all arguments to nettest itself. That run_nettest function could
include a leading common "-t" arg that is parsed at the top of
fcnal-test.sh.
--
Regards,
Leonard
