unittests stuck in test_ocxl_afu_attach()

[sharkcz]

I have a problem with the unittests, it gets stuck in test_ocxl_afu_attach() (almost) every time. See the output below. The environment is libocxl from github master, Fedora 28 with kernel-5.1.0-0.rc5.git0.1.fc31.op.1.ppc64le, HW is Power9 Talos II bare metal.

[dan@talos libocxl]$ make test V=1                                                                                                                                                     :master
sudo testobj/unittests
Starting test AFU:init
OK
Starting test AFU:ocxl_afu_alloc
OK
Starting test AFU:device_matches
OK
Starting test AFU:populate_metadata
OK
Starting test AFU:getters
   unique: 140735475091664, success, outsize: 24
OK
Starting test AFU:get_afu_by_path
OK
Starting test AFU:afu_open
   unique: 140735475091664, success, outsize: 24
OK
Starting test AFU:ocxl_afu_open_specific
OK
Starting test AFU:ocxl_afu_open_from_dev
OK
Starting test AFU:ocxl_afu_open
OK
Starting test AFU:ocxl_afu_attach
   unique: 140735475089536, success, outsize: 24

^C^C
make: *** [Makefile:84: test] Zabit (SIGKILL)

output of ps shows

...
11377 pts/1    S+     0:00 sudo testobj/unittests
11379 pts/1    Dl+    0:00 testobj/unittests
...

and trying to attach either strace or gdb leads nowhere, they get stuck as well.

[deece]

The use of fuse makes these tests flakey as
best, there is also a race condition in fuse that i have not been able to work around that I suspect you are hitting.

I would suggest testing with real hardware (like the memcpy AFU) instead

[sharkcz]

Thanks for info, I too suspected fuse ... I'll see what I can do, my goal is to have some basic testing for distro builds, so relying on real hw might be difficult.

[ajdlinux]

@deece we need to document that

[deece]

I think a better choice would be to disable any subtests that are subject to the races

[deece]

Hmm, it's not that simple, the tests work the first time through, then hang in the open call on subsequent runs.

rmmod cuse before the run fixes the problem, but adding it to the script may mess up other users of cuse.