prov/efa: Set g_device_list to NULL after free

[zachdworkin]

fi_info has a double free issue. The efa g_device_list isn't being set to NULL after being freed so the next time efa_device_list_finalize is called it will be freed again (double free). Setting it to NULL removes this.

[shijin-aws]

@zachdworkin Is it a new issue after b7b9dd6?

[zachdworkin]

Yes its a new issue after that commit. I rebased before finding this bug. The way I found it was running the fi_info binary with any provider and saw the double free report. Libfabric must be configured so that efa is built (HAVE_EFA 1 in config.log) to reproduce

[shijin-aws]

@zachdworkin Thanks for catching it ! Let me run with ASAN to reproduce

[zachdworkin]

@shijin-aws no problem! Seems simple enough. What does ASAN stand for?

[shijin-aws]

ASAN is address sanitizer that we used to catch illegal memory access and double free error. But I couldn't reproduce the double free report: fi_info -p tcp just runs ok... how do you get that double free report?

[shijin-aws]

I think I can see where the bug is introduced:

Before b7b9dd6, efa_prov is never returned in the error path, so EFA_FIN is never called . But now efa_prov is returned anyway and EFA_FIN will be called in the process reaper again. So if g_device_list is not initialized and reset as NULL, EFA_FIN can see a garbage efa_device_list and double free it.

[zachdworkin]

steps to reproduce:
git reset --hard upstream/main
make distclean
./autogen.sh
./configure --prefix=install/location
make -j install

(check HAVE_EFA 1 in config.log)

install/location/bin/fi_info -p verbs
double free error

if run with -p efa, shm, tcp, etc the error still appears and using gdb to follow it leads to the changes in this patch

[shijin-aws]

I think this PR is only a bandaid as we have more clean up issue here - The EFA_INI code block needs to move the initialization code to the getinfo. I need to think about whether I need to revert that commit now before i can have a good fix shortly

[shijin-aws]

@zachdworkin can you try if #10807 fixes the issue for you?

[zachdworkin]

@shijin-aws #10807 fixes it. Should I close this and let that PR be the one to fix it?

[shijin-aws]

Yes, you can close this one