Memory Leak in Comm::wait_any_for() when killing the waiting actor during the wait

[henricasanova]

I am not sure this can be fixed in any way, but I have a simulation in which actors do wait_any_for() a lot, and also kill each other brutally a lot. As a result, in wait_any_for() the dynar allocated as

xbt_dynar_t comms = xbt_dynar_new(sizeof(simgrid::kernel::activity::ActivityImpl*), [](void* ptr) {
intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
});

leaks if the actor is killed why it is waiting (the dynar is freed only at the end of the wait_any_for() code).

Again, this seems difficult to fix and after all killing actors brutally may not be good design, but I thought I'd mention it just in case Mt can do some magic ;)

[mquinson]

I would not be surprised if this was fixed in between. The line now reads

  std::unique_ptr<kernel::activity::CommImpl* []> rcomms(new kernel::activity::CommImpl*[comms->size()]);

so I think that RAII is now allowing to free the resource when the exception is used to (1) kill the actor (2) clean all data allocated on the stack.

@henricasanova if you have a easy way to reproduce this bug (even if not a MWE), could you please check whether I am right? Thanks.

[mquinson]

Hello @henricasanova, if you get bored in the future, could you please check whether the issue got fixed in between, as I think? If not, could we maybe close the bug since we fail to reproduce and think it's fixed anyway?

Thanks,

[henricasanova]

Sorry to have missed this. It's been 3 years now that I've posted this. I think I was able to reproduce it using an old MWE about this, and valgrind confirms that there are no leaks. So I believe it's safe to close. Doing so right now.

[mquinson]

Thanks!