> + child_tid = atomic_fetch_or(&mgr->managed_tid, 0);
> + /*
> + * Check if the child has already terminated by this point. If not, wait
> + * for the child to exit. As long as the trampoline is not killed by
> + * a signal, the kernel guarantees that the memory at &mgr->managed_tid
> + * will be cleared, and a FUTEX_WAKE at that address will triggered.
> + */
> + if (child_tid != 0) {
> + ret = syscall(SYS_futex, &mgr->managed_tid, FUTEX_WAIT,
> + child_tid, NULL, NULL, 0);
> + assert(ret == 0 && "clone manager futex should always succeed");
> + }
A note for any reviewers/maintainers: While doing some additional
testing today, I discovered there is a bug in this section of the
patch. The child process can exit between the `atomic_fetch` and start
of the `futex(FUTEX_WAIT)` call, causing the kernel to respond with an
`EAGAIN` error, which will be caught by the assert and crash the
program. I have a patch for this. I suspect there will be comments on
this change, so I'm holding off on re-sending the series until initial
reviews have been done. I just wanted to make maintainers aware to
avoid the possibility of this bug being merged in the (very) unlikely
case there are no comments.
