Should probably have been an obvious side effect :-)
Child process gets open file descriptors to parent's epoll/timerfd,
and it's totally OK to just close() all of them, but it's absolutely NOT
OK to run destructors - they modify the kernel state of epoll/timerfd
before destroying. So, basically, when we destroy the client in the child
process, we break it in the parent too. This also means that cluster_client_t
doesn't support fork(). :-)
Do all NBD configuration in the child process, after the last fork.
Why? It's needed because there is a race condition in the Linux kernel nbd driver
in nbd_add_socket() - it saves `current` task pointer as `nbd->task_setup` and
then rechecks if the new `current` is the same. Problem is that if that process
is already dead, `current` may be freed and then replaced by another process
with the same pointer value. So the check passes and NBD allows a different process
to set up a device which is already set up. Proper fix would have to be done in the
kernel code, but the workaround is obviously to perform NBD setup from the process
which will then actually call NBD_DO_IT. That process stays alive during the whole
time of NBD device execution and the (nbd->task_setup != current) check always
works correctly, and we don't accidentally break previous NBD devices while setting
up a new device. Forking to check every device is of course rather slow, so we also
do an additional check by calling list_mapped() before searching for a free NBD device.
Vitaliy Filippov