On Tue, Oct 2, 2018 at 1:19 AM Dmitry Vyukov <dvyu...@google.com> wrote: > > On Tue, Oct 2, 2018 at 7:49 AM, Eric Dumazet <eduma...@google.com> wrote: > > > Does inet_frag_kill() hold fq->lock? I am missing how inet_frag_kill() > and inet_frags_exit_net() are synchronized. > Since you use smp_store_release()/READ_ONCE() they seem to run in > parallel. But then isn't it possible that inet_frag_kill() reads > nf->dead == 0, then inet_frags_exit_net() sets nf->dead, and then we > have the same race on concurrent removal? Or, isn't it possible that > inet_frag_kill() reads nf->dead == 1, but does not set > INET_FRAG_HASH_DEAD yet, and then inet_frags_free_cb() misses the > INET_FRAG_HASH_DEAD flag? >
Yes this is kind of implied in my patch. I put the smp_store_release() and READ_ONCE exactly to document the possible races. This was the reason for my attempt in V1, doing a walk, but Herbert said walk was not designed for doing deletes. Proper synch will need a synchronize_rcu(), and thus a future conversion in net-next because we can not really add new synchronize_rcu() calls in an (struct pernet_operations.)exit() without considerable performance hit of netns dismantles. So this will require a conversion of all inet_frags_exit_net() callers to .exit_batch() to mitigate the cost. I thought of synchronize_rcu_bh() but this beast is going away soon anyway.
