On Wed, May 29, 2019 at 7:40 AM Herbert Xu <herb...@gondor.apana.org.au> wrote:
>
> On Tue, May 28, 2019 at 06:31:00AM -0700, Eric Dumazet wrote:
> >
> > This smp_store_release() is a left over of the first version of the patch,
> > where
> > there was no rcu grace period enforcement.
> >
> > I do not believe there is harm letting this, but if you disagree
> > please send a patch ;)
>
> I see now that it is actually relying on the barrier/locking
> semantics of call_rcu vs. rcu_read_lock. So the smp_store_release
> and READ_ONCE are simply unnecessary and could be confusing to
> future readers.
>
> ---8<---
> The smp_store_release call in fqdir_exit cannot protect the setting
> of fqdir->dead as claimed because its memory barrier is only
> guaranteed to be one-way and the barrier precedes the setting of
> fqdir->dead.
>
> IOW it doesn't provide any barriers between fq->dir and the following
> hash table destruction.
>
> In fact, the code is safe anyway because call_rcu does provide both
> the memory barrier as well as a guarantee that when the destruction
> work starts executing all RCU readers will see the updated value for
> fqdir->dead.
>
> Therefore this patch removes the unnecessary smp_store_release call
> as well as the corresponding READ_ONCE on the read-side in order to
> not confuse future readers of this code. Comments have been added
> in their places.
>
> Signed-off-by: Herbert Xu <herb...@gondor.apana.org.au>
>
> diff --git a/net/ipv4/inet_fragment.c b/net/ipv4/inet_fragment.c
> index 2b816f1ebbb4..35e9784fab4e 100644
> --- a/net/ipv4/inet_fragment.c
> +++ b/net/ipv4/inet_fragment.c
> @@ -193,10 +193,12 @@ void fqdir_exit(struct fqdir *fqdir)
> {
> fqdir->high_thresh = 0; /* prevent creation of new frags */
>
> - /* paired with READ_ONCE() in inet_frag_kill() :
> - * We want to prevent rhashtable_remove_fast() calls
> + fqdir->dead = true;
> +
> + /* call_rcu is supposed to provide memory barrier semantics,
> + * separating the setting of fqdir->dead with the destruction
> + * work. This implicit barrier is paired with inet_frag_kill().
> */
> - smp_store_release(&fqdir->dead, true);
>
> INIT_RCU_WORK(&fqdir->destroy_rwork, fqdir_rwork_fn);
> queue_rcu_work(system_wq, &fqdir->destroy_rwork);
> @@ -214,10 +216,12 @@ void inet_frag_kill(struct inet_frag_queue *fq)
>
> fq->flags |= INET_FRAG_COMPLETE;
> rcu_read_lock();
> - /* This READ_ONCE() is paired with smp_store_release()
> - * in inet_frags_exit_net().
> + /* The RCU read lock provides a memory barrier
> + * guaranteeing that if fqdir->dead is false then
> + * the hash table destruction will not start until
> + * after we unlock. Paired with inet_frags_exit_net().
> */
> - if (!READ_ONCE(fqdir->dead)) {
> + if (!fqdir->dead) {
If fqdir->dead read/write are concurrent, then this still needs to be
READ_ONCE/WRITE_ONCE. Ordering is orthogonal to atomicity.
> rhashtable_remove_fast(&fqdir->rhashtable, &fq->node,
> fqdir->f->rhash_params);
> refcount_dec(&fq->refcnt);
