[PATCH v3/3] replication: handle replication shutdown correctly.

[Vladimir Davydov]

On Wed, Aug 15, 2018 at 07:13:28PM +0300, Serge Petrenko wrote:
> diff --git a/src/box/replication.cc b/src/box/replication.cc
> index 48956d2ed..083ae6407 100644
> --- a/src/box/replication.cc
> +++ b/src/box/replication.cc
> @@ -91,13 +91,6 @@ replication_init(void)
>  	latch_create(&replicaset.applier.order_latch);
>  }
>  
> -void
> -replication_free(void)
> -{
> -	free(replicaset.replica_by_id);
> -	fiber_cond_destroy(&replicaset.applier.cond);
> -}
> -
>  void
>  replica_check_id(uint32_t replica_id)
>  {
> @@ -242,6 +235,42 @@ replica_clear_applier(struct replica *replica)
>  	trigger_clear(&replica->on_applier_state);
>  }
>  
> +void
> +replication_free(void)
> +{
> +	struct replica *replica, *next;
> +
> +	replica_hash_foreach_safe(&replicaset.hash, replica, next) {
> +		if (replica->id == instance_id) {
> +			replica_hash_remove(&replicaset.hash, replica);
> +			/*
> +			 * Local replica doesn't have neither applier
> +			 * nor relay, so ignore it.
> +			 */
> +			continue;
> +		}
> +		if (replica->applier != NULL) {
> +			replica_clear_applier(replica);
> +			/*
> +			 * We're exiting, so control won't be passed
> +			 * to appliers and we don't need to stop them.
> +			 */
> +		}

You don't need this code either. I want this loop to be as simple as

	/*
	 * <explain why>
	 */
	replicaset_foreach(replica)
		relay_cancel(replica->relay);

Then you wouldn't even need to move the definition of replication_free.

> +		if (replica->id != REPLICA_ID_NIL) {
> +			/*
> +			 * Relay threads keep sending messages
> +			 * to tx via cbus upon shutdown, which
> +			 * could lead to segfaults. So cancel
> +			 * them.
> +			 */
> +			relay_cancel(replica->relay);
> +		}
> +	}
> +
> +	free(replicaset.replica_by_id);
> +	fiber_cond_destroy(&replicaset.applier.cond);
> +}