[Tarantool-patches] [PATCH 1/5] [tosquash] replication: fix multiple rollbacks

[Serge Petrenko]

03.07.2020 02:40, Vladislav Shpilevoy пишет:
> The problem was that if several transactions time out in one
> event loop iteration, the all will write rollback. Moreover, they
> will do that in a weird order, starting from the oldest, not in
> a reversed order.
>
> This patch makes limbo write only one rollback at once.
> ---
>   src/box/txn_limbo.c                 | 25 +++++++++++++++++++++++++
>   test/replication/qsync_basic.result |  2 +-
>   2 files changed, 26 insertions(+), 1 deletion(-)
>
> diff --git a/src/box/txn_limbo.c b/src/box/txn_limbo.c
> index 0402664cb..2cb687f4d 100644
> --- a/src/box/txn_limbo.c
> +++ b/src/box/txn_limbo.c
> @@ -44,6 +44,13 @@ txn_limbo_create(struct txn_limbo *limbo)
>   	limbo->got_rollback = false;
>   }
>   
> +static inline struct txn_limbo_entry *
> +txn_limbo_first_entry(struct txn_limbo *limbo)
> +{
> +	return rlist_first_entry(&limbo->queue, struct txn_limbo_entry,
> +				 in_queue);
> +}
> +
>   struct txn_limbo_entry *
>   txn_limbo_append(struct txn_limbo *limbo, uint32_t id, struct txn *txn)
>   {
> @@ -150,6 +157,24 @@ txn_limbo_wait_complete(struct txn_limbo *limbo, struct txn_limbo_entry *entry)
>   	bool timed_out = fiber_yield_timeout(txn_limbo_confirm_timeout(limbo));
>   	fiber_set_cancellable(cancellable);
>   	if (timed_out) {
> +		assert(!txn_limbo_is_empty(limbo));
> +		if (txn_limbo_first_entry(limbo) != entry) {
> +			/*
> +			 * If this is not a first entry in the
> +			 * limbo, it is definitely not a first
> +			 * timed out entry. And since it managed
> +			 * to time out too, it means there is
> +			 * currently another fiber writing
> +			 * rollback. Wait when it will finish and
> +			 * wake us up.
> +			 */

Why isn't it the first timed out? Is it because once previous entry was 
confirmed, it
is removed from the queue immediately?
Looks fragile.

> +			bool cancellable = fiber_set_cancellable(false);
> +			fiber_yield();
> +			fiber_set_cancellable(cancellable);
> +			assert(txn_limbo_entry_is_complete(entry));
> +			goto complete;
> +		}
> +
>   		txn_limbo_write_rollback(limbo, entry);
>   		struct txn_limbo_entry *e, *tmp;
>   		rlist_foreach_entry_safe_reverse(e, &limbo->queue,
> diff --git a/test/replication/qsync_basic.result b/test/replication/qsync_basic.result
> index cdecf00e8..32deb2ac3 100644
> --- a/test/replication/qsync_basic.result
> +++ b/test/replication/qsync_basic.result
> @@ -272,7 +272,7 @@ box.cfg{replication_synchro_timeout = 0.001, replication_synchro_quorum = 3}
>    | ...
>   f = fiber.create(box.space.sync.replace, box.space.sync, {6}) s:replace{6}
>    | ---
> - | - error: Quorum collection for a synchronous transaction is timed out
> + | - error: A rollback for a synchronous transaction is received
>    | ...
>   f:status()
>    | ---

-- 
Serge Petrenko