Re: [tarantool-patches] [PATCH v2] replication: implement replication_shutdown()

[Vladimir Davydov <vdavydov.dev@gmail.com>]

On Mon, Aug 06, 2018 at 10:09:17AM +0300, Serge Petrenko wrote:
> Relay threads keep using tx upon shutdown, which leads to occasional
> segmentation faults and assertion fails (e.g. in replication test
> suite).
> 
> Fix this by implementing replication_shutdown and relay_halt functions.
> replication_shutdown calls relay_halt to stop every relay thread that is
> using tx.
> 
> Closes #3485
> ---
> https://github.com/tarantool/tarantool/issues/3485
> https://github.com/tarantool/tarantool/tree/sergepetrenko/gh-3485-replication-shutdown
> 
> Changes in v2:
>   - instead of setting tx_in_use flag
>     in relay and checking it in tx, send a
>     message from relay to tx to set the flag.
> 
>  src/box/box.cc         |  2 +-
>  src/box/relay.cc       | 85 ++++++++++++++++++++++++++++++++++++++++++++++++--
>  src/box/relay.h        | 10 ++++++
>  src/box/replication.cc | 30 ++++++++++++++++++
>  src/box/replication.h  |  6 ++++
>  5 files changed, 129 insertions(+), 4 deletions(-)

I see some discrepancy between the patch submitted for review and the
code pushed to the branch:

vlad@esperanza tarantool$ date
Wed Aug  8 19:12:29 MSK 2018
vlad@esperanza tarantool$ git remote update origin
Fetching origin
vlad@esperanza tarantool$ git status
On branch sergepetrenko/gh-3485-replication-shutdown
Your branch is up-to-date with 'origin/sergepetrenko/gh-3485-replication-shutdown'.
nothing to commit, working tree clean
vlad@esperanza tarantool$ git show --oneline --stat
920dc83a replication: implement replication_shutdown()
 src/box/box.cc               |  2 +-
 src/box/relay.cc             | 85 ++++++++++++++++++++++++++++++++++++++++++--
 src/box/relay.h              | 10 ++++++
 src/box/replication.cc       | 30 ++++++++++++++++
 src/box/replication.h        |  6 ++++
 test/replication/gc.result   | 12 +++++++
 test/replication/gc.test.lua |  4 +++
 7 files changed, 145 insertions(+), 4 deletions(-)

Are you trying to conceal what you intend to do to the tests?

> diff --git a/src/box/relay.cc b/src/box/relay.cc
> index 4cacbc840..60cb11932 100644
> --- a/src/box/relay.cc
> +++ b/src/box/relay.cc
> @@ -63,6 +63,9 @@ struct relay_status_msg {
>  	struct relay *relay;
>  	/** Replica vclock. */
>  	struct vclock vclock;
> +	/** A flag to notify tx on creation / before removal
> +	 * of tx_pipe/relay_pipe. */

Malformed comment style.

> +	bool tx_in_use;
>  };
>  
>  /**
> @@ -82,10 +85,22 @@ struct relay_gc_msg {
>  	struct vclock vclock;
>  };
>  
> +/**
> + * Cbus message sent by tx thread to stop relay on shutdown.
> + */
> +struct relay_halt_msg {
> +	/** Parent. */
> +	struct cmsg msg;
> +	/** Relay instance. */
> +	struct relay *relay;
> +};
> +
>  /** State of a replication relay. */
>  struct relay {
>  	/** The thread in which we relay data to the replica. */
>  	struct cord cord;
> +	/** The main fiber in cord to be canceled upon relay halt. */
> +	struct fiber *main_fiber;
>  	/** Replica connection */
>  	struct ev_io io;
>  	/** Request sync */
> @@ -120,6 +135,11 @@ struct relay {
>  	struct cpipe tx_pipe;
>  	/** A pipe from 'tx' thread to 'relay' */
>  	struct cpipe relay_pipe;
> +	/**
> +	 * A flag indicating that we executed relay_subscribe_f and
> +	 * have tx_pipe and relay_pipe ready.
> +	 */
> +	bool tx_in_use;

I rather dislike this flag. I think it should be a part of cbus
subsystem. Say,

  bool cpipe_is_created(cpipe) { return cpipe->endpoint != NULL }

or something like that.

>  	/** Status message */
>  	struct relay_status_msg status_msg;
>  	/**
> @@ -152,6 +172,12 @@ relay_get_state(const struct relay *relay)
>  	return relay->state;
>  }
>  
> +bool
> +relay_uses_tx(const struct relay *relay)
> +{
> +	return relay->tx_in_use;
> +}
> +

No point in exporting this function. You can check that relay_pipe is
available right in relay_halt.

>  const struct vclock *
>  relay_vclock(const struct relay *relay)
>  {
> @@ -198,6 +224,40 @@ relay_start(struct relay *relay, int fd, uint64_t sync,
>  	relay->state = RELAY_FOLLOW;
>  }
>  
> +static void
> +relay_main_fiber_halt(struct cmsg *msg)
> +{
> +	struct relay_halt_msg *m = (struct relay_halt_msg *)msg;
> +	struct relay *relay = m->relay;
> +
> +	assert(relay->main_fiber != NULL);
> +	fiber_cancel(relay->main_fiber);
> +	relay->main_fiber = NULL;

Can't you simply use fiber() here? AFAIU cbus messages are processed by
the "main" relay fiber anyway, see relay_subscribe_f.

BTW this code isn't covered by any test:

  https://coveralls.io/builds/18341862/source?filename=src/box/relay.cc#L228

Please make sure it is.

> +
> +	free(m);
> +}
> +
> +void
> +relay_halt(struct relay *relay)

'halt' is an exact synonym of 'stop', and we already have relay_stop().
I think this one should be called relay_cancel() or relay_abort() or
something like that. Please come up with a better name to avoid
confusion.

> +{
> +	assert(relay->state == RELAY_FOLLOW);
> +
> +	static const struct cmsg_hop route[] ={
> +		{relay_main_fiber_halt, NULL}
> +	};
> +	struct relay_halt_msg *m = (struct relay_halt_msg *)malloc(sizeof(*m));
> +	if (m == NULL) {
> +		/*
> +		 * Out of memory during shutdown. Do nothing.
> +		 */
> +		say_warn("failed to allocate relay halt message");
> +		return;
> +	}
> +	cmsg_init(&m->msg, route);
> +	m->relay = relay;
> +	cpipe_push(&relay->relay_pipe, &m->msg);
> +}

AFAIR cpipe_push() doesn't necessarily flushes the input.
cpipe_flush_input(), may be?

Also, I think you should wait for the relay thread to exit (see
cord_join), otherwise you may proceed to tx destruction while it
can still access tx data.

> diff --git a/src/box/replication.cc b/src/box/replication.cc
> index 48956d2ed..9b4968777 100644
> --- a/src/box/replication.cc
> +++ b/src/box/replication.cc
> @@ -398,6 +398,36 @@ replica_on_applier_state_f(struct trigger *trigger, void *event)
>  	fiber_cond_signal(&replicaset.applier.cond);
>  }
>  
> +void
> +replication_shutdown()

So now we have replication_shutdown and replication_free, both public,
and replication_free is only used in replication_shutdown. Not good.

We typically call subsys constructor something_init() and destructor
something_free(). That said, I guess this code should be a part of
replication_free() and replication_shutdown() shouldn't exist.

> +{
> +	struct replica *replica, *next;
> +
> +	replica_hash_foreach_safe(&replicaset.hash, replica, next) {

> +		if (replica->id == instance_id)
> +			continue;

Why? Don't we want to delete all replicas, including ourselves?

> +		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.
> +			 */
> +		}
> +		if (replica->id != REPLICA_ID_NIL) {
> +			if (relay_get_state(replica->relay) == RELAY_FOLLOW &&
> +			    relay_uses_tx(replica->relay)) {
> +				replica->id = REPLICA_ID_NIL;
> +				relay_halt(replica->relay);
> +			}
> +		} else {
> +			replica_hash_remove(&replicaset.hash, replica);
> +			replica_delete(replica);

Don't we want to delete all replicas here? Can we?

What about replicas on the replicaset.anon list. Shouldn't we delete
them, too?

> +		}
> +	}
> +
> +	replication_free();
> +}