[sheepdog] [PATCH v2] sheep: fix incorrect status transition during joining cluster
MORITA Kazutaka
morita.kazutaka at gmail.com
Mon Jun 17 18:00:44 CEST 2013
At Thu, 13 Jun 2013 01:02:29 +0900,
Hitoshi Mitake wrote:
>
> Current sheep set jm->cluster_status SD_STATUS_OK if
> 1. itself is a first node to join the cluster, and
> 2. a number of nodes of latest epoch is 1
> in sd_check_join_cb().
>
> This is an invalid behavior. For example, the behavior allows such a
> situation:
> 1. create a cluster with 2 nodes
> 2. format the cluster with --copies 2
> 3. kill single sheep
> 4. shutdown the cluster
> 5. launch single sheep
> 6. the status of the cluster is SD_STATUS_OK
Can you add a testcase to reproduce this problem?
>
> This patch solves the problem. The second condition is now checked
> with have_enough_zones() and the check is done in sd_join_handler().
>
> Signed-off-by: Hitoshi Mitake <mitake.hitoshi at lab.ntt.co.jp>
> ---
>
> v2:
> - call have_enough_zones() in sd_join_handler(), not in
> sd_check_join_cb(). The current_vnode_info isn't ready when
> sd_check_join_cb() called first.
> - do master transfer in cluster_running_check().
>
> sheep/group.c | 47 +++++++++++++++++++++++++++++------------------
> 1 files changed, 29 insertions(+), 18 deletions(-)
>
> diff --git a/sheep/group.c b/sheep/group.c
> index f74ef10..f0be421 100644
> --- a/sheep/group.c
> +++ b/sheep/group.c
> @@ -633,6 +633,11 @@ static int cluster_running_check(struct join_message *jm)
> {
> int ret;
>
> + if (sys->epoch < jm->epoch) {
> + sd_iprintf("master transfer, %d vs %d", sys->epoch, jm->epoch);
> + return CJ_RES_MASTER_TRANSFER;
> + }
> +
It is not obvious how this solves the problem you described in the
commit log. Can you explain it?
> /*
> * When the joining node is newly created and we are not waiting for
> * join we do not need to check anything.
> @@ -994,38 +999,40 @@ enum cluster_join_result sd_check_join_cb(const struct sd_node *joining,
> }
>
> if (node_is_local(joining)) {
> - struct sd_node entries[SD_MAX_NODES];
> - int nr_entries;
> - uint32_t epoch;
> -
> /*
> - * If I'm the first sheep joins in corosync, I
> + * If I'm the first sheep joins in the cluster, I
> * becomes the master without sending JOIN.
> */
> -
> - sd_printf(SDOG_DEBUG, "%s", node_to_str(&sys->this_node));
> + sd_dprintf("%s", node_to_str(&sys->this_node));
>
> jm->cluster_status = sys->status;
>
> - epoch = get_latest_epoch();
> + uint32_t epoch = get_latest_epoch();
Please submit these kinds of clean-ups as an isolated patch.
> if (!epoch)
> return CJ_RES_SUCCESS;
>
> if (sys->status != SD_STATUS_WAIT_FOR_JOIN) {
> - sd_eprintf("unexpected cluster status 0x%x",
> - sys->status);
> - return CJ_RES_FAIL;
> + if (sys->status == SD_STATUS_KILLED ||
> + sys->status == SD_STATUS_SHUTDOWN)
> + return CJ_RES_FAIL;
> +
> + /*
> + * This condition is not an error, a bug.
> + *
> + * sys->status is set as SD_STATUS_WAIT_FOR_JOIN in
> + * create_cluster() if epoch is not zero. And the status
> + * cannot be changed between create_cluster() and
> + * sd_check_join_cb() (exceptions are KILLED and
> + * SHUTDOWN).
> + */
> + panic("unexpected cluster status 0x%x"
> + " (expected one is 0x%x, wait for join)",
> + sys->status, SD_STATUS_WAIT_FOR_JOIN);
This looks like an unrelated change.
> }
>
> - nr_entries = epoch_log_read(epoch, entries, sizeof(entries));
> - if (nr_entries == -1)
> - return CJ_RES_FAIL;
> -
> sys->epoch = epoch;
> jm->ctime = get_cluster_ctime();
>
> - if (nr_entries == 1)
> - jm->cluster_status = SD_STATUS_OK;
> return CJ_RES_SUCCESS;
> }
>
> @@ -1139,7 +1146,11 @@ void sd_join_handler(const struct sd_node *joined,
>
> if (node_is_local(joined))
> /* this output is used for testing */
> - sd_printf(SDOG_DEBUG, "join Sheepdog cluster");
> + sd_dprintf("join Sheepdog cluster");
Again, please don't include an irrelevant change for us to review your
patch easily. Can you split this patch?
> +
> + if (sys->epoch && have_enough_zones())
> + sys->status = SD_STATUS_OK;
> +
This is not obvious too. Can you explain why we can set sys->status
to SD_STATUS_OK in this case? I think it's better to move this
assinment into update_cluster_info() if possible.
Thanks,
Kazutaka
More information about the sheepdog
mailing list