[Cluster-devel] [PATCH] decrease expected_replies when remote node is down [V2]

[Dongmao Zhang <dmzhang@suse.com>]

dear list,

how about this patch?

any comments there?



? 2014?02?26? 19:20, dongmao zhang ??:
> dear list,
> I think the last patch is broken, because
> it has line:thisfd->bits.localsock.finished = 0
> which may cause pre_post_thread ends.
> This is the revised patch. So any comments?
> 
>> when clvmd is waiting for remote node's relies, but
>> the remote node could be possiblly closed at the same
>> time. This could result in a request timeout in mainloop.
>> The default timeout is 60 seconds. This is still too large.
>>
>> https://bugzilla.novell.com/show_bug.cgi?id=837538
>> https://bugzilla.novell.com/show_bug.cgi?id=862633
>> clvmd users usually meet this when they call LVM command(like vgscan)
>> in one node, while other nodes's clvmd is closing.
>>
>> My solution to this is to decrement in-flight request's
>> expected_replies. The command could finish much faster in this
>> situation.
> ---
>   daemons/clvmd/clvmd-corosync.c |    7 ++++-
>   daemons/clvmd/clvmd.c          |   52 +++++++++++++++++++++++++++++++++++++++-
>   daemons/clvmd/clvmd.h          |    2 +
>   3 files changed, 59 insertions(+), 2 deletions(-)
> 
> diff --git a/daemons/clvmd/clvmd-corosync.c b/daemons/clvmd/clvmd-corosync.c
> index e68cb73..8e642ca 100644
> --- a/daemons/clvmd/clvmd-corosync.c
> +++ b/daemons/clvmd/clvmd-corosync.c
> @@ -251,11 +251,16 @@ static void corosync_cpg_confchg_callback(cpg_handle_t handle,
>   		ninfo = dm_hash_lookup_binary(node_hash,
>   					      (char *)&left_list[i].nodeid,
>   					      COROSYNC_CSID_LEN);
> -		if (ninfo)
> +		if (ninfo) {
>   			ninfo->state = NODE_DOWN;
> +			char name[MAX_CLUSTER_MEMBER_NAME_LEN];
> +			sprintf(name, "%x", ninfo->nodeid);
> +			decrease_inflight_expected_reply(name);
> +		}
>   	}
>   
>   	num_nodes = member_list_entries;
> +
>   }
>   
>   static int _init_cluster(void)
> diff --git a/daemons/clvmd/clvmd.c b/daemons/clvmd/clvmd.c
> index fc8cf6e..bd056f2 100644
> --- a/daemons/clvmd/clvmd.c
> +++ b/daemons/clvmd/clvmd.c
> @@ -1697,6 +1697,56 @@ static void process_remote_command(struct clvm_header *msg, int msglen, int fd,
>   	free(replyargs);
>   }
>   
> +void decrease_inflight_expected_reply(char *nodename)
> +{
> +	struct local_client * thisfd;
> +	struct node_reply *reply;
> +
> +	DEBUGLOG("remote node %s down", nodename);
> +
> +	for (thisfd = &local_client_head; thisfd != NULL;
> +	     thisfd = thisfd->next) {
> +		/* in-flight request */
> +		if (thisfd->type == LOCAL_SOCK
> +		    && thisfd->bits.localsock.sent_out
> +		    && thisfd->bits.localsock.in_progress
> +		    && ! thisfd->bits.localsock.finished
> +		    && thisfd->bits.localsock.expected_replies >
> +		    thisfd->bits.localsock.num_replies) {
> +
> +			pthread_mutex_lock(&thisfd->bits.localsock.reply_mutex);
> +
> +			reply = thisfd->bits.localsock.replies;
> +			while (reply && strcmp(reply->node, nodename) != 0) {
> +				reply = reply->next;
> +			}
> +			/* if the remote down server has replies,do not decrease the expected_replies */
> +			if (reply)
> +				continue;
> +
> +			thisfd->bits.localsock.expected_replies--;
> +			DEBUGLOG
> +			    ("remote node down, decrement the expected replies to (%ld),num_replies(%ld)",
> +			     thisfd->bits.localsock.expected_replies,
> +			     thisfd->bits.localsock.num_replies)
> +
> +			if (thisfd->bits.localsock.expected_replies <= thisfd->bits.localsock.num_replies) {
> +				/* tell pre_and_post thread to finish */
> +				if (thisfd->bits.localsock.threadid) {
> +					thisfd->bits.localsock.all_success = 0;
> +					pthread_mutex_lock(&thisfd->bits.localsock.mutex);
> +					thisfd->bits.localsock.state = POST_COMMAND;
> +					pthread_cond_signal(&thisfd->bits.localsock.cond);
> +					pthread_mutex_unlock(&thisfd->bits.localsock.mutex);
> +				}
> +			}
> +			pthread_mutex_unlock(&thisfd->bits.localsock.reply_mutex);
> +
> +		}
> +	}
> +
> +}
> +
>   /* Add a reply to a command to the list of replies for this client.
>      If we have got a full set then send them to the waiting client down the local
>      socket */
> @@ -1738,7 +1788,7 @@ static void add_reply_to_list(struct local_client *client, int status,
>   		 client->bits.localsock.expected_replies);
>   
>   	/* If we have the whole lot then do the post-process */
> -	if (++client->bits.localsock.num_replies ==
> +	if (++client->bits.localsock.num_replies >=
>   	    client->bits.localsock.expected_replies) {
>   		/* Post-process the command */
>   		if (client->bits.localsock.threadid) {
> diff --git a/daemons/clvmd/clvmd.h b/daemons/clvmd/clvmd.h
> index 0f837dd..52e4e61 100644
> --- a/daemons/clvmd/clvmd.h
> +++ b/daemons/clvmd/clvmd.h
> @@ -112,6 +112,8 @@ extern int do_post_command(struct local_client *client);
>   extern void cmd_client_cleanup(struct local_client *client);
>   extern int add_client(struct local_client *new_client);
>   
> +
> +extern void decrease_inflight_expected_reply();
>   extern void clvmd_cluster_init_completed(void);
>   extern void process_message(struct local_client *client, char *buf,
>   			    int len, const char *csid);
>