[ClusterLabs] [Cluster-devel] DLM connection channel switch take too long time (> 5mins)

[Gang He]

Hi Feldhost,


>>> 
> Hello Gang He,
> 
> which type of corosync rrp_mode you use? Passive or Active? 
clvm1:/etc/corosync # cat corosync.conf  | grep rrp_mode
        rrp_mode:       passive

Did you try test both?
No, only this mode. 
Also, what kernel version you use? I see some SCTP fixes in latest kernels.
clvm1:/etc/corosync # uname -r
4.4.114-94.11-default
It looks that sock->ops->connect() function is blocked for too long time before return, under broken network situation. 
In normal network, sock->ops->connect() function returns very quickly.

Thanks
Gang

> 
>> On 8 Mar 2018, at 08:52, Gang He <ghe at suse.com> wrote:
>> 
>> Hello list and David Teigland,
>> 
>> I got a problem under a two rings cluster, the problem can be reproduced 
> with the below steps.
>> 1) setup a two rings cluster with two nodes.
>> e.g. 
>> clvm1(nodeid 172204569)  addr_list eth0 10.67.162.25 eth1 192.168.152.240
>> clvm2(nodeid 172204570)  addr_list eth0 10.67.162.26 eth1 192.168.152.103
>> 
>> 2) the whole cluster works well, then I put eth0 down on node clvm2, and 
> restart pacemaker service on that node.
>> ifconfig eth0 down
>> rcpacemaker restart
>> 
>> 3) the whole cluster still work well (that means corosync is very smooth to 
> switch to the other ring).
>> Then, I can mount ocfs2 file system on node clvm2 quickly with the command 
>> mount /dev/sda /mnt/ocfs2 
>> 
>> 4) Next, I do the same mount on node clvm1, the mount command will be hanged 
> for about 5 mins, and finally the mount command is done.
>> But, if we setup a ocfs2 file system resource in pacemaker,
>> the pacemaker resource agent will consider ocfs2 file system resource 
> startup failure before this command returns,
>> the pacemaker will fence node clvm1. 
>> This problem is impacting our customer's estimate, since they think the two 
> rings can be switched smoothly.
>> 
>> According to this problem, I can see the mount command is hanged with the 
> below back trace,
>> clvm1:/ # cat /proc/6688/stack
>> [<ffffffffa04b8f2d>] new_lockspace+0x92d/0xa70 [dlm]
>> [<ffffffffa04b92d9>] dlm_new_lockspace+0x69/0x160 [dlm]
>> [<ffffffffa04db758>] user_cluster_connect+0xc8/0x350 [ocfs2_stack_user]
>> [<ffffffffa0483872>] ocfs2_cluster_connect+0x192/0x240 [ocfs2_stackglue]
>> [<ffffffffa0577efc>] ocfs2_dlm_init+0x31c/0x570 [ocfs2]
>> [<ffffffffa05c2983>] ocfs2_fill_super+0xb33/0x1200 [ocfs2]
>> [<ffffffff8120e130>] mount_bdev+0x1a0/0x1e0
>> [<ffffffff8120ea1a>] mount_fs+0x3a/0x170
>> [<ffffffff81228bf2>] vfs_kern_mount+0x62/0x110
>> [<ffffffff8122b123>] do_mount+0x213/0xcd0
>> [<ffffffff8122bed5>] SyS_mount+0x85/0xd0
>> [<ffffffff81614b0a>] entry_SYSCALL_64_fastpath+0x1e/0xb6
>> [<ffffffffffffffff>] 0xffffffffffffffff
>> 
>> The root cause is in sctp_connect_to_sock() function in lowcomms.c,
>> 1075
>> 1076         log_print("connecting to %d", con->nodeid);
>> 1077
>> 1078         /* Turn off Nagle's algorithm */
>> 1079         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
>> 1080                           sizeof(one));
>> 1081
>> 1082         result = sock->ops->connect(sock, (struct sockaddr *)&daddr, 
> addr_len,
>> 1083                                    O_NONBLOCK);  <<= here, this invoking 
> will cost > 5 mins before return ETIMEDOUT(-110).
>> 1084         printk(KERN_ERR "sctp_connect_to_sock connect: %d\n", result);
>> 1085
>> 1086         if (result == -EINPROGRESS)
>> 1087                 result = 0;
>> 1088         if (result == 0)
>> 1089                 goto out;
>> 
>> Then, I want to know if this problem was found/fixed before? 
>> it looks DLM can not switch the second ring very quickly, this will impact 
> the above application (e.g. CLVM, ocfs2) to create a new lock space before 
> it's startup.
>> 
>> Thanks
>> Gang
>> 
>> 
>> _______________________________________________
>> Users mailing list: Users at clusterlabs.org 
>> https://lists.clusterlabs.org/mailman/listinfo/users 
>> 
>> Project Home: http://www.clusterlabs.org 
>> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf 
>> Bugs: http://bugs.clusterlabs.org