[Gluster-users] I/O error on replicated volume

Jonathan Heese jheese at inetu.net
Fri Mar 27 05:34:30 UTC 2015 

On Mar 27, 2015, at 1:29 AM, "Mohammed Rafi K C" <rkavunga at redhat.com<mailto:rkavunga at redhat.com>> wrote:


When we change the transport from x to y, it should reflect in all the vol files. But unfortunately, the volume set command failed to change in nfs server, (of course it is a bug).  I had clearly mentioned in my previous mails, that changing the volume files using the volume set command is not recommended, i suggested this, just to check whether tcp work fine or not.

The reason why you are getting rdma connection error is because , now bricks are running through tcp, so the brick process will be listening on socket port. But nfs-server asked for an rdma connection, so they are trying to connect from rdma port to tcp port. Obviously the connection will be rejected.

Okay, thanks for the thorough explanation there.

Now that we know that TCP does function without the original I/O errors (from the start of this thread), how do you suggest that I proceed?

Do I have to wait for a subsequent release to rid myself of this bug?

Would it be feasible for me to switch from RDMA to TCP in a more permanent fashion (maybe wipe the cluster and start over?)?

Thanks.

Regards,
Jon Heese

Regards
Rafi KC

On 03/27/2015 12:28 AM, Jonathan Heese wrote:

Rafi,


Here is my nfs-server.vol file:


[root at duke ~]# cat /var/lib/glusterd/nfs/nfs-server.vol
volume gluster_disk-client-0
    type protocol/client
    option send-gids true
    option password 562ab460-7754-4b5a-82e6-18ed6c130786
    option username ad5d5754-cf02-4b96-9f85-ff3129ae0405
    option transport-type rdma
    option remote-subvolume /bricks/brick1
    option remote-host duke-ib
end-volume

volume gluster_disk-client-1
    type protocol/client
    option send-gids true
    option password 562ab460-7754-4b5a-82e6-18ed6c130786
    option username ad5d5754-cf02-4b96-9f85-ff3129ae0405
    option transport-type rdma
    option remote-subvolume /bricks/brick1
    option remote-host duchess-ib
end-volume

volume gluster_disk-replicate-0
    type cluster/replicate
    subvolumes gluster_disk-client-0 gluster_disk-client-1
end-volume

volume gluster_disk-dht
    type cluster/distribute
    subvolumes gluster_disk-replicate-0
end-volume

volume gluster_disk-write-behind
    type performance/write-behind
    subvolumes gluster_disk-dht
end-volume

volume gluster_disk
    type debug/io-stats
    option count-fop-hits off
    option latency-measurement off
    subvolumes gluster_disk-write-behind
end-volume

volume nfs-server
    type nfs/server
    option nfs3.gluster_disk.volume-id 2307a5a8-641e-44f4-8eaf-7cc2b704aafd
    option rpc-auth.addr.gluster_disk.allow *
    option nfs.drc off
    option nfs.nlm on
    option nfs.dynamic-volumes on
    subvolumes gluster_disk
end-volume


I see that "transport-type rdma" is listed a couple times here, but "gluster volume info" indicates that the volume is using the tcp transport:


[root at duke ~]# gluster volume info gluster_disk

Volume Name: gluster_disk
Type: Replicate
Volume ID: 2307a5a8-641e-44f4-8eaf-7cc2b704aafd
Status: Started
Number of Bricks: 1 x 2 = 2
Transport-type: tcp
Bricks:
Brick1: duke-ib:/bricks/brick1
Brick2: duchess-ib:/bricks/brick1
Options Reconfigured:
config.transport: tcp


Please let me know if you need any further information from me to determine how to correct this discrepancy.


Also, I feel compelled to ask: Since the TCP connections are going over the InfiniBand connections between the two Gluster servers (based on the hostnames which are pointed to the IB IPs via hosts files), are there any (significant) drawbacks to using TCP instead of RDMA here?  Thanks.


Regards,

Jon Heese


________________________________
From: Mohammed Rafi K C <rkavunga at redhat.com><mailto:rkavunga at redhat.com>
Sent: Monday, March 23, 2015 3:29 AM
To: Jonathan Heese
Cc: gluster-users
Subject: Re: [Gluster-users] I/O error on replicated volume


On 03/23/2015 11:28 AM, Jonathan Heese wrote:
On Mar 23, 2015, at 1:20 AM, "Mohammed Rafi K C" <rkavunga at redhat.com<mailto:rkavunga at redhat.com>> wrote:


On 03/21/2015 07:49 PM, Jonathan Heese wrote:

Mohamed,


I have completed the steps you suggested (unmount all, stop the volume, set the config.transport to tcp, start the volume, mount, etc.), and the behavior has indeed changed.


[root at duke ~]# gluster volume info

Volume Name: gluster_disk
Type: Replicate
Volume ID: 2307a5a8-641e-44f4-8eaf-7cc2b704aafd
Status: Started
Number of Bricks: 1 x 2 = 2
Transport-type: tcp
Bricks:
Brick1: duke-ib:/bricks/brick1
Brick2: duchess-ib:/bricks/brick1
Options Reconfigured:
config.transport: tcp

[root at duke ~]# gluster volume status
Status of volume: gluster_disk
Gluster process                                         Port    Online  Pid
------------------------------------------------------------------------------
Brick duke-ib:/bricks/brick1                            49152   Y       16362
Brick duchess-ib:/bricks/brick1                         49152   Y       14155
NFS Server on localhost                                 2049    Y       16374
Self-heal Daemon on localhost                           N/A     Y       16381
NFS Server on duchess-ib                                2049    Y       14167
Self-heal Daemon on duchess-ib                          N/A     Y       14174

Task Status of Volume gluster_disk
------------------------------------------------------------------------------
There are no active volume tasks


I am no longer seeing the I/O errors during prolonged periods of write I/O that I was seeing when the transport was set to rdma. However, I am seeing this message on both nodes every 3 seconds (almost exactly):


==> /var/log/glusterfs/nfs.log <==
[2015-03-21 14:17:40.379719] W [rdma.c:1076:gf_rdma_cm_event_handler] 0-gluster_disk-client-1: cma event RDMA_CM_EVENT_REJECTED, error 8 (me:10.10.10.1:1023 peer:10.10.10.2:49152)


Is this something to worry about?

If you are not using nfs to export the volumes, there is nothing to worry.

I'm using the native glusterfs FUSE component to mount the volume locally on both servers -- I assume that you're referring to the standard NFS protocol stuff, which I'm not using here.

Incidentally, I would like to keep my logs from filling up with junk if possible.  Is there something I can do to get rid of these (useless?) error messages?

If i understand correctly, you are getting this enormous log message from nfs log only, all other logs and everything are fine now, right ? If that is the case, and you are not at all using nfs for exporting the volume, as  a workaround you can disable nfs for your volume or cluster. (gluster v set nfs.disable on). This will turnoff your gluster nfs server, and you will no longer get those log messages.



Any idea why there are rdma pieces in play when I've set my transport to tcp?

there should not be any piece of rdma,if possible, can you paste the volfile for nfs server. You can find the volfile in /var/lib/glusterd/nfs/nfs-server.vol or /usr/local/var/lib/glusterd/nfs/nfs-server.vol

I will get this for you when I can.  Thanks.

If you can make it, that will be great help to understand the problem.


Rafi KC


Regards,
Jon Heese

Rafi KC

The actual I/O appears to be handled properly and I've seen no further errors in the testing I've done so far.


Thanks.


Regards,

Jon Heese


________________________________
From: gluster-users-bounces at gluster.org<mailto:gluster-users-bounces at gluster.org> <gluster-users-bounces at gluster.org><mailto:gluster-users-bounces at gluster.org> on behalf of Jonathan Heese <jheese at inetu.net><mailto:jheese at inetu.net>
Sent: Friday, March 20, 2015 7:04 AM
To: Mohammed Rafi K C
Cc: gluster-users
Subject: Re: [Gluster-users] I/O error on replicated volume

Mohammed,

Thanks very much for the reply.  I will try that and report back.

Regards,
Jon Heese

On Mar 20, 2015, at 3:26 AM, "Mohammed Rafi K C" <rkavunga at redhat.com<mailto:rkavunga at redhat.com>> wrote:


On 03/19/2015 10:16 PM, Jonathan Heese wrote:
Hello all,

Does anyone else have any further suggestions for troubleshooting this?

To sum up: I have a 2 node 2 brick replicated volume, which holds a handful of iSCSI image files which are mounted and served up by tgtd (CentOS 6) to a handful of devices on a dedicated iSCSI network.  The most important iSCSI clients (initiators) are four VMware ESXi 5.5 hosts that use the iSCSI volumes as backing for their datastores for virtual machine storage.

After a few minutes of sustained writing to the volume, I am seeing a massive flood (over 1500 per second at times) of this error in /var/log/glusterfs/mnt-gluster-disk.log:
[2015-03-16 02:24:07.582801] W [fuse-bridge.c:2242:fuse_writev_cbk] 0-glusterfs-fuse: 635358: WRITE => -1 (Input/output error)

When this happens, the ESXi box fails its write operation and returns an error to the effect of “Unable to write data to datastore”.  I don’t see anything else in the supporting logs to explain the root cause of the i/o errors.

Any and all suggestions are appreciated.  Thanks.


>From the mount logs, i assume that your volume transport type is rdma. There are some known issues for rdma in 3.5.3, and the patch for to address those issues are already send to upstream [1]. From the logs, I'm not sure and it is hard to tell you whether this problem is something related to rdma transport or not. To make sure that the tcp transport is works well in this scenario, if possible can you try to reproduce the same using tcp type volumes. You can change the transport type of volume by doing the following step ( not recommended in normal use case).

1) unmount every client
2) stop the volume
3) run gluster volume set volname config.transport tcp
4) start the volume again
5) mount the clients

[1] : http://goo.gl/2PTL61

Regards
Rafi KC

Jon Heese
Systems Engineer
INetU Managed Hosting
P: 610.266.7441 x 261
F: 610.266.7434
www.inetu.net<https://www.inetu.net/>
** This message contains confidential information, which also may be privileged, and is intended only for the person(s) addressed above. Any unauthorized use, distribution, copying or disclosure of confidential and/or privileged information is strictly prohibited. If you have received this communication in error, please erase all copies of the message and its attachments and notify the sender immediately via reply e-mail. **

From: Jonathan Heese
Sent: Tuesday, March 17, 2015 12:36 PM
To: 'Ravishankar N'; gluster-users at gluster.org<mailto:gluster-users at gluster.org>
Subject: RE: [Gluster-users] I/O error on replicated volume

Ravi,

The last lines in the mount log before the massive vomit of I/O errors are from 22 minutes prior, and seem innocuous to me:

[2015-03-16 01:37:07.126340] E [client-handshake.c:1760:client_query_portmap_cbk] 0-gluster_disk-client-0: failed to get the port number for remote subvolume. Please run 'gluster volume status' on server to see if brick process is running.
[2015-03-16 01:37:07.126587] W [rdma.c:4273:gf_rdma_disconnect] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_notify+0x13f) [0x7fd9c557bccf] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_handle_reply+0xa5) [0x7fd9c557a995] (-->/usr/lib64/glusterfs/3.5.3/xlator/protocol/client.so(client_query_portmap_cbk+0x1ea) [0x7fd9c0d8fb9a]))) 0-gluster_disk-client-0: disconnect called (peer:10.10.10.1:24008)
[2015-03-16 01:37:07.126687] E [client-handshake.c:1760:client_query_portmap_cbk] 0-gluster_disk-client-1: failed to get the port number for remote subvolume. Please run 'gluster volume status' on server to see if brick process is running.
[2015-03-16 01:37:07.126737] W [rdma.c:4273:gf_rdma_disconnect] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_notify+0x13f) [0x7fd9c557bccf] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_handle_reply+0xa5) [0x7fd9c557a995] (-->/usr/lib64/glusterfs/3.5.3/xlator/protocol/client.so(client_query_portmap_cbk+0x1ea) [0x7fd9c0d8fb9a]))) 0-gluster_disk-client-1: disconnect called (peer:10.10.10.2:24008)
[2015-03-16 01:37:10.730165] I [rpc-clnt.c:1729:rpc_clnt_reconfig] 0-gluster_disk-client-0: changing port to 49152 (from 0)
[2015-03-16 01:37:10.730276] W [rdma.c:4273:gf_rdma_disconnect] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_notify+0x13f) [0x7fd9c557bccf] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_handle_reply+0xa5) [0x7fd9c557a995] (-->/usr/lib64/glusterfs/3.5.3/xlator/protocol/client.so(client_query_portmap_cbk+0x1ea) [0x7fd9c0d8fb9a]))) 0-gluster_disk-client-0: disconnect called (peer:10.10.10.1:24008)
[2015-03-16 01:37:10.739500] I [rpc-clnt.c:1729:rpc_clnt_reconfig] 0-gluster_disk-client-1: changing port to 49152 (from 0)
[2015-03-16 01:37:10.739560] W [rdma.c:4273:gf_rdma_disconnect] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_notify+0x13f) [0x7fd9c557bccf] (-->/usr/lib64/libgfrpc.so.0(rpc_clnt_handle_reply+0xa5) [0x7fd9c557a995] (-->/usr/lib64/glusterfs/3.5.3/xlator/protocol/client.so(client_query_portmap_cbk+0x1ea) [0x7fd9c0d8fb9a]))) 0-gluster_disk-client-1: disconnect called (peer:10.10.10.2:24008)
[2015-03-16 01:37:10.741883] I [client-handshake.c:1677:select_server_supported_programs] 0-gluster_disk-client-0: Using Program GlusterFS 3.3, Num (1298437), Version (330)
[2015-03-16 01:37:10.744524] I [client-handshake.c:1462:client_setvolume_cbk] 0-gluster_disk-client-0: Connected to 10.10.10.1:49152, attached to remote volume '/bricks/brick1'.
[2015-03-16 01:37:10.744537] I [client-handshake.c:1474:client_setvolume_cbk] 0-gluster_disk-client-0: Server and Client lk-version numbers are not same, reopening the fds
[2015-03-16 01:37:10.744566] I [afr-common.c:4267:afr_notify] 0-gluster_disk-replicate-0: Subvolume 'gluster_disk-client-0' came back up; going online.
[2015-03-16 01:37:10.744627] I [client-handshake.c:450:client_set_lk_version_cbk] 0-gluster_disk-client-0: Server lk version = 1
[2015-03-16 01:37:10.753037] I [client-handshake.c:1677:select_server_supported_programs] 0-gluster_disk-client-1: Using Program GlusterFS 3.3, Num (1298437), Version (330)
[2015-03-16 01:37:10.755657] I [client-handshake.c:1462:client_setvolume_cbk] 0-gluster_disk-client-1: Connected to 10.10.10.2:49152, attached to remote volume '/bricks/brick1'.
[2015-03-16 01:37:10.755676] I [client-handshake.c:1474:client_setvolume_cbk] 0-gluster_disk-client-1: Server and Client lk-version numbers are not same, reopening the fds
[2015-03-16 01:37:10.761945] I [fuse-bridge.c:5016:fuse_graph_setup] 0-fuse: switched to graph 0
[2015-03-16 01:37:10.762144] I [client-handshake.c:450:client_set_lk_version_cbk] 0-gluster_disk-client-1: Server lk version = 1
[2015-03-16 01:37:10.762279] I [fuse-bridge.c:3953:fuse_init] 0-glusterfs-fuse: FUSE inited with protocol versions: glusterfs 7.22 kernel 7.14
[2015-03-16 01:59:26.098670] W [fuse-bridge.c:2242:fuse_writev_cbk] 0-glusterfs-fuse: 292084: WRITE => -1 (Input/output error)
…

I’ve seen no indication of split-brain on any files at any point in this (ever since downdating from 3.6.2 to 3.5.3, which is when this particular issue started):
[root at duke gfapi-module-for-linux-target-driver-]# gluster v heal gluster_disk info
Brick duke.jonheese.local:/bricks/brick1/
Number of entries: 0

Brick duchess.jonheese.local:/bricks/brick1/
Number of entries: 0

Thanks.

Jon Heese
Systems Engineer
INetU Managed Hosting
P: 610.266.7441 x 261
F: 610.266.7434
www.inetu.net<https://www.inetu.net/>
** This message contains confidential information, which also may be privileged, and is intended only for the person(s) addressed above. Any unauthorized use, distribution, copying or disclosure of confidential and/or privileged information is strictly prohibited. If you have received this communication in error, please erase all copies of the message and its attachments and notify the sender immediately via reply e-mail. **

From: Ravishankar N [mailto:ravishankar at redhat.com]
Sent: Tuesday, March 17, 2015 12:35 AM
To: Jonathan Heese; gluster-users at gluster.org<mailto:gluster-users at gluster.org>
Subject: Re: [Gluster-users] I/O error on replicated volume


On 03/17/2015 02:14 AM, Jonathan Heese wrote:
Hello,

So I resolved my previous issue with split-brains and the lack of self-healing by dropping my installed glusterfs* packages from 3.6.2 to 3.5.3, but now I've picked up a new issue, which actually makes normal use of the volume practically impossible.

A little background for those not already paying close attention:
I have a 2 node 2 brick replicating volume whose purpose in life is to hold iSCSI target files, primarily for use to provide datastores to a VMware ESXi cluster.  The plan is to put a handful of image files on the Gluster volume, mount them locally on both Gluster nodes, and run tgtd on both, pointed to the image files on the mounted gluster volume. Then the ESXi boxes will use multipath (active/passive) iSCSI to connect to the nodes, with automatic failover in case of planned or unplanned downtime of the Gluster nodes.

In my most recent round of testing with 3.5.3, I'm seeing a massive failure to write data to the volume after about 5-10 minutes, so I've simplified the scenario a bit (to minimize the variables) to: both Gluster nodes up, only one node (duke) mounted and running tgtd, and just regular (single path) iSCSI from a single ESXi server.

About 5-10 minutes into migration a VM onto the test datastore, /var/log/messages on duke gets blasted with a ton of messages exactly like this:
Mar 15 22:24:06 duke tgtd: bs_rdwr_request(180) io error 0x1781e00 2a -1 512 22971904, Input/output error

And /var/log/glusterfs/mnt-gluster_disk.log gets blased with a ton of messages exactly like this:
[2015-03-16 02:24:07.572279] W [fuse-bridge.c:2242:fuse_writev_cbk] 0-glusterfs-fuse: 635299: WRITE => -1 (Input/output error)


Are there any messages in the mount log from AFR about split-brain just before the above line appears?
Does `gluster v heal <VOLNAME> info` show any files? Performing I/O on files that are in split-brain fail with EIO.

-Ravi

And the write operation from VMware's side fails as soon as these messages start.

I don't see any other errors (in the log files I know of) indicating the root cause of these i/o errors.  I'm sure that this is not enough information to tell what's going on, but can anyone help me figure out what to look at next to figure this out?

I've also considered using Dan Lambright's libgfapi gluster module for tgtd (or something similar) to avoid going through FUSE, but I'm not sure whether that would be irrelevant to this problem, since I'm not 100% sure if it lies in FUSE or elsewhere.

Thanks!

Jon Heese
Systems Engineer
INetU Managed Hosting
P: 610.266.7441 x 261
F: 610.266.7434
www.inetu.net<https://www.inetu.net/>
** This message contains confidential information, which also may be privileged, and is intended only for the person(s) addressed above. Any unauthorized use, distribution, copying or disclosure of confidential and/or privileged information is strictly prohibited. If you have received this communication in error, please erase all copies of the message and its attachments and notify the sender immediately via reply e-mail. **




_______________________________________________

Gluster-users mailing list

Gluster-users at gluster.org<mailto:Gluster-users at gluster.org>

http://www.gluster.org/mailman/listinfo/gluster-users




_______________________________________________
Gluster-users mailing list
Gluster-users at gluster.org<mailto:Gluster-users at gluster.org>
http://www.gluster.org/mailman/listinfo/gluster-users




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20150327/8695f68e/attachment.html>

More information about the Gluster-users mailing list