Re: libgfapi failover problem on replica bricks

[Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>]

On 08/06/2014 12:27 PM, Roman wrote:

Yesterday I've reproduced this situation two times.

 

The setup:

1. Hardware and network

   a. Disks INTEL SSDSC2BB240G4

   b1. Network cards: X540-AT2

   b2. Netgear 10g switch

2. Software setup:

   a. OS: Debian wheezy

   b. Glusterfs: 3.4.4 (latest 3.4.4 from gluster repository)

   c. Promox VE with update glusterfs from gluster repository

3. Software Configuration

   a. create replicated volume with cluster.self-heal-daemon: off; nfs.disable: off; network.ping-timeout: 2 opts

   b. mount it on proxmox VE (via proxmox gui, it mouts with these opts: stor1:HA-fast-150G-PVE1 on /mnt/pve/FAST-TESt type fuse.glusterfs (rw,relatime,user_id=0,group_id=0,default_permissions,allow_other,max_read=131072)   )

   c. install VM with qcow2 or raw disk image.

   d. disable port / remove network cable from one of storage servers

   e. wait and put cable back

   f. keep waiting for sync (pointless, it won't ever start)

   g. disable another port for second server (or remove cable from second server)

   h. profit.

Maybe I could use 3.5.2 from debian sid (testing) repository to test with?

Sure, you can go ahead. I will just write one document about maintaining VMs on gluster from the perspective of replication.

Pranith

2014-08-06 9:39 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

Roman,
    The file went into split-brain. I think we should do these tests with 3.5.2. Where monitoring the heals is easier. Let me also come up with a document about how to do this testing you are trying to do.

Humble/Niels,
    Do we have debs available for 3.5.2? In 3.5.1 there was packaging issue where /usr/bin/glfsheal is not packaged along with the deb. I think that should be fixed now as well?

Pranith

On 08/06/2014 11:52 AM, Roman wrote:

good morning, 

root@stor1:~# getfattr -d -m. -e hex /exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

getfattr: Removing leading '/' from absolute path names

# file: exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

trusted.afr.HA-fast-150G-PVE1-client-0=0x000000000000000000000000

trusted.afr.HA-fast-150G-PVE1-client-1=0x000001320000000000000000

trusted.gfid=0x23c79523075a4158bea38078da570449

getfattr: Removing leading '/' from absolute path names

# file: exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

trusted.afr.HA-fast-150G-PVE1-client-0=0x000000040000000000000000

trusted.afr.HA-fast-150G-PVE1-client-1=0x000000000000000000000000

trusted.gfid=0x23c79523075a4158bea38078da570449

2014-08-06 9:20 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

On 08/06/2014 11:30 AM, Roman wrote:

Also, this time files are not the same!

root@stor1:~# md5sum /exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

32411360c53116b96a059f17306caeda  /exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

root@stor2:~# md5sum /exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

65b8a6031bcb6f5fb3a11cb1e8b1c9c9  /exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

What is the getfattr output?

Pranith

2014-08-05 16:33 GMT+03:00 Roman <romeo.r@xxxxxxxxx>:

Nope, it is not working. But this time it went a bit other way

root@gluster-client:~# dmesg

Segmentation fault

I was not able even to start the VM after I done the tests

Could not read qcow2 header: Operation not permitted

And it seems, it never starts to sync files after first disconnect. VM survives first disconnect, but not second (I waited around 30 minutes). Also, I've got network.ping-timeout: 2 in volume settings, but logs react on first disconnect around 30 seconds. Second was faster, 2 seconds.

Reaction was different also:

slower one:

[2014-08-05 13:26:19.558435] W [socket.c:514:__socket_rwv] 0-glusterfs: readv failed (Connection timed out)

[2014-08-05 13:26:19.558485] W [socket.c:1962:__socket_proto_state_machine] 0-glusterfs: reading from socket failed. Error (Connection timed out), peer (10.250.0.1:24007)

[2014-08-05 13:26:21.281426] W [socket.c:514:__socket_rwv] 0-HA-fast-150G-PVE1-client-0: readv failed (Connection timed out)

[2014-08-05 13:26:21.281474] W [socket.c:1962:__socket_proto_state_machine] 0-HA-fast-150G-PVE1-client-0: reading from socket failed. Error (Connection timed out), peer (10.250.0.1:49153)

[2014-08-05 13:26:21.281507] I [client.c:2098:client_rpc_notify] 0-HA-fast-150G-PVE1-client-0: disconnected

the fast one:

2014-08-05 12:52:44.607389] C [client-handshake.c:127:rpc_client_ping_timer_expired] 0-HA-fast-150G-PVE1-client-1: server 10.250.0.2:49153 has not responded in the last 2 seconds, disconnecting.

[2014-08-05 12:52:44.607491] W [socket.c:514:__socket_rwv] 0-HA-fast-150G-PVE1-client-1: readv failed (No data available)

[2014-08-05 12:52:44.607585] E [rpc-clnt.c:368:saved_frames_unwind] (-->/usr/lib/x86_64-linux-gnu/libgfrpc.so.0(rpc_clnt_notify+0xf8) [0x7fcb1b4b0558] (-->/usr/lib/x86_64-linux-gnu/libgfrpc.so.0(rpc_clnt_connection_cleanup+0xc3) [0x7fcb1b4aea63] (-->/usr/lib/x86_64-linux-gnu/libgfrpc.so.0(saved_frames_destroy+0xe) [0x7fcb1b4ae97e]))) 0-HA-fast-150G-PVE1-client-1: forced unwinding frame type(GlusterFS 3.3) op(LOOKUP(27)) called at 2014-08-05 12:52:42.463881 (xid=0x381883x)

[2014-08-05 12:52:44.607604] W [client-rpc-fops.c:2624:client3_3_lookup_cbk] 0-HA-fast-150G-PVE1-client-1: remote operation failed: Transport endpoint is not connected. Path: / (00000000-0000-0000-0000-000000000001)

[2014-08-05 12:52:44.607736] E [rpc-clnt.c:368:saved_frames_unwind] (-->/usr/lib/x86_64-linux-gnu/libgfrpc.so.0(rpc_clnt_notify+0xf8) [0x7fcb1b4b0558] (-->/usr/lib/x86_64-linux-gnu/libgfrpc.so.0(rpc_clnt_connection_cleanup+0xc3) [0x7fcb1b4aea63] (-->/usr/lib/x86_64-linux-gnu/libgfrpc.so.0(saved_frames_destroy+0xe) [0x7fcb1b4ae97e]))) 0-HA-fast-150G-PVE1-client-1: forced unwinding frame type(GlusterFS Handshake) op(PING(3)) called at 2014-08-05 12:52:42.463891 (xid=0x381884x)

[2014-08-05 12:52:44.607753] W [client-handshake.c:276:client_ping_cbk] 0-HA-fast-150G-PVE1-client-1: timer must have expired

[2014-08-05 12:52:44.607776] I [client.c:2098:client_rpc_notify] 0-HA-fast-150G-PVE1-client-1: disconnected

I've got SSD disks (just for an info).

Should I go and give a try for 3.5.2?

2014-08-05 13:06 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

reply along with gluster-users please :-). May be you are hitting 'reply' instead of 'reply all'?

Pranith

On 08/05/2014 03:35 PM, Roman wrote:

To make sure and clean, I've created another VM with raw format and goint to repeat those steps. So now I've got two VM-s one with qcow2 format and other with raw format. I will send another e-mail shortly.

2014-08-05 13:01 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

On 08/05/2014 03:07 PM, Roman wrote:

really, seems like the same file

stor1:

a951641c5230472929836f9fcede6b04  /exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

stor2:

a951641c5230472929836f9fcede6b04  /exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

one thing I've seen from logs, that somehow proxmox VE is connecting with wrong version to servers?

[2014-08-05 09:23:45.218550] I [client-handshake.c:1659:select_server_supported_programs] 0-HA-fast-150G-PVE1-client-0: Using Program GlusterFS 3.3, Num (1298437), Version (330)

It is the rpc (over the network data structures) version, which is not changed at all from 3.3 so thats not a problem. So what is the conclusion? Is your test case working now or not?

Pranith

but if I issue:

root@pve1:~# glusterfs -V

glusterfs 3.4.4 built on Jun 28 2014 03:44:57

seems ok.

server  use 3.4.4 meanwhile

[2014-08-05 09:23:45.117875] I [server-handshake.c:567:server_setvolume] 0-HA-fast-150G-PVE1-server: accepted client from stor1-9004-2014/08/05-09:23:45:93538-HA-fast-150G-PVE1-client-1-0 (version: 3.4.4)

[2014-08-05 09:23:49.103035] I [server-handshake.c:567:server_setvolume] 0-HA-fast-150G-PVE1-server: accepted client from stor1-8998-2014/08/05-09:23:45:89883-HA-fast-150G-PVE1-client-0-0 (version: 3.4.4)

if this could be the reason, of course.

I did restart the Proxmox VE yesterday (just for an information)

2014-08-05 12:30 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

On 08/05/2014 02:33 PM, Roman wrote:

Waited long enough for now, still different sizes and no logs about healing :(

stor1 

# file: exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

trusted.afr.HA-fast-150G-PVE1-client-0=0x000000000000000000000000

trusted.afr.HA-fast-150G-PVE1-client-1=0x000000000000000000000000

trusted.gfid=0xf10ad81b58484bcd9b385a36a207f921

root@stor1:~# du -sh /exports/fast-test/150G/images/127/

1.2G    /exports/fast-test/150G/images/127/

stor2

# file: exports/fast-test/150G/images/127/vm-127-disk-1.qcow2

trusted.afr.HA-fast-150G-PVE1-client-0=0x000000000000000000000000

trusted.afr.HA-fast-150G-PVE1-client-1=0x000000000000000000000000

trusted.gfid=0xf10ad81b58484bcd9b385a36a207f921

root@stor2:~# du -sh /exports/fast-test/150G/images/127/

1.4G    /exports/fast-test/150G/images/127/

According to the changelogs, the file doesn't need any healing. Could you stop the operations on the VMs and take md5sum on both these machines?

Pranith

2014-08-05 11:49 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

On 08/05/2014 02:06 PM, Roman wrote:

Well, it seems like it doesn't see the changes were made to the volume ? I created two files 200 and 100 MB (from /dev/zero) after I disconnected the first brick. Then connected it back and got these logs:

[2014-08-05 08:30:37.830150] I [glusterfsd-mgmt.c:1584:mgmt_getspec_cbk] 0-glusterfs: No change in volfile, continuing

[2014-08-05 08:30:37.830207] I [rpc-clnt.c:1676:rpc_clnt_reconfig] 0-HA-fast-150G-PVE1-client-0: changing port to 49153 (from 0)

[2014-08-05 08:30:37.830239] W [socket.c:514:__socket_rwv] 0-HA-fast-150G-PVE1-client-0: readv failed (No data available)

[2014-08-05 08:30:37.831024] I [client-handshake.c:1659:select_server_supported_programs] 0-HA-fast-150G-PVE1-client-0: Using Program GlusterFS 3.3, Num (1298437), Version (330)

[2014-08-05 08:30:37.831375] I [client-handshake.c:1456:client_setvolume_cbk] 0-HA-fast-150G-PVE1-client-0: Connected to 10.250.0.1:49153, attached to remote volume '/exports/fast-test/150G'.

[2014-08-05 08:30:37.831394] I [client-handshake.c:1468:client_setvolume_cbk] 0-HA-fast-150G-PVE1-client-0: Server and Client lk-version numbers are not same, reopening the fds

[2014-08-05 08:30:37.831566] I [client-handshake.c:450:client_set_lk_version_cbk] 0-HA-fast-150G-PVE1-client-0: Server lk version = 1

[2014-08-05 08:30:37.830150] I [glusterfsd-mgmt.c:1584:mgmt_getspec_cbk] 0-glusterfs: No change in volfile, continuing

this line seems weird to me tbh.

I do not see any traffic on switch interfaces between gluster servers, which means, there is no syncing between them.

I tried to ls -l the files on the client and servers to trigger the healing, but seems like no success. Should I wait more?

Yes, it should take around 10-15 minutes. Could you provide 'getfattr -d -m. -e hex <file-on-brick>' on both the bricks.

Pranith

2014-08-05 11:25 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

On 08/05/2014 01:10 PM, Roman wrote:

Ahha! For some reason I was not able to start the VM anymore, Proxmox VE told me, that it is not able to read the qcow2 header due to permission is denied for some reason. So I just deleted that file and created a new VM. And the nex message I've got was this:

Seems like these are the messages where you took down the bricks before self-heal. Could you restart the run waiting for self-heals to complete before taking down the next brick?

Pranith

[2014-08-05 07:31:25.663412] E [afr-self-heal-common.c:197:afr_sh_print_split_brain_log] 0-HA-fast-150G-PVE1-replicate-0: Unable to self-heal contents of '/images/124/vm-124-disk-1.qcow2' (possible split-brain). Please delete the file from all but the preferred subvolume.- Pending matrix:  [ [ 0 60 ] [ 11 0 ] ]

[2014-08-05 07:31:25.663955] E [afr-self-heal-common.c:2262:afr_self_heal_completion_cbk] 0-HA-fast-150G-PVE1-replicate-0: background  data self-heal failed on /images/124/vm-124-disk-1.qcow2

2014-08-05 10:13 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

I just responded to your earlier mail about how the log looks. The log comes on the mount's logfile

Pranith

On 08/05/2014 12:41 PM, Roman wrote:

Ok, so I've waited enough, I think. Had no any traffic on switch ports between servers. Could not find any suitable log message about completed self-heal (waited about 30 minutes). Plugged out the other server's UTP cable this time and got in the same situation:

root@gluster-test1:~# cat /var/log/dmesg

-bash: /bin/cat: Input/output error

brick logs:

[2014-08-05 07:09:03.005474] I [server.c:762:server_rpc_notify] 0-HA-fast-150G-PVE1-server: disconnecting connectionfrom pve1-27649-2014/08/04-13:27:54:720789-HA-fast-150G-PVE1-client-0-0

[2014-08-05 07:09:03.005530] I [server-helpers.c:729:server_connection_put] 0-HA-fast-150G-PVE1-server: Shutting down connection pve1-27649-2014/08/04-13:27:54:720789-HA-fast-150G-PVE1-client-0-0

[2014-08-05 07:09:03.005560] I [server-helpers.c:463:do_fd_cleanup] 0-HA-fast-150G-PVE1-server: fd cleanup on /images/124/vm-124-disk-1.qcow2

[2014-08-05 07:09:03.005797] I [server-helpers.c:617:server_connection_destroy] 0-HA-fast-150G-PVE1-server: destroyed connection of pve1-27649-2014/08/04-13:27:54:720789-HA-fast-150G-PVE1-client-0-0

2014-08-05 9:53 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

Do you think it is possible for you to do these tests on the latest version 3.5.2? 'gluster volume heal <volname> info' would give you that information in versions > 3.5.1.
Otherwise you will have to check it from either the logs, there will be self-heal completed message on the mount logs (or) by observing 'getfattr -d -m. -e hex <image-file-on-bricks>'

Pranith

On 08/05/2014 12:09 PM, Roman wrote:

Ok, I understand. I will try this shortly.

How can I be sure, that healing process is done, if I am not able to see its status?

2014-08-05 9:30 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

Mounts will do the healing, not the self-heal-daemon. The problem I feel is that whichever process does the healing has the latest information about the good bricks in this usecase. Since for VM usecase, mounts should have the latest information, we should let the mounts do the healing. If the mount accesses the VM image either by someone doing operations inside the VM or explicit stat on the file it should do the healing.

Pranith.

On 08/05/2014 10:39 AM, Roman wrote:

Hmmm, you told me to turn it off. Did I understood something wrong? After I issued the command you've sent me, I was not able to watch the healing process, it said, it won't be healed, becouse its turned off.

2014-08-05 5:39 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

You didn't mention anything about self-healing. Did you wait until the self-heal is complete?

Pranith

On 08/04/2014 05:49 PM, Roman wrote:

Hi!

Result is pretty same. I set the switch port down for 1st server, it was ok. Then set it up back and set other server's port off. and it triggered IO error on two virtual machines: one with local root FS but network mounted storage. and other with network root FS. 1st gave an error on copying to or from the mounted network disk, other just gave me an error for even reading log.files.

cat: /var/log/alternatives.log: Input/output error

then I reset the kvm VM and it said me, there is no boot device. Next I virtually powered it off and then back on and it has booted.

By the way, did I have to start/stop volume?

>> Could you do the following and test it again?

>> gluster volume set <volname> cluster.self-heal-daemon off

>>Pranith

2014-08-04 14:10 GMT+03:00 Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx>:

On 08/04/2014 03:33 PM, Roman wrote:

Hello!

Facing the same problem as mentioned here:

http://supercolony.gluster.org/pipermail/gluster-users/2014-April/039959.html

my set up is up and running, so i'm ready to help you back with feedback.

setup:

proxmox server as client

2 gluster physical  servers

server side and client side both running atm 3.4.4 glusterfs from gluster repo.

the problem is:

1. craeted replica bricks.

2. mounted in proxmox (tried both promox ways: via GUI and fstab (with backup volume line), btw while mounting via fstab I'm unable to launch a VM without cache, meanwhile direct-io-mode is enabled in fstab line)

3. installed VM

4. bring one volume down - ok

5. bringing up, waiting for sync is done.

6. bring other volume down - getting IO errors on VM guest and not able to restore the VM after I reset the VM via host. It says (no bootable media). After I shut it down (forced) and bring back up, it boots.

Could you do the following and test it again?
gluster volume set <volname> cluster.self-heal-daemon off

Pranith

Need help. Tried 3.4.3, 3.4.4.

Still missing pkg-s for 3.4.5 for debian and 3.5.2 (3.5.1 always gives a healing error for some reason)

--
Best regards,
Roman.

_______________________________________________
Gluster-users mailing list
Gluster-users@xxxxxxxxxxx
http://supercolony.gluster.org/mailman/listinfo/gluster-users

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

--
Best regards,
Roman.

_______________________________________________
Gluster-users mailing list
Gluster-users@xxxxxxxxxxx
http://supercolony.gluster.org/mailman/listinfo/gluster-users