Hide Forgot
Description of problem: Ran 'service clvmd stop' and even though it reported that it was stopped... it wasn't. Deactivating VG vg: 0 logical volume(s) in volume group "vg" now active [ OK ] Stopping clvm: [ OK ] type level name id state fence 0 default 00010004 none [1 2 3 4 5] dlm 1 clvmd 00020004 none [3 4] This then will cause the cman init script to fail as well since the clvmd service is still registered. Version-Release number of selected component (if applicable): [root@morph-04 ~]# uname -ar Linux morph-04 2.6.17-1.2519.4.14.el5 #1 SMP Thu Aug 24 19:31:14 EDT 2006 i686 i686 i386 GNU/Linux [root@morph-04 ~]# rpm -q lvm2-cluster lvm2-cluster-2.02.09-1.0.RHEL5 How reproducible: fairly reproducable
'service clvmd stop' does stop the clvmd daemon. Try the following after 'service clvmd stop' [root@morph-05 ~]# ps -ef | grep clvmd root 2542 2520 0 13:22 pts/1 00:00:00 grep clvmd [root@morph-05 ~]# service clvmd status clvmd is stopped active volumes: (none) But 'cman_tool services' still shows clvmd bound to the dlm. [root@morph-05 ~]# cman_tool services type level name id state fence 0 default 00010004 none [1 2 3 4 5] dlm 1 clvmd 00020004 none [5]
FWIW, also reproduced this on x86_64.
There's a FIXME in the dlm to add AUTOFREE like in RHEL4. Not sure if we need to implement that or if libdlm can do without it. [I've largely worked out how to add it but would like to avoid it if possible since it's a little unpleasant adding more special cases for user lockspaces to the normal paths.]
Under most circumstances clvmd should close the lockspace when it shuts down. If this isn't happening then we need to know how clvmd is being stopped. Killing it -9 will obviously do this! The library can't implement autofree because it doesn't know who else has the lockspace open (only the kernel does) - libraries live in the same process address space as the calling program and are no more privileged (what do you think this is? VMS?). In any case that wouldn't help as the clvmd lockspace is not marked AUTOFREE anyway - that's for the default lockspace and is never used by anthing else. One reason why it hasn't been added: it's not really needed. There's nothing wrong with leaving the clvmd lockspace open when the daemon isn't running (or there shouldn't be), it can happen under the same circumstances on RHCS4. If it's preventing shutdown then we need to look at why it's doing so - because it shouldn't.
Oops, I should have checked clvmd more carefully before assuming it wanted to use autofree. It looks like doing a 'cman_tool leave' should cause clvmd to release the lockspace. The question is then how the clvmd/cman init scripts should shut down a node. It looks like we're currently doing: > service clvmd stop > service cman stop the clvmd daemon exits in the first step so it never gets the shutdown callback from cman_tool leave in the second step, so it never releasing the lockspace, unless I'm still confused about all this. Perhaps the clvmd daemon should remain running after the first step and then when clvmd gets a shutdown callback from the second step, it would release the lockspace and exit.
Dave, that sounds completely backwards. I would expect to see no traces of clvmd after running "service clvmd stop." It should release its lockspace before clvmd exits.
If clvmd is sent a SIGTERM or SIGINT then it should shut down cleanly taking the lockspace with it. Only a SIGKILL (or a crash) should leave the lockspace open.
What I'm saying here I suppose is that if clvmd is being killed -KILL then don't, If it's crashing then I want to know where (either a core file, strace or similar), If it's being killed -TERM but still leaving the lockspace open then there might be something in the DLM that's failing the remove_lockspace call - I'm not sure. Looking at the information in comment #1 it seems that this is not that common (on a 5 node cluster it has removed the lockspace correctly on 3 or 4 nodes). My additional comment in #4 also still stands. Why is this a problem ?
OK that wasn't hard, so far. clvmd is calling dlm_release_lockspace, and I can see i nthe log for dlm_controld that it is making an attempt to remove the lockspace. but it doesn't go away: 1156861126 groupd callback: finish clvmd (unused) 1156861129 groupd callback: stop clvmd 1156861129 write "0" to "/sys/kernel/dlm/clvmd/control" 1156861129 groupd callback: start clvmd count 3 members 19 16 14 1156861129 dir_member 19 1156861129 dir_member 18 1156861129 dir_member 14 1156861129 dir_member 16 1156861129 set_members rmdir "/sys/kernel/config/dlm/cluster/spaces/clvmd/nodes/ 18" 1156861129 write "1" to "/sys/kernel/dlm/clvmd/control" 1156861129 groupd callback: finish clvmd (unused) 1156861130 groupd callback: stop clvmd 1156861130 write "0" to "/sys/kernel/dlm/clvmd/control" 1156861130 groupd callback: start clvmd count 2 members 16 14 1156861130 dir_member 19 1156861130 dir_member 14 1156861130 dir_member 16 1156861130 set_members rmdir "/sys/kernel/config/dlm/cluster/spaces/clvmd/nodes/ 19" 1156861130 write "1" to "/sys/kernel/dlm/clvmd/control" 1156861130 groupd callback: finish clvmd (unused) group_tool shows the lockspace still in existance, and the directory /sys/kernel/config/dlm/cluster/spaces/clvmd/nodes still exists (with node numbers 14 and 16 (us) in it.
Created attachment 135144 [details] daemon logs. nodes 14 & 16 failed to remove the lockspace
It looks like the dlm was probably doing recovery on 14 and 16. We'd want to look at kernel messages and info under /sys/kernel/dlm for hints about why dlm recovery wasn't completing.
Created attachment 135151 [details] dmesg logs & sysfs dlm files
sequence of events that should be taking place: - clvmd calls libdlm: dlm_release_lockspace() - libdlm writes DLM_USER_REMOVE_LOCKSPACE to dlm control device - dlm-kernel: device_remove_lockspace() - dlm-kernel: dlm_release_lockspace() - dlm-kernel: do_uevent() does "offline" uevent that should be seen by dlm_controld and reported in debug output: uevent: offline@/kernel/dlm/clvmd - dlm-kernel: do_uevent() waits for ack from dlm_controld through sysfs - dlm_controld leaves the group, after which no "clvmd" group should appear in group_tool's output - dlm_controld writes to /sys/kernel/dlm/clvmd/event to tell the kernel that the lockspace has been left - do_uevent() returns - dlm_release_lockspace() cleans up the lockspace in the kernel - device_remove_lockspace() returns - libdlm write on control device returns - libdlm dlm_release_lockspace() returns - clvmd continues Looking at the logs, it's evident that dlm_controld on 14 & 16 have not yet gotten a uevent from the kernel. It would be interesting to see where the clvmd process is waiting and whether libdlm has called into the kernel yet for removing the lockspace.
clvmd has finished up cleanly and gone home. I'm 98% certain that the release lockspace calls has completed to the library's satisfaction at least because I can see no abnormal finish in the clvmd logs and all the locks have been tidily released.
dlm-kernel was ignoring the FORCEFREE flag from the lib and not sending a force value to dlm_release_lockspace(), so any locks in the lockspace were causing dlm_release_lockspace() to fail and return EBUSY. Patch sent.
I can confirm that this fixes the bug for me, at least.
The cluster-5.0 flag is going away....
A package has been built which should help the problem described in this bug report. This report is therefore being closed with a resolution of CURRENTRELEASE. You may reopen this bug report if the solution does not work for you.
Moving all RHCS ver 5 bugs to RHEL 5 so we can remove RHCS v5 which never existed.