[Pacemaker] heartbeat stop hangs sometimes

[Andrew Beekhof]

On Mon, Feb 22, 2010 at 5:10 PM, Lars Ellenberg
<lars.ellenberg at linbit.com> wrote:
> On Mon, Feb 22, 2010 at 01:00:29PM +0100, Markus M. wrote:
>> Hello,
>>
>> sometimes "heartbeat stop" seems to hang (latest packets from
>> clusterlabs.org, RHEL5 x86_64, 2-node cluster with only one node
>> running).
>>
>> The last lines from ha-debug are like this:
>>
>> Feb 22 12:52:48 dbprod21 ccm: [24053]: info: client (pid=24058) removed from ccm
>> Feb 22 12:52:48 dbprod21 crmd: [24058]: info: do_ha_control: Disconnected from Heartbeat
>> Feb 22 12:52:48 dbprod21 crmd: [24058]: info: do_cib_control: Disconnecting CIB
>> Feb 22 12:52:48 dbprod21 cib: [24054]: info: cib_process_readwrite: We are now in R/O mode
>> Feb 22 12:52:48 dbprod21 crmd: [24058]: info: crmd_cib_connection_destroy: Connection to the CIB terminated...
>> Feb 22 12:52:48 dbprod21 cib: [24054]: WARN: send_ipc_message: IPC Channel to 24058 is not connected
>> Feb 22 12:52:48 dbprod21 crmd: [24058]: info: do_exit: Performing A_EXIT_0 - gracefully exiting the CRMd
>> Feb 22 12:52:48 dbprod21 cib: [24054]: WARN: send_via_callback_channel: Delivery of reply to client 24058/d9c9c281-4f38-46d8-b83e-54135f6c75e9 failed
>> Feb 22 12:52:48 dbprod21 crmd: [24058]: info: free_mem: Dropping I_TERMINATE: [ state=S_STOPPING cause=C_FSA_INTERNAL origin=do_stop ]
>> Feb 22 12:52:48 dbprod21 cib: [24054]: WARN: do_local_notify: A-Sync reply to crmd failed: reply failed
>> Feb 22 12:52:48 dbprod21 crmd: [24058]: info: do_exit: [crmd] stopped (0)
>> Feb 22 12:52:48 dbprod21 heartbeat: [24040]: info: killing /usr/lib64/heartbeat/attrd process group 24057 with signal 15
>
> Yep.
> I've seen this too,
> a few times.
> Apparently attrd sometimes "ignores" a term signal.
> Once an additional signal comes in, or any message is processed, i.e.
> once the mainloop() actually _processes_ the signal,
> it is handled and attrd dies.
>
> Unfortunately I don't see exactly where this signal is "lost", though.
> The signal is delivered, the flag is raised, mainloop should recognize
> and handle it...
> And I don't yet have a reliable way to reproduce it, either.
> If you have, let us know!
>
> Maybe the following helps (sorry, patch is likely not whitespace clean)
>
> diff -r 1a6d0f690c3e lib/common/mainloop.c
> --- a/lib/common/mainloop.c     Thu Feb 18 22:36:49 2010 +0100
> +++ b/lib/common/mainloop.c     Mon Feb 22 17:09:31 2010 +0100
> @@ -191,7 +191,12 @@
>     CRM_ASSERT(sizeof(crm_signal_t) > sizeof(GSource));
>     source = g_source_new(&crm_signal_funcs, sizeof(crm_signal_t));
>
> -    crm_signals[sig] = (crm_signal_t*)mainloop_setup_trigger(source, G_PRIORITY_HIGH, NULL, NULL);
> +    crm_signals[sig] = (crm_signal_t*)mainloop_setup_trigger(source,
> +                   /* TERM is higher priority than other signals,
> +                    * signals are higher priority than other ipc.
> +                    * yes, minus: smaller is "higher". */
> +                   G_PRIORITY_HIGH - (sig == SIGTERM ? 2 : 1),
> +                   NULL, NULL);
>     CRM_ASSERT(crm_signals[sig] != NULL);
>
>     crm_signals[sig]->handler = dispatch;

I've applied a similar patch to stable.
Also in stable is a patch that waits up to 2.5 minutes for post-crmd
clients to terminate.
So either way we should have this resolved.