[Pacemaker] [Problem] The attrd does not sometimes stop.

[Dejan Muhamedagic]

Hi,

On Mon, Nov 14, 2011 at 01:17:37PM +0100, Lars Ellenberg wrote:
> On Mon, Nov 14, 2011 at 11:58:09AM +1100, Andrew Beekhof wrote:
> > On Mon, Nov 7, 2011 at 8:39 AM, Lars Ellenberg
> > <lars.ellenberg at linbit.com> wrote:
> > > On Thu, Nov 03, 2011 at 01:49:46AM +1100, Andrew Beekhof wrote:
> > >> On Tue, Oct 18, 2011 at 12:19 PM,  <renayama19661014 at ybb.ne.jp> wrote:
> > >> > Hi,
> > >> >
> > >> > We sometimes fail in a stop of attrd.
> > >> >
> > >> > Step1. start a cluster in 2 nodes
> > >> > Step2. stop the first node.(/etc/init.d/heartbeat stop.)
> > >> > Step3. stop the second node after time passed a little.(/etc/init.d/heartbeat
> > >> > stop.)
> > >> >
> > >> > The attrd catches the TERM signal, but does not stop.
> > >>
> > >> There's no evidence that it actually catches it, only that it is sent.
> > >> I've seen it before but never figured out why it occurs.
> > >
> > > I had it once tracked down almost to where it occurs, but then got distracted.
> > > Yes the signal was delivered.
> > >
> > > I *think* it had to do with attrd doing a blocking read,
> > > or looping in some internal message delivery function too often.
> > >
> > > I had a quick look at the code again now, to try and remember,
> > > but I'm not sure.
> > >
> > > I *may* be that, because
> > > xmlfromIPC(IPC_Channel * ch, int timeout) calls
> > >    msg = msgfromIPC_timeout(ch, MSG_ALLOWINTR, timeout, &ipc_rc);
> > >
> > > And MSG_ALLOWINTR will cause msgfromIPC_ll() to
> > >        IPC_INTR:
> > >                if ( allow_intr){
> > >                        goto startwait;
> > >
> > > Depending on the frequency of deliverd signals, it may cause this goto
> > > startwait loop to never exit, because the timeout always starts again
> > > from the full passed in timeout.
> > >
> > > If only one signal is deliverd, it may still take 120 seconds
> > > (MAX_IPC_DELAY from crm.h) to be actually processed, as the signal
> > > handler only raises a flag for the next mainloop iteration.
> > >
> > > If a (non-fatal) signal is delivered every few seconds,
> > > then the goto loop will never timeout.
> > >
> > > Please someone check this for plausibility ;-)
> > 
> > Most plausible explanation I've heard so far... still odd that only
> > attrd is affected.
> > So what do we do about it?
> 
> Reproduce, and confirm that this is what people are seeing.
> 
> Make attrd non-blocking?
> 
> Fix the ipc layer to not restart the full timeout,
> but only the remaining partial time?

Lars and I made a quick patch for cluster-glue (attached).
Hideo-san, is there a way for you to verify if it helps? The
patch is not perfect and under unfavourable circumstances it may
still take a long time for the caller to exit, but it'd be good
to know if this is the right spot.

Cheers,

Dejan

> -- 
> : Lars Ellenberg
> : LINBIT | Your Way to High Availability
> : DRBD/HA support and consulting http://www.linbit.com
> 
> _______________________________________________
> Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> 
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cl_msg_timeout.patch
Type: text/x-diff
Size: 1230 bytes
Desc: not available
URL: <https://lists.clusterlabs.org/pipermail/pacemaker/attachments/20111114/296dec66/attachment-0003.bin>