[ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
Hi guys, ocassionally stopping a VirtualDomain resource via "crm resource stop" does not work, and in the end the node is fenced, which is ugly. I had a look at the RA to see what it does. After trying to stop the domain via "virsh shutdown ..." in a configurable time it switches to "virsh destroy". i assume "virsh destroy" send a sigkill to the respective process. But when the host is doing heavily IO it's possible that the process is in "D" state (uninterruptible sleep) in which it can't be finished with a SIGKILL. The the node the domain is running on is fenced due to that. I digged deeper and found out that the signal is often delivered a bit later (just some seconds) and the process is killed, but pacemaker already decided to fence the node. It's all about this excerp in the RA: force_stop() { local out ex translate local status=0 ocf_log info "Issuing forced shutdown (destroy) request for domain ${DOMAIN_NAME}." out=$(LANG=C virsh $VIRSH_OPTIONS destroy ${DOMAIN_NAME} 2>&1) ex=$? translate=$(echo $out|tr 'A-Z' 'a-z') echo >&2 "$translate" case $ex$translate in *"error:"*"domain is not running"*|*"error:"*"domain not found"*|\ *"error:"*"failed to get domain"*) : ;; # unexpected path to the intended outcome, all is well [!0]*) ocf_exit_reason "forced stop failed" return $OCF_ERR_GENERIC ;; 0*) while [ $status != $OCF_NOT_RUNNING ]; do VirtualDomain_status status=$? done ;; esac return $OCF_SUCCESS } I'm thinking about the following: How about to let the script wait a bit after "virsh destroy". I saw that usually it just takes some seconds that "virsh destroy" is successfull. I'm thinking about this change: ocf_log info "Issuing forced shutdown (destroy) request for domain ${DOMAIN_NAME}." out=$(LANG=C virsh $VIRSH_OPTIONS destroy ${DOMAIN_NAME} 2>&1) ex=$? sleep (10)< (or maybe configurable) translate=$(echo $out|tr 'A-Z' 'a-z') What do you think ? Bernd -- Bernd Lentes Systemadministration Institute for Metabolism and Cell Death (MCD) Building 25 - office 122 HelmholtzZentrum München bernd.len...@helmholtz-muenchen.de phone: +49 89 3187 1241 phone: +49 89 3187 3827 fax: +49 89 3187 2294 http://www.helmholtz-muenchen.de/mcd stay healthy Helmholtz Zentrum München Helmholtz Zentrum Muenchen Deutsches Forschungszentrum fuer Gesundheit und Umwelt (GmbH) Ingolstaedter Landstr. 1 85764 Neuherberg www.helmholtz-muenchen.de Aufsichtsratsvorsitzende: MinDir.in Prof. Dr. Veronika von Messling Geschaeftsfuehrung: Prof. Dr. med. Dr. h.c. Matthias Tschoep, Kerstin Guenther Registergericht: Amtsgericht Muenchen HRB 6466 USt-IdNr: DE 129521671 ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
22.10.2020 23:29, Lentes, Bernd пишет: > Hi guys, > > ocassionally stopping a VirtualDomain resource via "crm resource stop" does > not work, and in the end the node is fenced, which is ugly. > I had a look at the RA to see what it does. After trying to stop the domain > via "virsh shutdown ..." in a configurable time it switches to "virsh > destroy". > i assume "virsh destroy" send a sigkill to the respective process. But when > the host is doing heavily IO it's possible that the process is in "D" state > (uninterruptible sleep) > in which it can't be finished with a SIGKILL. The the node the domain is > running on is fenced due to that. > I digged deeper and found out that the signal is often delivered a bit later > (just some seconds) and the process is killed, but pacemaker already decided > to fence the node. > It's all about this excerp in the RA: > > force_stop() > { > local out ex translate > local status=0 > > ocf_log info "Issuing forced shutdown (destroy) request for domain > ${DOMAIN_NAME}." > out=$(LANG=C virsh $VIRSH_OPTIONS destroy ${DOMAIN_NAME} 2>&1) > ex=$? > translate=$(echo $out|tr 'A-Z' 'a-z') > echo >&2 "$translate" > case $ex$translate in > *"error:"*"domain is not running"*|*"error:"*"domain not > found"*|\ > *"error:"*"failed to get domain"*) > : ;; # unexpected path to the intended outcome, all > is well > [!0]*) > ocf_exit_reason "forced stop failed" > return $OCF_ERR_GENERIC ;; > 0*) > while [ $status != $OCF_NOT_RUNNING ]; do > VirtualDomain_status > status=$? > done ;; > esac > return $OCF_SUCCESS > } > > I'm thinking about the following: > How about to let the script wait a bit after "virsh destroy". I saw that > usually it just takes some seconds that "virsh destroy" is successfull. > I'm thinking about this change: > > ocf_log info "Issuing forced shutdown (destroy) request for domain > ${DOMAIN_NAME}." > out=$(LANG=C virsh $VIRSH_OPTIONS destroy ${DOMAIN_NAME} 2>&1) > ex=$? > sleep (10)< (or maybe configurable) > translate=$(echo $out|tr 'A-Z' 'a-z') > > > What do you think ? > It makes no difference. You wait 10 seconds before parsing output of "virsh destroy", that's all. It does not change output itself, so if output indicates that "virsh destroy" failed, it will still indicate that after 10 seconds. Either you need to repeat "virsh destroy" in a loop, or virsh itself should be more robust. ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
- On Oct 23, 2020, at 7:11 AM, Andrei Borzenkov arvidj...@gmail.com wrote: >> >> ocf_log info "Issuing forced shutdown (destroy) request for domain >> ${DOMAIN_NAME}." >> out=$(LANG=C virsh $VIRSH_OPTIONS destroy ${DOMAIN_NAME} 2>&1) >> ex=$? >> sleep (10)< (or maybe configurable) >> translate=$(echo $out|tr 'A-Z' 'a-z') >> >> >> What do you think ? >> > > > It makes no difference. You wait 10 seconds before parsing output of > "virsh destroy", that's all. It does not change output itself, so if > output indicates that "virsh destroy" failed, it will still indicate > that after 10 seconds. > > Either you need to repeat "virsh destroy" in a loop, or virsh itself > should be more robust. Hi Andrei, yes, you are right. I saw it alread after sending the E-Mail. I will change that. Bernd Helmholtz Zentrum München Helmholtz Zentrum Muenchen Deutsches Forschungszentrum fuer Gesundheit und Umwelt (GmbH) Ingolstaedter Landstr. 1 85764 Neuherberg www.helmholtz-muenchen.de Aufsichtsratsvorsitzende: MinDir.in Prof. Dr. Veronika von Messling Geschaeftsfuehrung: Prof. Dr. med. Dr. h.c. Matthias Tschoep, Kerstin Guenther Registergericht: Amtsgericht Muenchen HRB 6466 USt-IdNr: DE 129521671 ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
- On Oct 23, 2020, at 5:06 PM, Strahil Nikolov hunter86...@yahoo.com wrote: > why don't you work with something like this: 'op stop interval =300 > timeout=600'. > The stop operation will timeout at your requirements without modifying the > script. > > Best Regards, > Strahil Nikolov But when the timeout has run out the RA tries to kill the machine with a "virsh destroy". And if that does not work (what is occasionally my problem) because the domain is in uninterruptable sleep (D state) the RA gives a $OCF_ERR_GENERIC back which cause pacemaker to fence the lazy node. Or am i wrong ? Where is the benefit of the shorter interval ? The return value of the "virsh destroy" operation is set immediately. And it's -ne 0 when the "virsh destroy" didn't suceed. No matter if the domain stops 20 sec. later, the return value is not changed. and send to the LRM so the cluster wants to stonith that node. Surprisingly if the virsh destroy is successfull the RA waits until the domain isn't running anymore: force_stop { ... 0*) while [ $status != $OCF_NOT_RUNNING ]; do VirtualDomain_status status=$? done ;; I need someting like that which waits for some time (maybe 30s) if the domain nevertheless stops although "virsh destroy" gaves an error back. Because the SIGKILL is delivered if the process wakes up from D state. For this amount of time the RA has to wait and to take care that the the return value is zero if the domain stopped or is -ne 0 if also the waiting didn't help. Bernd Helmholtz Zentrum München Helmholtz Zentrum Muenchen Deutsches Forschungszentrum fuer Gesundheit und Umwelt (GmbH) Ingolstaedter Landstr. 1 85764 Neuherberg www.helmholtz-muenchen.de Aufsichtsratsvorsitzende: MinDir.in Prof. Dr. Veronika von Messling Geschaeftsfuehrung: Prof. Dr. med. Dr. h.c. Matthias Tschoep, Kerstin Guenther Registergericht: Amtsgericht Muenchen HRB 6466 USt-IdNr: DE 129521671 ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
why don't you work with something like this: 'op stop interval =300 timeout=600'. The stop operation will timeout at your requirements without modifying the script. Best Regards, Strahil Nikolov В четвъртък, 22 октомври 2020 г., 23:30:08 Гринуич+3, Lentes, Bernd написа: Hi guys, ocassionally stopping a VirtualDomain resource via "crm resource stop" does not work, and in the end the node is fenced, which is ugly. I had a look at the RA to see what it does. After trying to stop the domain via "virsh shutdown ..." in a configurable time it switches to "virsh destroy". i assume "virsh destroy" send a sigkill to the respective process. But when the host is doing heavily IO it's possible that the process is in "D" state (uninterruptible sleep) in which it can't be finished with a SIGKILL. The the node the domain is running on is fenced due to that. I digged deeper and found out that the signal is often delivered a bit later (just some seconds) and the process is killed, but pacemaker already decided to fence the node. It's all about this excerp in the RA: force_stop() { local out ex translate local status=0 ocf_log info "Issuing forced shutdown (destroy) request for domain ${DOMAIN_NAME}." out=$(LANG=C virsh $VIRSH_OPTIONS destroy ${DOMAIN_NAME} 2>&1) ex=$? translate=$(echo $out|tr 'A-Z' 'a-z') echo >&2 "$translate" case $ex$translate in *"error:"*"domain is not running"*|*"error:"*"domain not found"*|\ *"error:"*"failed to get domain"*) : ;; # unexpected path to the intended outcome, all is well [!0]*) ocf_exit_reason "forced stop failed" return $OCF_ERR_GENERIC ;; 0*) while [ $status != $OCF_NOT_RUNNING ]; do VirtualDomain_status status=$? done ;; esac return $OCF_SUCCESS } I'm thinking about the following: How about to let the script wait a bit after "virsh destroy". I saw that usually it just takes some seconds that "virsh destroy" is successfull. I'm thinking about this change: ocf_log info "Issuing forced shutdown (destroy) request for domain ${DOMAIN_NAME}." out=$(LANG=C virsh $VIRSH_OPTIONS destroy ${DOMAIN_NAME} 2>&1) ex=$? sleep (10) < (or maybe configurable) translate=$(echo $out|tr 'A-Z' 'a-z') What do you think ? Bernd -- Bernd Lentes Systemadministration Institute for Metabolism and Cell Death (MCD) Building 25 - office 122 HelmholtzZentrum München bernd.len...@helmholtz-muenchen.de phone: +49 89 3187 1241 phone: +49 89 3187 3827 fax: +49 89 3187 2294 http://www.helmholtz-muenchen.de/mcd stay healthy Helmholtz Zentrum München Helmholtz Zentrum Muenchen Deutsches Forschungszentrum fuer Gesundheit und Umwelt (GmbH) Ingolstaedter Landstr. 1 85764 Neuherberg www.helmholtz-muenchen.de Aufsichtsratsvorsitzende: MinDir.in Prof. Dr. Veronika von Messling Geschaeftsfuehrung: Prof. Dr. med. Dr. h.c. Matthias Tschoep, Kerstin Guenther Registergericht: Amtsgericht Muenchen HRB 6466 USt-IdNr: DE 129521671 ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/ ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
23.10.2020 21:08, Lentes, Bernd пишет: > > Surprisingly if the virsh destroy is successfull the RA waits until the > domain isn't running anymore: > ... > > I need someting like that which waits for some time (maybe 30s) if the domain > nevertheless stops although > "virsh destroy" gaves an error back. Because the SIGKILL is delivered if the > process wakes up from D state. So why not ignore virsh error and just wait always? You probably need to retain "domain not found" exit condition still. ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
- On Oct 23, 2020, at 8:45 PM, Valentin Vidić vvi...@valentin-vidic.from.hr wrote: > On Fri, Oct 23, 2020 at 08:08:31PM +0200, Lentes, Bernd wrote: >> But when the timeout has run out the RA tries to kill the machine with a >> "virsh >> destroy". >> And if that does not work (what is occasionally my problem) because the >> domain >> is in uninterruptable sleep (D state) the RA gives a $OCF_ERR_GENERIC back >> which >> cause pacemaker to fence the lazy node. Or am i wrong ? > > What does the log look like when this happens? > /var/log/cluster/corosync.log: VirtualDomain(vm_amok)[8998]: 2020/09/27_22:34:11 INFO: Issuing graceful shutdown request for domain vm_amok. VirtualDomain(vm_amok)[8998]: 2020/09/27_22:37:06 INFO: Issuing forced shutdown (destroy) request for domain vm_amok. Sep 27 22:37:11 [11282] ha-idg-2 lrmd: warning: child_timeout_callback: vm_amok_stop_0 process (PID 8998) timed out Sep 27 22:37:11 [11282] ha-idg-2 lrmd: warning: operation_finished: vm_amok_stop_0:8998 - timed out after 18ms timeout of the domain is 180 sec. /var/log/libvirt/libvirtd.log (time is UTC): 2020-09-27 20:37:21.489+: 18583: error : virProcessKillPainfully:401 : Failed to terminate process 14037 with SIGKILL: Device or resource busy 2020-09-27 20:37:21.505+: 6610: error : virNetSocketWriteWire:1852 : Cannot write data: Broken pipe 2020-09-27 20:37:31.962+: 6610: error : qemuMonitorIO:719 : internal error: End of file from qemu monitor SIGKILL didn't work. Nevertheless the process is finished 20 seconds later after destroy, surely because it woke up from D and received the signal. /var/log/cluster/corosync.log on the DC: Sep 27 22:37:11 [3580] ha-idg-1 crmd: warning: status_from_rc: Action 93 (vm_amok_stop_0) on ha-idg-2 failed (target: 0 vs. rc: 1): Error Stop (also sigkill) failed Sep 27 22:37:11 [3579] ha-idg-1pengine: notice: native_stop_constraints: Stop of failed resource vm_amok is implicit after ha-idg-2 is fenced cluster decides to fence the node although resource is stopped 10 seconds later atop log: 14037 - S 261% /usr/bin/qemu-system-x86_64 -machine accel=kvm -name guest=vm_amok,debug-threads=on -S -object secret,id=masterKey0 ... PID of the domain is 14037 14037 - E 0% worker (at 22:37:31) domain has stoppped Bernd Helmholtz Zentrum München Helmholtz Zentrum Muenchen Deutsches Forschungszentrum fuer Gesundheit und Umwelt (GmbH) Ingolstaedter Landstr. 1 85764 Neuherberg www.helmholtz-muenchen.de Aufsichtsratsvorsitzende: MinDir.in Prof. Dr. Veronika von Messling Geschaeftsfuehrung: Prof. Dr. med. Dr. h.c. Matthias Tschoep, Kerstin Guenther Registergericht: Amtsgericht Muenchen HRB 6466 USt-IdNr: DE 129521671 ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
- On Oct 23, 2020, at 11:11 PM, arvidjaar arvidj...@gmail.com wrote: >> I need someting like that which waits for some time (maybe 30s) if the domain >> nevertheless stops although >> "virsh destroy" gaves an error back. Because the SIGKILL is delivered if the >> process wakes up from D state. > > So why not ignore virsh error and just wait always? You probably need to > retain "domain not found" exit condition still. That's my plan. Bernd Helmholtz Zentrum München Helmholtz Zentrum Muenchen Deutsches Forschungszentrum fuer Gesundheit und Umwelt (GmbH) Ingolstaedter Landstr. 1 85764 Neuherberg www.helmholtz-muenchen.de Aufsichtsratsvorsitzende: MinDir.in Prof. Dr. Veronika von Messling Geschaeftsfuehrung: Prof. Dr. med. Dr. h.c. Matthias Tschoep, Kerstin Guenther Registergericht: Amtsgericht Muenchen HRB 6466 USt-IdNr: DE 129521671 ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
On Fri, Oct 23, 2020 at 08:08:31PM +0200, Lentes, Bernd wrote: > But when the timeout has run out the RA tries to kill the machine with a > "virsh destroy". > And if that does not work (what is occasionally my problem) because the domain > is in uninterruptable sleep (D state) the RA gives a $OCF_ERR_GENERIC back > which > cause pacemaker to fence the lazy node. Or am i wrong ? What does the log look like when this happens? -- Valentin ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
Re: [ClusterLabs] VirtualDomain does not stop via "crm resource stop" - modify RA ?
- On Oct 23, 2020, at 11:18 PM, Bernd Lentes bernd.len...@helmholtz-muenchen.de wrote: > - On Oct 23, 2020, at 11:11 PM, arvidjaar arvidj...@gmail.com wrote: > > >>> I need someting like that which waits for some time (maybe 30s) if the >>> domain >>> nevertheless stops although >>> "virsh destroy" gaves an error back. Because the SIGKILL is delivered if the >>> process wakes up from D state. >> >> So why not ignore virsh error and just wait always? You probably need to >> retain "domain not found" exit condition still. Hi, here is my rewritten RA: https://hmgubox2.helmholtz-muenchen.de/index.php/s/iYjRyJiWb5XNfXm As i'm not a coder feedback is welcome. Bernd Helmholtz Zentrum München Helmholtz Zentrum Muenchen Deutsches Forschungszentrum fuer Gesundheit und Umwelt (GmbH) Ingolstaedter Landstr. 1 85764 Neuherberg www.helmholtz-muenchen.de Aufsichtsratsvorsitzende: MinDir.in Prof. Dr. Veronika von Messling Geschaeftsfuehrung: Prof. Dr. med. Dr. h.c. Matthias Tschoep, Kerstin Guenther Registergericht: Amtsgericht Muenchen HRB 6466 USt-IdNr: DE 129521671 ___ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/