The reproducer is not detailed enough for me to reproduce anything (mind, I'm a VDSM ignoramus).  Even if it had more detail, a reproducer that works several layers up the stack requires me to peel off the layers myself to get to the layer I'm supposed to debug.  I commonly need the reporter's help to do that when it's more than one layer.

Your libvirt logs *might* be enough for me to make sense of the bug.  Please provide a reproducer in terms qemu-kvm, or at least your libvirt logs.

QE cann't reproduce this bug. This is the detailed steps as below. please let me know if my steps have any problems.

1) enabled hugepage on host
#echo 10240 >  /sys/devices/system/node/node0/hugepages/hugepages-2048kB/nr_hugepages

2) Boot RHEL7.6 guest with hugepage, This is part of xml file.

<memory unit='KiB'>4194304</memory>
  <currentMemory unit='KiB'>4194304</currentMemory>
  <memoryBacking>
    <hugepages>
      <page size='2048' unit='KiB'/>
    </hugepages>
    <locked/>
  </memoryBacking>

3. run io stress script inside guest.

#cat io-stress.sh
dd if=/dev/random of=file1 iflag=fullblock bs=1M count=2000 &
dd if=/dev/random of=file2 iflag=fullblock bs=1M count=2000 &
dd if=/dev/random of=file3 iflag=fullblock bs=1M count=2000 &
dd if=/dev/random of=file4 iflag=fullblock bs=1M count=2000 &
dd if=/dev/random of=file5 iflag=fullblock bs=1M count=2000 &

4. Run continuously the qemu monitor commands
while true;do virsh qemu-monitor-command rhel7 --hmp "info qtree";done

5. shutdown guest on host
#virsh destroy rhel7

Result:

step 4 & 5 work well.

(In reply to Markus Armbruster from comment #2)
> The reproducer is not detailed enough for me to reproduce anything (mind,
> I'm a VDSM ignoramus).  Even if it had more detail, a reproducer that works
> several layers up the stack requires me to peel off the layers myself to get
> to the layer I'm supposed to debug.  I commonly need the reporter's help to
> do that when it's more than one layer.
> 
> Your libvirt logs *might* be enough for me to make sense of the bug.  Please
> provide a reproducer in terms qemu-kvm, or at least your libvirt logs.

Sorry Markus, I thought my original bug on RHV will help in understanding the issue.

I am giving a reproducer steps on a KVM directly (No RHV).

Steps to Reproduce:

[1] I used slowfs (https://github.com/nirs/slowfs) on the directory where the image is located to get some delay to the write operation. The value I set is "0.001" for the "write". So VIR_DOMAIN_DESTROY_GRACEFUL will take some time to complete as it will flush the i/o and that would take some time if I have slow storage.

[2] Started a while loop to execute qemu-monitor command every second.

while true;do virsh qemu-monitor-command rhel7.5 --hmp "info qtree";date;sleep 1;done

[3] In the VM, ran the stress script mentioned in comment #3.

[4] When I have large dirty pages in the VM, ran "virsh destroy rhel7.5 --graceful"

[5] The qemu-monitor got stuck immediately after I run the destroy and got exited only when the domain was destroyed successfully.

===
The query which was called before the "destroy" got completed successfully.

2019-01-16 13:58:25.548+0000: 32959: info : qemuMonitorIOWrite:551 : QEMU_MONITOR_IO_WRITE: mon=0x7f49e0017930 buf={"execute":"human-monitor-command","arguments":{"command-line":"info qtree"},"id":"libvirt-254"}^M
2019-01-16 13:58:25.552+0000: 32963: debug : qemuDomainObjExitMonitorInternal:7074 : Exited monitor (mon=0x7f49e0017930 vm=0x7f4994193c40 name=rhel7.5)
2019-01-16 13:58:25.552+0000: 32963: debug : qemuDomainObjEndJob:6935 : Stopping job: query (async=none vm=0x7f4994193c40 name=rhel7.5)

However, the next one got blocked.

===
2019-01-16 13:58:26.570+0000: 32964: info : qemuMonitorSend:1083 : QEMU_MONITOR_SEND_MSG: mon=0x7f49e0017930 msg={"execute":"human-monitor-command","arguments":{"command-line":"info qtree"},"id":"libvirt-255"}^M
2019-01-16 13:58:26.570+0000: 32959: info : qemuMonitorIOWrite:551 : QEMU_MONITOR_IO_WRITE: mon=0x7f49e0017930 buf={"execute":"human-monitor-command","arguments":{"command-line":"info qtree"},"id":"libvirt-255"}^M

destroy called...

2019-01-16 13:58:26.751+0000: 32960: debug : virDomainDestroyFlags:524 : dom=0x7f49f0001360, (VM: name=rhel7.5, uuid=85ea4436-0295-474c-83d8-9c85ca7dafae), flags=0x1


2019-01-16 13:58:47.080+0000: 32962: debug : qemuDomainObjExitMonitorInternal:7074 : Exited monitor (mon=0x7f49e0017930 vm=0x7f4994193c40 name=rhel7.5)
2019-01-16 13:58:47.080+0000: 32962: debug : qemuDomainObjEndJob:6935 : Stopping job: query (async=none vm=0x7f4994193c40 name=rhel7.5)
===

So during the "destroy", the monitor is not responding to the commands and this is the issue we are facing here and would like to know if this can be improved.

In RHV, there are many periodic tasks running these monitor commands. The issue was observed in a customer environment where the "destroy" of a VM took 59 seconds which caused the monitoring queue in RHV to get full and made all the VMs "not responding" and this was reported in bug 1660451.


Attaching the libvirtd debug log.

Created attachment 1521112 [details]
libvirt log

QE cann't reproduce this bug yet. This is detailed steps. please correct me if my steps is wrong, Thanks.

1. Boot RHEL7.6 guest with virsh.

2. run slowfs inside guest e.g

#slowfs -c slowfs.cfg.example /realfs /slowfs

3. run io stress script seems comment3.

#./io-stress.sh

4.Started a while loop to execute qemu-monitor command every second.

#while true;do virsh qemu-monitor-command rhel7 --hmp "info qtree";date;sleep 1;done

5."virsh destroy rhel7 --graceful"

result:

step5 is successful.

Highlight: The "destroy command" doesn't have any delay(response is very fast) in my testing. I can't reproduce it like comment4(e.g it took 59 seconds...)

(In reply to FuXiangChun from comment #6)
> QE cann't reproduce this bug yet. This is detailed steps. please correct me
> if my steps is wrong, Thanks.
> 
> 1. Boot RHEL7.6 guest with virsh.
> 
> 2. run slowfs inside guest e.g
> 
> #slowfs -c slowfs.cfg.example /realfs /slowfs
> 

You have to run slowfs on the directory where the image file is located. I also played with parameters "vm.dirty_background_ratio", "vm.dirty_ratio" and "vm.dirty_expire_centisecs" within the VM to get a large cache  before shutdown.