On Fri, Apr 23, 2021 at 1:12 PM Boris Behrens <bb@kervyn.de> wrote:

Am Fr., 23. Apr. 2021 um 13:00 Uhr schrieb Ilya Dryomov <idryomov@gmail.com>:

...

On Fri, Apr 23, 2021 at 12:46 PM Boris Behrens <bb@kervyn.de> wrote:

...

Am Fr., 23. Apr. 2021 um 12:16 Uhr schrieb Ilya Dryomov <idryomov@gmail.com>:

...

On Fri, Apr 23, 2021 at 12:03 PM Boris Behrens <bb@kervyn.de> wrote:

...

Am Fr., 23. Apr. 2021 um 11:52 Uhr schrieb Ilya Dryomov <idryomov@gmail.com>:

...

This snippet confirms my suspicion. Unfortunately without a verbose log from that VM from three days ago (i.e. when it got into this state) it's hard to tell what exactly went wrong.

The problem is that the VM doesn't consider itself to be the rightful owner of the lock and so when "rbd snap create" requests the lock from it in order to make a snapshot, the VM just ignores the request because even though it owns the lock, its record appears to be of sync.

I'd suggest to kick it by restarting osd36. If the VM is active, it should reacquire the lock and hopefully update its internal record as expected. If "rbd snap create" still hangs after that, it would mean that we have a reproducer and can gather logs on the VM side.

What version of qemu/librbd and ceph is in use (both on the VM side and on the side you are running "rbd snap create"?

I just stopped the OSD, waited some seconds and started it again. I still can't create snapshots.

Ceph version is 14.2.18 accross the board qemu is 4.1.0-1 as we use krbd, the kernel version is 5.2.9-arch1-1-ARCH

How can I gather more logs to debug it?

Are you saying that this image is mapped and the lock is held by the kernel client? It doesn't look that way from the logs you shared.

We use krbd instead of librbd (at least this is what I think I know), but qemu is doing the kvm/rbd stuff.

I'm going to assume that by "qemu is doing the kvm/rbd stuff", you mean that you are using the librbd driver inside qemu and that this image is opened by qemu (i.e. that driver). If you don't know what access method is being used, debugging this might be challenging ;)

Let's start with the same output: "rbd lock ls", "rbd status" and "rbd snap create --debug-ms=1 --debug-rbd=20". It should be different after osd36 was restarted.

Here is the new one: https://pastebin.com/6qTsJK6W Ah ok, this CPU node still got the old thing and uses librbd to work with rbd instead of krbd.

Sorry, I forgot that simply restarting the OSD doesn't trigger the code path that I'm hoping would cause librbd inside the VM to update its state. I took a look at the code and I think there are a couple of ways to do it (listed in the order of preference): - cut the network between the VM and the cluster for more than 30 seconds; it should be done externally so that to the VM it looks like a long network blip - stop the VM process for more than 30 seconds $ PID=<pid of qemu-system-x86_64 process> $ kill -STOP $PID && sleep 40 && kill -CONT $PID - stop the osd36 process for more than 30 seconds with "nodown" flag set $ ceph osd set nodown $ PID=<pid of osd36 process> $ kill -STOP $PID && sleep 40 && kill -CONT $PID $ ceph osd unset nodown Thanks, Ilya