Re: update_balloon_size_func blocked for more than 120 seconds

[David Hildenbrand]

On Thu, Nov 11, 2021 at 11:49 PM Luis Chamberlain <mcgrof@xxxxxxxxxx> wrote:
>
> I get the following splats with a kvm guest in idle, after a few seconds
> it starts:
>
> [  242.412806] INFO: task kworker/6:2:271 blockedfor more than 120 seconds.
> [  242.415790]       Tainted: G            E     5.15.0-next-20211111 #68
> [  242.417755] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> [  242.418332] task:kworker/6:2     state:D stack:    0 pid:  271 ppid: 2 flags:0x00004000
> [  242.418954] Workqueue: events_freezable update_balloon_size_func [virtio_balloon]
> [  242.419518] Call Trace:
> [  242.419709]  <TASK>
> [  242.419873]  __schedule+0x2fd/0x990
> [  242.420142]  schedule+0x4e/0xc0
> [  242.420382]  tell_host+0xaa/0xf0 [virtio_balloon]
> [  242.420757]  ? do_wait_intr_irq+0xa0/0xa0
> [  242.421065]  update_balloon_size_func+0x2c9/0x2e0 [virtio_balloon]
> [  242.421527]  process_one_work+0x1e5/0x3c0
> [  242.421833]  worker_thread+0x50/0x3b0
> [  242.422204]  ? rescuer_thread+0x370/0x370
> [  242.422507]  kthread+0x169/0x190
> [  242.422754]  ? set_kthread_struct+0x40/0x40
> [  242.423073]  ret_from_fork+0x1f/0x30
> [  242.423347]  </TASK>
>
> And this goes on endlessly. The last one says it blocked for more than 1208
> seconds. This was not happening until the last few weeks but I see no
> relevant recent commits for virtio_balloon, so the related change could
> be elsewhere.

We're stuck somewhere in:

wq: update_balloon_size_func()->fill_balloon()->tell_host()

Most probably in wait_event().


I am no waitqueue expert, but my best guess would be that we're
waiting more than 2 minutes
on a host reply with TASK_UNINTERRUPTIBLE. At least that's my interpretation,

In case we're really stuck for more than 2 minutes, the hypervisor
might not be processing our
requests anymore -- or it's not getting processed for some other reason (or the
waitqueue is not getting woken up due do some other BUG).

IIUC, we can sleep longer via wait_event_interruptible(), TASK_UNINTERRUPTIBLE
seems to be the issue that triggers the warning. But by changing that
might just be hiding the fact that
we're waiting more than 2 minutes on a reply.

>
> I could bisect but first I figured I'd check to see if someone already
> had spotted this.

Bisecting would be awesome, then we might at least know if this is a
guest or a hypervisor issue.

Note that the environment matters: the hypervisor seems to be
requesting the guest to inflate
the balloon right when booting up. So you might not be able to
reproduce in a different environment.