On Fri, Jan 22, 2021 at 2:24 PM Michal Privoznik <mprivozn@xxxxxxxxxx> wrote:
On 1/22/21 12:09 PM, Nikolay Shirokovskiy wrote:
> On Fri, Jan 22, 2021 at 12:45 PM Michal Privoznik <mprivozn@xxxxxxxxxx>
> wrote:
>
>> If libvirtd is sent SIGTERM while it is still initializing, it
>> may crash. The following scenario was observed (using 'stress' to
>> slow down CPU so much that the window where the problem exists is
>> bigger):
>>
>> 1) The main thread is already executing virNetDaemonRun() and is
>> in virEventRunDefaultImpl().
>> 2) The thread that's supposed to run daemonRunStateInit() is
>> spawned already, but daemonRunStateInit() is in its very early
>> stage (in the stack trace I see it's executing
>> virIdentityGetSystem()).
>>
>> If SIGTERM (or any other signal that we don't override handler
>> for) arrives at this point, the main thread jumps out from
>> virEventRunDefaultImpl() and enters virStateShutdownPrepare()
>> (via shutdownPrepareCb which was set earlier). This iterates
>> through stateShutdownPrepare() callbacks of state drivers and
>> reaching qemuStateShutdownPrepare() eventually only to
>> dereference qemu_driver. But since thread 2) has not been
>> scheduled/not proceeded yet, qemu_driver was not allocated yet.
>>
>> Solution is simple - just check if qemu_driver is not NULL. But
>> doing so only in qemuStateShutdownPrepare() would push the
>> problem down to virStateShutdownWait(), well
>> qemuStateShutdownWait(). Therefore, duplicate the trick there
>> too.
>>
>
> I guess this is a partial solution. Initialization may be in a state when
> qemu_driver is initialized but qemu_driver->workerPool is still NULL
> for example.
Yes.
>
> Maybe we'd better delay shutdown until initialization is finished?
I'm not exactly sure how to achieve that. Do you have a hint? Also, part
of qemu driver state init is autostarting domains (which may take ages).
I'm thinking of adding a new variable 'initialized' to virnetdaemon. It can be
set by call from daemonRunStateInit after initialization is finished.
And we should call shutdownPrepare/shutdownWait only when both 'quit' and
'initialized' are true. It will require another pipe pair probably to wake up
event loop thread when 'initialized' it set to true.
As initialization can take a lot of time (autostart as you mentioned) we
can arm finishTimer right when quit is set without waiting for initialization
to finish. This way we exit ungracefully just as in other cases when shutdown
finishing takes too much time. Libvirtd has a goal to handle crashes at
any time so this exit should be fine.
Nikolay
