Bug#791944: /etc/init.d/sendsigs kills systemd-udevd upon shutdown, causing dmsetup to hang

[Felipe Sateler]

On Mon, Jun 5, 2017 at 2:32 PM, Pali Rohár <pali.rohar at gmail.com> wrote:
> On Monday 05 June 2017 20:23:52 Michael Biebl wrote:
>> Am 05.06.2017 um 13:12 schrieb Pali Rohár:
>> >> Am 05.06.2017 um 09:59 schrieb Pali Rohár:
>> >>> /etc/rc0.d/K08cryptdisks
>> >>> /etc/rc0.d/K09cryptdisks-early
>> >>> /etc/rc0.d/K10udev
>> >
>> > Problem is that in K08 and K09 phase is udev not running, but
>> > socket is present. udev is probably killed by
>> > /etc/rc0.d/K03sendsigs which is before K08 and K09. Above patch
>> > modify K10 which is *after* K08 and K09, therefore it has
>> > absolutely no effect on fixing this issue.
>>
>> Ah, ok. The info about sendsigs being ordered before udev on shutdown
>> was missing.
>
> Anyway, if you are patching K10udev and whole freeze happens in
> K08cryptdisks and K09cryptdisks-early, then such patch is noop as whole
> freeze happens before patch line is executed. sendsigs does not change
> behaviour of patch.
>
>> I guess we need both then:
>> - udev using the sendsigs.omit.d interface to avoid being killed by
>> /etc/init.d/sendsigs
>> and
>> - removing the socket on stop and making sure udev is stopped in 0 /
>> 6.
>>
>> Anyone willing to work on a patch?
>
> I think the first thing which needs to be done is to define order
> sendsigs, cryptdisks and udev. Should be udev really stopped after
> cryptdisks? Or not? If not, then udev should be stopped *before*
> sendsigs. Otherwise sendsigs needs to be patched to not kill udev.


If the udev pid is added to sendsigs.omit.d as noted by Michael then
sendsigs should stop killing udev. So no need to patch sendsigs.

>
> But other question is what happens in system when udev is still running
> even after unmounting filesystem and when basically all other remaining
> services were already killed by sendsigs...

Udev would still be stopped, just later. I note that under systemd
udev is not stopped until the final killing spree, so hopefully this
delay in stopping will not introduce problems.

-- 

Saludos,
Felipe Sateler