Re: Early ACPI events prevent subsequent ACPI functionality on xen 4.3 + HVM domU

[Diana Crisan <dcrisan@flexiant.com>]

George,

On 21/05/13 17:09, George Dunlap wrote:
> On 05/21/2013 04:59 PM, Alex Bligh wrote:
>> George,
>>
>> --On 21 May 2013 16:23:00 +0100 George Dunlap
>> <george.dunlap@eu.citrix.com> wrote:
>>
>>> On 05/21/2013 04:16 PM, Alex Bligh wrote:
>>>> George,
>>>>
>>>> --On 21 May 2013 14:39:55 +0100 George Dunlap
>>>> <George.Dunlap@eu.citrix.com> wrote:
>>>>
>>>>> So this appears to be an xl toolstack thing.  I managed to reproduce
>>>>> your results using "xl shutdown -F [domain]"; but if you then do "xl
>>>>> trigger [domian] power", the domain shuts down as normal.
>>>>
>>>> OK. But doesn't the power thing yank the power rather than send a
>>>> clean shutdown?
>>>
>>> No -- if you push the button just once on most modern hardware it will
>>> send an ACPI "poweroff" event that the OS handles gracefully. That's
>>> what gets sent when you do "xl trigger [domain] power".  If the OS
>>> ignores it (either on real hardware or virtual hardware) nothing 
>>> happens.
>>> On real hardware you have to then hold down the button for 5 seconds 
>>> for
>>> a hard-shutdown, with xl you have to do "xl destroy".
>>
>> OK, great. I am guessing the reason why 'xl shutdown' doesn't do
>> that is to cope with (a) non-HVM domains, and (b) old fashioned
>> HVM domains with PV support but not ACPI support. Correct?
>>
>>>> We are using libxl here (admittedly having looked carefully at
>>>> the xl code for guidance) and get the same problem.
>>>
>>> The relevant xl code is here:
>>>
>>>      rc=libxl_domain_shutdown(ctx, domid);
>>>      if (rc == ERROR_NOPARAVIRT) {
>>>          if (fallback_trigger) {
>>>              fprintf(stderr, "PV control interface not available:"
>>>                      " sending ACPI power button event.\n");
>>>              rc = libxl_send_trigger(ctx, domid, 
>>> LIBXL_TRIGGER_POWER, 0);
>>>          } else {
>>>              fprintf(stderr, "PV control interface not available:"
>>>                      " external graceful shutdown not possible.\n");
>>>              fprintf(stderr, "Use \"-F\" to fallback to ACPI power
>>> event.\n");
>>>          }
>>>      }
>>>
>>> It looks like libxl_domain_shutdown() will only use the PV shutdown
>>> method.  If that method is not available, then it will call
>>> libxl_send_trigger().
>>
>> OK. Well if that's the case, why is it not working? i.e. how
>> is using 'xl trigger power' a workaround? As our testing was
>> on xl (as well as libxl). Perhaps libxl_domain_shutdown
>> is returning an error other than ERROR_NOPARAVIRT or no error
>> at all?
>
> xl/libxl is working as designed. xl tried libxl_domain_shutdown() 
> first.  libxl saw that there were PV drivers available, so it sent the 
> PV shutdown signal over xenstore and returned success.  libxl and xl 
> have no way of knowing that the signal was never received, so it never 
> falls back to ACPI.
>
> "xl trigger power" is a work-around because when you detect that a 
> guest is stuck, you should be able to issue an ACPI shutdown and get a 
> clean shutdown, even when the PV shutdown path is stuck.
>
> (In fact, as a short-term solution, you might consider just replacing 
> "xl shutdown $d" with "xl trigger $d power" in your control framework.)
>

I am having problems replicating the short-term solution reliably on our 
end, even with xl/libxl. It seems that issuing an "xl trigger dom power" 
does not make the domain shutdown, even after a 10 minute wait. 
Subsequent ones tend to do the same and after issuing quite a few I 
managed to get the domU to power off.

This happens even if I do not issue the shutdown early in the boot 
process, but rather issue directly an xl trigger dom power.

>> Does xl trigger reset do the same thing (i.e. a graceful shutdown)?
>> If not we'd also (I think) have to recode reboots to do a power trigger
>> then a restart.
>
> Hmm, this is what I get:
>
> # xl trigger h0 reset
> libxl: error: libxl.c:4639:libxl_send_trigger: Send trigger 'reset' 
> failed: Function not implemented
>
> YMMV...
>
>  -George

--
Diana Crisan