Re: [Xen-devel] linux 4.4 Regression: 100% cpu usage on idle pv guest under Xen with single vcpu.

[Boris Ostrovsky <boris.ostrovsky@oracle.com>]

On 12/01/2015 06:30 PM, Sander Eikelenboom wrote:
> On 2015-12-02 00:19, Boris Ostrovsky wrote:
>> On 12/01/2015 06:00 PM, Sander Eikelenboom wrote:
>>> On 2015-12-01 23:47, Boris Ostrovsky wrote:
>>>> On 11/30/2015 05:55 PM, Sander Eikelenboom wrote:
>>>>> On 2015-11-30 23:54, Boris Ostrovsky wrote:
>>>>>> On 11/30/2015 04:46 PM, Sander Eikelenboom wrote:
>>>>>>> On 2015-11-30 22:45, Konrad Rzeszutek Wilk wrote:
>>>>>>>> On Sat, Nov 28, 2015 at 04:47:43PM +0100, Sander Eikelenboom 
>>>>>>>> wrote:
>>>>>>>>> Hi all,
>>>>>>>>>
>>>>>>>>> I have just tested a 4.4-rc2 kernel (current linus tree) + the 
>>>>>>>>> tip tree
>>>>>>>>> pulled on top.
>>>>>>>>>
>>>>>>>>> Running this kernel under Xen on PV-guests with multiple vcpus 
>>>>>>>>> goes well (on
>>>>>>>>> idle < 10% cpu usage),
>>>>>>>>> but a guest with only a single vcpu doesn't idle at all, it 
>>>>>>>>> seems a kworker
>>>>>>>>> thread is stuck:
>>>>>>>>> root       569 98.0  0.0      0     0 ?        R 16:02 12:47
>>>>>>>>> [kworker/0:1]
>>>>>>>>>
>>>>>>>>> Running a 4.3 kernel works fine with a single vpcu, bisecting 
>>>>>>>>> would probably
>>>>>>>>> quite painful since there were some breakages this merge 
>>>>>>>>> window with respect
>>>>>>>>> to Xen pv-guests.
>>>>>>>>>
>>>>>>>>> There are some differences in the diff's from booting a 4.3, 
>>>>>>>>> 4.4-single,
>>>>>>>>> 4.4-multi cpu boot:
>>>>>>>>
>>>>>>>> Boris has been tracking a bunch of them. I am attaching the 
>>>>>>>> latest set of
>>>>>>>> patches I've to carry on top of v4.4-rc3.
>>>>>>>
>>>>>>> Hi Konrad,
>>>>>>>
>>>>>>> i will test those, see if it fixes all my issues and report back
>>>>>>
>>>>>> They shouldn't help you ;-( (and I just saw a message from you 
>>>>>> confirming this)
>>>>>>
>>>>>> The first one fixes a 32-bit bug (on bare metal too). The second 
>>>>>> fixes
>>>>>> a fatal bug for 32-bit PV guests. The other two are code
>>>>>> improvements/cleanup.
>>>>>
>>>>> One of these patches also fixes a bug i was having with a 
>>>>> pci-passthrough device in
>>>>> a HVM that wasn't working (depending on which dom0-kernel i was 
>>>>> using (4.3 or 4.4)),
>>>>> but didn't report yet.
>>>>>
>>>>> Fingers crossed but i think this pv-guest single vcpu issue is the 
>>>>> last i'm troubled by for now ;)
>>>>
>>>> I could not reproduce this, including with your kernel config file.
>>>
>>> Hmm that's unpleasant :-\
>>>
>>> Hmm other strange thing is it doesn't seem to affect dom0 (which is 
>>> also a PV guest), but only unprivileged ones
>>> All unprivileged pv-guests seem to have the irq issue, but only with 
>>> a single vcpu i see to get the stuck kworker thread that got my 
>>> attention, with a 2 vcpu that doesn't seem to happen, but you still 
>>> get the dmesg output and warnings about hvc)
>>>
>>> Could it be that:
>>>
>>> arch/x86/include/asm/i8259.h
>>> static inline int nr_legacy_irqs(void)
>>> {
>>>         return legacy_pic->nr_legacy_irqs;
>>> }
>>>
>>> returns something different in some circumstances ?
>>
>> It should return 16 pre-8c058b0b9c34d8c8d7912880956543769323e2d8 and 0
>> after that commit.
>>
>> This is the last number that you see in
>>     NR_IRQS:4352 nr_irqs:48 0
>> line.
>>
>> I think you should be able to safely revert both
>> b4ff8389ed14b849354b59ce9b360bdefcdbf99c and
>> 8c058b0b9c34d8c8d7912880956543769323e2d8 and see if it makes any
>> difference.
>>
>>
>> -boris
>>
>
> That was already underway compiling :)
>
> And it does reveal that reverting both fixes the issue, no stuck 
> kworker thread .. and no:
>    genirq: Flags mismatch irq 8. 00000000 (hvc_console) vs. 00000000 
> (rtc0)
>    hvc_open: request_irq failed with rc -16.


Let me try it again tomorrow. Can you post your guest config file, Xen 
version and host HW (Intel or AMD)? 'xl info' maybe?

-boris


>
> What i did get was an conflict reverting 
> b4ff8389ed14b849354b59ce9b360bdefcdbf99c:
> arch/arm64/include/asm/irq.h, although that shouldn't matter because 
> we are on x86 and not on arm.
>
> -- 
> Sander
>
>
>>>
>>> -- Sander
>>>
>>>>
>>>> -boris
>>>
>>> _______________________________________________
>>> Xen-devel mailing list
>>> Xen-devel@lists.xen.org
>>> http://lists.xen.org/xen-devel