Re: [Qemu-devel] Re: kvm: savevm is broken for me

[Glauber Costa <glommer@redhat.com>]

On Fri, Jul 24, 2009 at 09:45:28PM +0200, Alexander Graf wrote:
>
> On 24.07.2009, at 21:42, Glauber Costa wrote:
>
>> On Fri, Jul 24, 2009 at 04:27:38PM +0200, Alexander Graf wrote:
>>>
>>> On 24.07.2009, at 16:20, Luiz Capitulino wrote:
>>>
>>>> On Fri, 24 Jul 2009 16:18:02 +0200
>>>> Alexander Graf <agraf@suse.de> wrote:
>>>>
>>>>>
>>>>> On 24.07.2009, at 15:54, Luiz Capitulino wrote:
>>>>>
>>>>>> On Fri, 24 Jul 2009 00:06:13 +0200
>>>>>> Alexander Graf <agraf@suse.de> wrote:
>>>>>>
>>>>>>> On 23.07.2009, at 23:58, Luiz Capitulino wrote:
>>>>>>>
>>>>>>>>
>>>>>>>> Hi there,
>>>>>>>>
>>>>>>>> If I try to 'savevm' with latest Anthony's tree (HEAD
>>>>>>>> 6f725c139ae975646c44ace77bf796318a5783da) QEMU will hang and I
>>>>>>>> know
>>>>>>>> it's not saving anything because vmstat shows no disk activity
>>>>>>>> (which
>>>>>>>> is the opposite behavior of when 'savevm' works).
>>>>>>>>
>>>>>>>> According to 'git bisect' the week's winner is:
>>>>>>>>
>>>>>>>> """
>>>>>>>> commit bd8367761236cd5c435598aeb2f1b8240c09b059
>>>>>>>> Author: Alexander Graf <agraf@suse.de>
>>>>>>>> Date:   Fri Jul 17 13:51:48 2009 +0200
>>>>>>>>
>>>>>>>> Fake dirty loggin when it's not there
>>>>>>>> """
>>>>>>>>
>>>>>>>> Indeed, reverting this makes 'savevm' work for me.
>>>>>>>
>>>>>>>
>>>>>>> Yikes. I have no idea why, but just revert the patch :-). I don't
>>>>>>> really need it anymore since we now have dirty logging in PPC.
>>>>>>>
>>>>>>> Maybe it's the if (is dirty logging really enabled?) in the
>>>>>>> beginning?
>>>>>>
>>>>>> Yes, it's the if. I have removed it and savevm works.
>>>>>>
>>>>>> Can you submit the fix please?
>>>>>
>>>>> Eh - what fix? You found that something doesn't set or evaluate the
>>>>> flags right. That still doesn't tell us who behaves incorrectly.
>>>>
>>>> Ah, okay. I thought you meant the 'if' _was_ the bug.
>>>>
>>>> Better to revert then?
>>>
>>> Even better yet to find out who's doing something wrong :-).
>>
>> It's actually quite simple.
>> you continue'd instead of break'ing during the dirty log loop,
>> which would put us in an infinite loop
>>
>> Patch follows:
>
> But who calls kvm_physical_sync_dirty_bitmap with !(mem->flags &  
> KVM_MEM_LOG_DIRTY_PAGES)? Sounds rather wrong to me.
I'd have to dig, but from code inspection,
it seems that ram_save_live does it, in its first interaction.

This can probably be improved, but I don't think it is wrong semantics
anyway. We can sync dirty bitmaps without having mem->flags & KVM_MEM_LOG_DIRTY_PAGES,
and just got nothing written to qemu dirty bitmap.

Anyway, even if we decide to fix ram_save_live to not do it,
getting into an endless loop in that case is buggy. So it is really orthogonal.