Re: [RFC/query] kvm async_pf anon pined pages migration

[Gu Zheng <guz.fnst@cn.fujitsu.com>]

On 10/10/2013 04:01 PM, Gleb Natapov wrote:

> On Thu, Oct 10, 2013 at 03:53:08PM +0800, Gu Zheng wrote:
>> Hi Gleb,
>>
>> On 10/10/2013 03:15 PM, Gleb Natapov wrote:
>>
>>> On Thu, Oct 10, 2013 at 03:05:58PM +0800, chai wen wrote:
>>>> On 10/08/2013 03:39 PM, Gleb Natapov wrote:
>>>>> On Tue, Oct 08, 2013 at 02:58:22PM +0800, chai wen wrote:
>>>>>> On 10/02/2013 12:04 AM, chaiwen wrote:
>>>>>>> On 09/30/2013 08:51 PM, Gleb Natapov wrote:
>>>>>>>> On Mon, Sep 30, 2013 at 06:03:07PM +0800, chai wen wrote:
>>>>>>>>> Hi all
>>>>>>>>>
>>>>>>>>> Async page fault in kvm currently pin user pages via get_user_pages.
>>>>>>>>> when doing page migration,the method can be found via
>>>>>>>>> page->mmapping->a_ops->migratepage to offline old pages and migrate to
>>>>>>>>> new pages. As to anonymous page there is no file mapping but a anon_vma.So
>>>>>>>>> the migration will fall back to some *default* migration method.Anon pages
>>>>>>>>> that have been pined in memory by some reasons could be failed in the migration
>>>>>>>>> processing because of some reasons like ref-count checking.
>>>>>>>>> (or I misunderstand some thing?)
>>>>>>>>>
>>>>>>>>> Now we want to make these anon pages in async_pf can be migrated, I try some
>>>>>>>>> ways.But there are still many problems. The following is one that replaceing
>>>>>>>>> the mapping of anon page arbitrarily and doing some thing based on it.
>>>>>>>>> Kvm-based virtual machine can works on this patch,but have no experience of
>>>>>>>>> offline pages because of the limitaion of resouces.I'll check it later.
>>>>>>>>>
>>>>>>>>> I don't know weather it is a right direction of this issue.
>>>>>>>>> All comments/criticize are welcomed.
>>>>>>>> The pinning is not mandatory and can (and probably should) be dropped, but
>>>>>>>> pinning that is done by async page faults is short lived. What problems
>>>>>>>> are you seeing that warrant the complexity of handling their migration?
>>>>>> Hi Gleb
>>>>>>
>>>>>> As to this issue, I still have some thing not very clear.
>>>>>> If pages pinning is successfully holding (although not mandatory) by
>>>>>> async page fault.
>>>>>> And at the same time page migration happens because of memory
>>>>>> hot-remove action.
>>>>>> It has 120*hz timeout setting in common page offline processing,
>>>>>> could it fail with
>>>>>> these async_pf pined pages migration ?
>>>>>> What's your opinion about this ?   If it may fail under this
>>>>>> circumstance, should we do
>>>>>> some thing on it ?
>>>>>>
>>>>> 120 seconds is more than enough time for pinning to go away, but as I
>>>>> said the pinning is not even necessary. Patch to remove it is welcomed.
>>>> Thank you for your clarification !  I've got it. we will still work on it.
>>>>
>>> Should be extremely easy. Drop FOLL_GET from GUP in async_pf_execute().
>>
>> One lower question, why pinning page is not necessary here?
>>
> The purpose of GUP here is to bring page from swap, the page itself is
> never used directly by async pf code. The page is used when guest
> accesses it next time, but that code path does its own GUP.

Got it, thanks for your explanation.:)

Regards,
Gu

> 
> --
> 			Gleb.
>