lkml.org 
[lkml]   [2021]   [Jul]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 1/5] mm/vmscan: put the redirtied MADV_FREE pages back to anonymous LRU list
From
Date
On 2021/7/15 3:43, John Hubbard wrote:
> On 7/14/21 4:48 AM, Matthew Wilcox wrote:
>> On Wed, Jul 14, 2021 at 07:36:57PM +0800, Miaohe Lin wrote:
>>> On 2021/7/13 21:34, Matthew Wilcox wrote:
>>>> On Tue, Jul 13, 2021 at 09:13:51PM +0800, Miaohe Lin wrote:
>>>>>>> When the MADV_FREE pages are redirtied before they could be reclaimed, the pages
>>>>>>> should be put back to anonymous LRU list by setting SwapBacked flag, thus the
>>>>>>> pages will be reclaimed in normal swapout way.
>>>>>>
>>>>>> Agreed. But the question is why this needs an explicit handling here
>>>>>> when we already do handle this case when trying to unmap the page.
>>>>>
>>>>> This makes me think more. It seems even the page_ref_freeze call is guaranteed to
>>>>> success as no one can grab the page refcnt after the page is successfully unmapped.
>>>>
>>>> NO!  This is wrong.  Every page can have its refcount speculatively raised
>>>> (and then lowered).  The two prime candidates for this are lockless GUP
>>>> and page cache lookups, but there can be others too.
>>>>
>>>
>>> Many thanks for pointing this out. My overlook! Sorry!
>>> So, it seems lockless GUP can redirty the MADV_FREE page. But is it ok to just release
>>> a redirtied MADV_FREE pages? Because we hold the last reference here and the page will
>>> be freed anyway...
>>
>> I don't see how lockless GUP can redirty the page.  It can grab the
>> refcount, thus making the refcount here two.  Then the call to freeze
>> here fails and the page stays on the list.  But the lockless GUP checks
>> the page is still in the page table (and discovers it isn't, so releases
>> the reference count).  Am I missing a path that lets lockless GUP dirty
>> the page?
>>
>
> If a device driver pins some pages using gup, and the device then uses dma
> to write to those pages, then you could get there. That story is part of the
> reasoning that led to creating pin_user_pages(), which btw does not yet
> fully solve that case.

Many thanks for your explanation.
So the similar scenario that is clarified in the __remove_mapping() is possible:

get_user_pages(&page);
[user mapping goes away]
write_to(page);
!PageDirty(page) [good]
SetPageDirty(page);
put_page(page);
!page_count(page) [good, discard it]

[oops, our write_to data is lost]

The page can be redirtied after the page is unmapped. And there is no way to restore the page
table as clean MADV_FREE page is simply cleared from page table via the try_to_unmap path.
Is it ok to just release the redirtied MADV_FREE pages here as we hold the last reference
and the page will be freed anyway... ?

>
> Basically, though, unless a non-CPU device has access to the page, it's
> hard to see how gup itself can lead to a page getting dirtied.
>
> thanks,

\
 
 \ /
  Last update: 2021-07-15 13:30    [W:0.107 / U:0.040 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site