Re: [RFC PATCH 5/5] Revert "mm/migrate: __unmap_and_move() push good newpage to LRU"

[David Hildenbrand <david@redhat.com>]

On 19.06.23 09:56, David Hildenbrand wrote:
> On 19.06.23 05:59, Yosry Ahmed wrote:
>> On Sun, Jun 18, 2023 at 7:00 PM Huang, Ying <ying.huang@intel.com> wrote:
>>>
>>> Hi, Yosry,
>>>
>>> Yosry Ahmed <yosryahmed@google.com> writes:
>>>
>>>> This reverts commit c3096e6782b733158bf34f6bbb4567808d4e0740.
>>>>
>>>> That commit made sure we immediately add the new page to the LRU before
>>>> remove_migration_ptes() is called in migrate_move_folio() (used to be
>>>> __unmap_and_move() back then), such that the rmap walk will rebuild the
>>>> correct mlock_count for the page again. This was needed because the
>>>> mlock_count was lost when the page is isolated. This is no longer the
>>>> case since mlock_count no longer overlays page->lru.
>>>>
>>>> Revert the commit (the code was foliated afterward the commit, so the
>>>> revert is updated as such).
>>>>
>>>> Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
>>>> ---
>>>>    mm/migrate.c | 24 +++++++++---------------
>>>>    1 file changed, 9 insertions(+), 15 deletions(-)
>>>>
>>>> diff --git a/mm/migrate.c b/mm/migrate.c
>>>> index 01cac26a3127..68f693731865 100644
>>>> --- a/mm/migrate.c
>>>> +++ b/mm/migrate.c
>>>> @@ -1279,19 +1279,6 @@ static int migrate_folio_move(free_page_t put_new_page, unsigned long private,
>>>>         if (unlikely(!is_lru))
>>>>                 goto out_unlock_both;
>>>
>>> The patch itself looks good to me!  Thanks!
>>>
>>> Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
>>
>> Thanks for taking a look!
>>
>>>
>>> And, it seems that we can remove the above 2 lines and "out_unlock_both"
>>> label now.  That can make the code simpler a little.  Right?
>>
>> I am not familiar with this code. If we remove the above condition
>> then pages that have is_lru == 0 (i.e __PageMovable(src) is true) and
>> page_was_mapped == 1 will call remove_migration_ptes(). This wouldn't
>> happen without removing the above 2 lines. If this combination is
>> impossible (is_lru == 0 && page_was_mapped == 1), then yeah we can
>> remove the above condition.
>>
>> It looks like __SetPageMovable() is only called by zsmalloc, z3fold,
>> and balloon_page_insert(). The former 2 will never have those pages
>> mapped into userspace. I am not familiar with balloon_page_insert(),
>> but my gut feeling is that those are pages used by the driver and are
>> also not mapped into userspace.
> 
> On XEN, there is xen_alloc_ballooned_pages(), which ends up mapping
> balloon-inflated pages into user space (for something like MMIO IIRC).
> But the XEN balloon does not use the balloon compaction framework, so
> __SetPageMovable() does not apply.
> 
> The other balloon_page_insert() users (VMware balloon, CMM,
> virtio-balloon) shouldn't be doing something like that.

Ah, and I remember they even can't, because in balloon_page_insert() we 
also do a __SetPageOffline(). And such typed pages cannot be mapped into 
user space (because the type overlays the mapcount).

-- 
Cheers,

David / dhildenb