Re: [External] Re: [PATCH] mm/page_alloc: speeding up the iteration of max_order

[Vlastimil Babka <vbabka@suse.cz>]

On 12/4/20 5:03 AM, Muchun Song wrote:
> On Fri, Dec 4, 2020 at 1:37 AM Vlastimil Babka <vbabka@suse.cz> wrote:
>>
>> On 12/2/20 1:18 PM, Muchun Song wrote:
>> > When we free a page whose order is very close to MAX_ORDER and greater
>> > than pageblock_order, it wastes some CPU cycles to increase max_order
>> > to MAX_ORDER one by one and check the pageblock migratetype of that page
>>
>> But we have to do that. It's not the same page, it's the merged page and the new
>> buddy is a different pageblock and we need to check if they have compatible
>> migratetypes and can merge, or we have to bail out. So the patch is wrong.
>>
>> > repeatedly especially when MAX_ORDER is much larger than pageblock_order.
>>
>> Do we have such architectures/configurations anyway?
>>
>> > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
>> > ---
>> >  mm/page_alloc.c | 4 +++-
>> >  1 file changed, 3 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> > index 141f12e5142c..959541234e1d 100644
>> > --- a/mm/page_alloc.c
>> > +++ b/mm/page_alloc.c
>> > @@ -1041,7 +1041,7 @@ static inline void __free_one_page(struct page *page,
>> >               pfn = combined_pfn;
>> >               order++;
>> >       }
>> > -     if (max_order < MAX_ORDER) {
> 
> If we free a page with order == MAX_ORDER - 1, it has no buddy.
> The following pageblock operation is also pointless.

OK, I see.

>> > +     if (max_order < MAX_ORDER && order < MAX_ORDER - 1) {

Yes, this makes sense, as in your other patch we shouldn't check the buddy when
order == MAX_ORDER - 1 already.

>> >               /* If we are here, it means order is >= pageblock_order.
>> >                * We want to prevent merge between freepages on isolate
>> >                * pageblock and normal pageblock. Without this, pageblock
>> > @@ -1062,6 +1062,8 @@ static inline void __free_one_page(struct page *page,
>> >                                               is_migrate_isolate(buddy_mt)))
>> >                               goto done_merging;
>> >               }
>> > +             if (unlikely(order != max_order - 1))
>> > +                     max_order = order + 1;
>> >               max_order++;

OK I see now what you want to do here. the "if" may be true if we already
entered the function with order > pageblock_order.
I think we could just simplfy the "if" and "max_order++" above to:

max_order = order + 2

which starts to get a bit ugly, so why not change max_order to be -1 (compared
to now) in the whole function:

max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
...
continue_merging:
        while (order < max_order) {
...
if (order < MAX_ORDER - 1) {
// it's redundant to keep checking max_order < MAX_ORDER - 1 here after your
change, right?
...

max_order = order + 1; // less weird than "+ 2"

Off by one errors, here we go!

>> Or maybe I just don't understand what this is doing. When is the new 'if' even
>> true? We just bailed out of "while (order < max_order - 1)" after the last
>> "order++", which means it should hold that "order == max_order - 1")?
> 
> No, I do not agree. The MAX_ORDER may be greater than 11.
> 
> # git grep "CONFIG_FORCE_MAX_ZONEORDER"
> # arch/arm/configs/imx_v6_v7_defconfig:CONFIG_FORCE_MAX_ZONEORDER=14
> # arch/powerpc/configs/85xx/ge_imp3a_defconfig:CONFIG_FORCE_MAX_ZONEORDER=17
> # arch/powerpc/configs/fsl-emb-nonhw.config:CONFIG_FORCE_MAX_ZONEORDER=13
> 
> Have you seen it? On some architecture, the MAX_ORDER
> can be 17. When we free a page with an order 16. Without this
> patch, the max_order should be increased one by one from 10 to
> 17.
> 
> Thanks.
> 
> 
>> Your description sounds like you want to increase max_order to MAX_ORDER in one
>> step, which as I explained would be wrong. But the implementation looks actually
>> like a no-op.
>>
>> >               max_order++;
>> >               goto continue_merging;
>> >       }
>> >
>>
> 
> 
> --
> Yours,
> Muchun
>