Re: [PATCH RFC] mm: don't miss the last page because of round-off error

[Konstantin Khlebnikov]

On Tue, Aug 21, 2018 at 8:15 PM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> On Tue, Aug 21, 2018 at 08:11:44AM +0300, Konstantin Khlebnikov wrote:
>> On Sat, Aug 18, 2018 at 4:22 AM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>> > On Fri, Aug 17, 2018 at 04:18:34PM -0700, Roman Gushchin wrote:
>> >> -                     scan = div64_u64(scan * fraction[file],
>> >> -                                      denominator);
>> >> +                     if (scan > 1)
>> >> +                             scan = div64_u64(scan * fraction[file],
>> >> +                                              denominator);
>> >
>> > Wouldn't we be better off doing a div_round_up?  ie:
>> >
>> >         scan = div64_u64(scan * fraction[file] + denominator - 1, denominator);
>> >
>> > although i'd rather hide that in a new macro in math64.h than opencode it
>> > here.
>>
>> All numbers here should be up to nr_pages * 200 and fit into unsigned long.
>> I see no reason for u64. If they overflow then u64 wouldn't help either.
>
> It is nr_pages * 200 * recent_scanned, where recent_scanned can be up
> to four times of what's on the LRUs. That can overflow a u32 with even
> small amounts of memory.

Ah, this thing is inverted because it aims to proportional reactivation rate
rather than the proportional pressure to reclaimable pages.
That's not obvious. I suppose this should be in comment above it.

Well, at least denominator should fit into unsigned long. So full
64/64 division is redundant.