Re: [PATCH 1/2 v2] f2fs: fix wrong calculation of block age

[qixiaoyu]

On Sat, Jan 28, 2023 at 11:35:34AM +0800, Chao Yu wrote:
> On 2023/1/16 11:08, qixiaoyu1 wrote:
> > Currently we wrongly calculate the new block age to
> > old * LAST_AGE_WEIGHT / 100.
> > 
> > Fix it to new * (100 - LAST_AGE_WEIGHT) / 100
> >                  + old * LAST_AGE_WEIGHT / 100.
> > 
> > Signed-off-by: qixiaoyu1 <qixiaoyu1@xxxxxxxxxx>
> > Signed-off-by: xiongping1 <xiongping1@xxxxxxxxxx>
> > ---
> > Change log v1 -> v2:
> >   - fix udiv
> > 
> >   fs/f2fs/extent_cache.c | 7 ++-----
> >   1 file changed, 2 insertions(+), 5 deletions(-)
> > 
> > diff --git a/fs/f2fs/extent_cache.c b/fs/f2fs/extent_cache.c
> > index 342af24b2f8c..ad5533f178fd 100644
> > --- a/fs/f2fs/extent_cache.c
> > +++ b/fs/f2fs/extent_cache.c
> > @@ -874,11 +874,8 @@ void f2fs_update_read_extent_tree_range_compressed(struct inode *inode,
> >   static unsigned long long __calculate_block_age(unsigned long long new,
> >   						unsigned long long old)
> >   {
> > -	unsigned long long diff;
> > -
> > -	diff = (new >= old) ? new - (new - old) : new + (old - new);
> > -
> > -	return div_u64(diff * LAST_AGE_WEIGHT, 100);
> > +	return div_u64(new, 100) * (100 - LAST_AGE_WEIGHT)
> > +		+ div_u64(old, 100) * LAST_AGE_WEIGHT;
> 
> How about updating as below to avoid lossing accuracy if new is less than 100?
> 
> return div_u64(new * (100 - LAST_AGE_WEIGHT), 100) +
> 		div_u64(old * LAST_AGE_WEIGHT, 100);
> 
> Thanks,
> 

We want to avoid overflow by doing the division first. To keep the accuracy, how
about updating as below:

	res = div_u64_rem(new, 100, &rem_new) * (100 - LAST_AGE_WEIGHT)
		+ div_u64_rem(old, 100, &rem_old) * LAST_AGE_WEIGHT;
	res += rem_new * (100 - LAST_AGE_WEIGHT) / 100 + rem_old * LAST_AGE_WEIGHT / 100;
	return res;

Thanks,

> >   }
> >   /* This returns a new age and allocated blocks in ei */