Re: [PATCH] mm: pgtable: Unlock pml without branches when !start_pte

[I Hsin Cheng]

On Mon, Feb 10, 2025 at 04:37:36PM -0800, Andrew Morton wrote:
> On Mon, 10 Feb 2025 18:09:48 +0800 I Hsin Cheng <richard120310@xxxxxxxxx> wrote:
> 
> > When !start_pte is true, the branch for "start_pte" in "out_ptl" label
> > section is surely false, and "ptl != pml" must be true since "ptl" is
> > NULL in this case.
> > 
> > It means both branches in "out_ptl" are redundant, only one thing to be
> > done is to unlock "pml", make it directly unlock "pml" and return in
> > this case.
> 
> Hopefully the compiler will skip the `if (start_pte)' test.
> 
> Generally, we try to avoid multiple function return points.  We could do
> 
> --- a/mm/pt_reclaim.c~mm-pgtable-unlock-pml-without-branches-when-start_pte
> +++ a/mm/pt_reclaim.c
> @@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *m
>  	pml = pmd_lock(mm, pmd);
>  	start_pte = pte_offset_map_rw_nolock(mm, pmd, addr, &pmdval, &ptl);
>  	if (!start_pte)
> -		goto out_ptl;
> +		goto out_unlock;
>  	if (ptl != pml)
>  		spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
>  
> @@ -67,5 +67,6 @@ out_ptl:
>  	if (start_pte)
>  		pte_unmap_unlock(start_pte, ptl);
>  	if (ptl != pml)
> +out_unlock:
>  		spin_unlock(pml);
>  }
> _
> 
> but that's really ugly.


Hi Andrew,

Thanks for your review!

>       if (ptl != pml)
> +out_unlock:
>               spin_unlock(pml);
>  }
> _
>
> but that's really ugly.

I agree. Would you be so nice to suggest some test method for me so I
can try to test how much benefit we can get from this?

If the case happens frequently enough I think it might be worth it?

Best regards,
I Hsin Cheng