Re: [PATCH 2/3] mm: gup: allow FOLL_PIN to scale in SMP

[John Hubbard <jhubbard@nvidia.com>]

On 5/6/21 4:25 PM, Peter Xu wrote:
> From: Andrea Arcangeli <aarcange@redhat.com>
> 
> has_pinned cannot be written by each pin-fast or it won't scale in
> SMP. This isn't "false sharing" strictly speaking (it's more like
> "true non-sharing"), but it creates the same SMP scalability
> bottleneck of "false sharing".
> 
> To verify the improvement a new "pin_fast.c" program was added to
> the will-it-scale benchmark.
...
> 
> This commits increases the SMP scalability of pin_user_pages_fast()
> executed by different threads of the same process by more than 4000%.
> 

Remarkable! I mean, yes, everyone knows that atomic writes are
"expensive", but this is a fun, dramatic example of just *how*
expensive they can get, once you start doing contended atomic writes.


Reviewed-by: John Hubbard <jhubbard@nvidia.com>

Other notes, that don't have any effect on the above reviewed-by
tag:

On the commit log, I will add a "+1" to the idea of deleting the
pin_fast.c contents from the commit log, and just providing a URL
instead. No need to put C programs in the commit log, IMHO, especially
when you have them elsewhere anyway.


thanks,
--
John Hubbard
NVIDIA


> Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
> Signed-off-by: Peter Xu <peterx@redhat.com>
> ---
>   mm/gup.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index 63a079e361a3d..8b513e1723b45 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -1292,7 +1292,7 @@ static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
>   		BUG_ON(*locked != 1);
>   	}
>   
> -	if (flags & FOLL_PIN)
> +	if (flags & FOLL_PIN && !atomic_read(&mm->has_pinned))
>   		atomic_set(&mm->has_pinned, 1);
>   
>   	/*
> @@ -2617,7 +2617,7 @@ static int internal_get_user_pages_fast(unsigned long start,
>   				       FOLL_FAST_ONLY)))
>   		return -EINVAL;
>   
> -	if (gup_flags & FOLL_PIN)
> +	if (gup_flags & FOLL_PIN && !atomic_read(&current->mm->has_pinned))
>   		atomic_set(&current->mm->has_pinned, 1);
>   
>   	if (!(gup_flags & FOLL_FAST_ONLY))
> 

thanks,
-- 
John Hubbard
NVIDIA