Re: [PATCH 4/8] mm/gup: add an assertion that the mmap lock is locked

[John Hubbard <jhubbard@nvidia.com>]

On 1/17/23 07:58, Jason Gunthorpe wrote:
> This is always required, but we can't have a proper unguarded assertion
> because of a shortcut in fork. So, cover as much as we can for now.
> 
> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> ---
>   mm/gup.c | 8 ++++++++
>   1 file changed, 8 insertions(+)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index 9e332e3f6ea8e2..d203e268793b9c 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -1350,6 +1350,14 @@ static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
>   			return -EAGAIN;
>   		lock_dropped = true;
>   		*locked = 1;
> +	} else if (flags & FOLL_PIN) {
> +		/*
> +		 * The mmap lock must be held when calling this function. This
> +		 * is true even for non-pin modes, but due to a shortcut in fork
> +		 * not taking the lock for the new mm we cannot check this
> +		 * comprehensively.

I get that fork doesn't lock the newly created mm. But I'm having
trouble finding the calling path from fork to __get_user_pages_locked()
that leads to this comment, can you provide a hint? Both the comment
and the commit log are rather coy about where this happens.

> +		 */
> +		mmap_assert_locked(mm);
>   	}
>   
>   	if (flags & FOLL_PIN)

thanks,
-- 
John Hubbard
NVIDIA