Re: [PATCH v2 6/9] mm: free user PTE page table pages

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Aug 19, 2021 at 11:18:55AM +0800, Qi Zheng wrote:

> diff --git a/mm/gup.c b/mm/gup.c
> index 2630ed1bb4f4..30757f3b176c 100644
> +++ b/mm/gup.c
> @@ -500,6 +500,9 @@ static struct page *follow_page_pte(struct vm_area_struct *vma,
>  	if (unlikely(pmd_bad(*pmd)))
>  		return no_page_table(vma, flags);
>  
> +	if (!pte_try_get(mm, pmd))
> +		return no_page_table(vma, flags);
> +
>  	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);

This is not good on a performance path, the pte_try_get() is
locking/locking the same lock that pte_offset_map_lock() is getting.

This would be much better if the map_lock infra could manage the
refcount itself.

I'm also not really keen on adding ptl level locking to all the
currently no-lock paths. If we are doing that then the no-lock paths
should rely on the ptl for alot more of their operations and avoid the
complicatred no-lock data access we have. eg 'pte_try_get()' should
also copy the pte_t under the lock.

Also, I don't really understand how this scheme works with
get_user_pages_fast.

Currently the zap triggers a TLB invalidation which synchronizes with
GUP fast, however this only makes the ptes non-present. The purpose is
to synchronize with the struct page refcount, not a pte refcount.

With this series the non-present PTEs are freed but how does this
synchronize with gup fast to avoid a use-after-free on the pte struct
page?

I agree with David, this series needs significant splitting to be
readable and a lot more explanation in the commit messages how all the
locking is working. Eg introducing the freeing should be a single
short patch at at end with a full explanation of the locking in all
the major scenarios.

Jason