Re: [PATCH v6 2/4] mm: Add functions to support extra actions on swap in/out

[Dave Hansen]

On 02/28/2017 10:35 AM, Khalid Aziz wrote:
> diff --git a/mm/memory.c b/mm/memory.c
> index 6bf2b47..b086c76 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -2658,6 +2658,7 @@ int do_swap_page(struct vm_fault *vmf)
>  	if (pte_swp_soft_dirty(vmf->orig_pte))
>  		pte = pte_mksoft_dirty(pte);
>  	set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
> +	arch_do_swap_page(vma->vm_mm, vmf->address, pte, vmf->orig_pte);
>  	vmf->orig_pte = pte;
>  	if (page == swapcache) {
>  		do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 91619fd..192c41a 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1538,6 +1538,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>  		swp_pte = swp_entry_to_pte(entry);
>  		if (pte_soft_dirty(pteval))
>  			swp_pte = pte_swp_mksoft_dirty(swp_pte);
> +		arch_unmap_one(mm, address, swp_pte, pteval);
>  		set_pte_at(mm, address, pte, swp_pte);
>  	} else if (PageAnon(page)) {
>  		swp_entry_t entry = { .val = page_private(page) };
> @@ -1571,6 +1572,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>  		swp_pte = swp_entry_to_pte(entry);
>  		if (pte_soft_dirty(pteval))
>  			swp_pte = pte_swp_mksoft_dirty(swp_pte);
> +		arch_unmap_one(mm, address, swp_pte, pteval);
>  		set_pte_at(mm, address, pte, swp_pte);
>  	} else
>  		dec_mm_counter(mm, mm_counter_file(page));

>From a core VM perspective, I'm fine with these hooks.  It's minimally
invasive.  It is missing some explanation in the *code* of why sparc is
doing this and when/why other architectures might want to use these
hooks.  I think that would be awfully nice.

I still think the _current_ SPARC implementation of these hooks is
pretty broken because it doesn't allow more than one ADI tag within a
given page.  But, fixing that is confined to sparc code and shouldn't
affect the core VM or these hooks.

I suspect these hooks are still quite incomplete.  For instance, I do
not think KSM goes through these paths.  Couldn't a process *lose* its
ADI tags when KSM merges an underlying physical page?

I think you need to resolve your outstanding issues (from your 0/4
patch) before anyone can really ack these.  I suspect solving your
issues will change the number and placement of these hooks.

There is no mention in these patches of the effectively reduced virtual
address space.  Why?