Re: [RFC] mm: Optimise put_pages_list()

[Mel Gorman]

On Mon, Oct 04, 2021 at 01:49:37PM +0100, Matthew Wilcox wrote:
> On Mon, Oct 04, 2021 at 10:10:37AM +0100, Mel Gorman wrote:
> > On Thu, Sep 30, 2021 at 05:32:58PM +0100, Matthew Wilcox (Oracle) wrote:
> > > Instead of calling put_page() one page at a time, pop pages off
> > > the list if there are other refcounts and pass the remainder
> > > to free_unref_page_list().  This should be a speed improvement,
> > > but I have no measurements to support that.  It's also not very
> > > widely used today, so I can't say I've really tested it.  I'm only
> > > bothering with this patch because I'd like the IOMMU code to use it
> > > https://lore.kernel.org/lkml/20210930162043.3111119-1-willy@xxxxxxxxxxxxx/
> > > 
> > > Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > 
> > I see your motivation but you need to check that all users of
> > put_pages_list (current and future) handle destroy_compound_page properly
> > or handle it within put_pages_list. For example, the release_pages()
> > user of free_unref_page_list calls __put_compound_page directly before
> > freeing. put_pages_list as it stands will call dstroy_compound_page but
> > free_unref_page_list does not destroy compound pages in free_pages_prepare
> 
> Quite right.  I was really only thinking about order-zero pages because
> there aren't any users of compound pages that call this.  But of course,
> we should be robust against future callers.  So the obvious thing to do
> is to copy what release_pages() does:
> 
> +++ b/mm/swap.c
> @@ -144,6 +144,10 @@ void put_pages_list(struct list_head *pages)
>         list_for_each_entry_safe(page, next, pages, lru) {
>                 if (!put_page_testzero(page))
>                         list_del(&page->lru);
> +               if (PageCompound(page)) {
> +                       list_del(&page->lru);
> +                       __put_compound_page(page);
> +               }
>         }
> 
>         free_unref_page_list(pages);

That would be the most straight-forward

> 
> But would it be better to have free_unref_page_list() handle compound
> pages itself?
> 
> +++ b/mm/page_alloc.c
> @@ -3427,6 +3427,11 @@ void free_unref_page_list(struct list_head *list)
> 
>         /* Prepare pages for freeing */
>         list_for_each_entry_safe(page, next, list, lru) {
> +               if (PageCompound(page)) {
> +                       __put_compound_page(page);
> +                       list_del(&page->lru);
> +                       continue;
> +               }
>                 pfn = page_to_pfn(page);
>                 if (!free_unref_page_prepare(page, pfn, 0)) {
>                         list_del(&page->lru);
> 
> (and delete the special handling from release_pages() in the same patch)

It's surprisingly tricky.

Minimally, that list_del should be before __put_compound_page or you'll
clobber whatever list the compound page destructor placed the free page on.
Take care with how you remove the special handling and leave a comment
explaining why __put_compound_page is not called and that PageLRU will be
cleared when it falls through to add the page to pages_to_free.  The tricky
part is memcg uncharging because if mem_cgroup_uncharge_list() is called
then the uncharging happens twice -- once in the destructor and again in
mem_cgroup_uncharge_list. I guess you could use two lists and splice them
after mem_cgroup_uncharge_list() and before free_unref_page_list.

-- 
Mel Gorman
SUSE Labs