Re: [PATCH v4 4/8] hugetlb: create remove_hugetlb_page() to separate functionality

[Michal Hocko]

On Tue 06-04-21 09:49:13, Mike Kravetz wrote:
> On 4/6/21 2:56 AM, Michal Hocko wrote:
> > On Mon 05-04-21 16:00:39, Mike Kravetz wrote:
[...]
> >> @@ -2298,6 +2312,7 @@ static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
> >>  		/*
> >>  		 * Freed from under us. Drop new_page too.
> >>  		 */
> >> +		remove_hugetlb_page(h, new_page, false);
> >>  		update_and_free_page(h, new_page);
> >>  		goto unlock;
> >>  	} else if (page_count(old_page)) {
> >> @@ -2305,6 +2320,7 @@ static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
> >>  		 * Someone has grabbed the page, try to isolate it here.
> >>  		 * Fail with -EBUSY if not possible.
> >>  		 */
> >> +		remove_hugetlb_page(h, new_page, false);
> >>  		update_and_free_page(h, new_page);
> >>  		spin_unlock(&hugetlb_lock);
> >>  		if (!isolate_huge_page(old_page, list))
> > 
> > the page is not enqued anywhere here so remove_hugetlb_page would blow
> > when linked list debugging is enabled.
> 
> I also thought this would be an issue.  However, INIT_LIST_HEAD would
> have been called for the page so,

OK, this is true for a freshly allocated hugetlb page (prep_new_huge_page.
It's a very sublte dependency though. In case somebody ever wants to
fortify linked lists and decides to check list_del on an empty list then
this would wait silently to blow up.

> Going forward, I agree it would be better to perhaps add a list_empty
> check so that things do not blow up if the debugging code is changed.

Yes this is less tricky then a bool flag or making more stages of the
tear down. 2 stages are more than enough IMHO.

> At one time I also thought of splitting the functionality in
> alloc_fresh_huge_page and prep_new_huge_page so that it would only
> allocate/prep the page but not increment nr_huge_pages.

We already have that distinction. alloc_buddy_huge_page is there to
allocate a fresh huge page without any hstate  accunting. Considering
that giga pages are not supported for the migration anyway, maybe this
would make Oscar's work slightly less tricky?
-- 
Michal Hocko
SUSE Labs