Re: [mm-unstable PATCH v4 1/9] mm/hugetlb: check gigantic_page_runtime_supported() in return_unused_surplus_pages()

[Mike Kravetz <mike.kravetz@oracle.com>]

On 07/06/22 11:04, Miaohe Lin wrote:
> On 2022/7/5 14:39, HORIGUCHI NAOYA(堀口 直也) wrote:
> > On Tue, Jul 05, 2022 at 10:16:39AM +0800, Miaohe Lin wrote:
> >> On 2022/7/4 9:33, Naoya Horiguchi wrote:
> >>> From: Naoya Horiguchi <naoya.horiguchi@nec.com>
> >>>
> >>> I found a weird state of 1GB hugepage pool, caused by the following
> >>> procedure:
> >>>
> >>>   - run a process reserving all free 1GB hugepages,
> >>>   - shrink free 1GB hugepage pool to zero (i.e. writing 0 to
> >>>     /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages), then
> >>>   - kill the reserving process.
> >>>
> >>> , then all the hugepages are free *and* surplus at the same time.
> >>>
> >>>   $ cat /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages
> >>>   3
> >>>   $ cat /sys/kernel/mm/hugepages/hugepages-1048576kB/free_hugepages
> >>>   3
> >>>   $ cat /sys/kernel/mm/hugepages/hugepages-1048576kB/resv_hugepages
> >>>   0
> >>>   $ cat /sys/kernel/mm/hugepages/hugepages-1048576kB/surplus_hugepages
> >>>   3
> >>>
> >>> This state is resolved by reserving and allocating the pages then
> >>> freeing them again, so this seems not to result in serious problem.
> >>> But it's a little surprising (shrinking pool suddenly fails).
> >>>
> >>> This behavior is caused by hstate_is_gigantic() check in
> >>> return_unused_surplus_pages(). This was introduced so long ago in 2008
> >>> by commit aa888a74977a ("hugetlb: support larger than MAX_ORDER"), and
> >>> at that time the gigantic pages were not supposed to be allocated/freed
> >>> at run-time.  Now kernel can support runtime allocation/free, so let's
> >>> check gigantic_page_runtime_supported() together.
> >>>
> >>> Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
> >>
> >> This patch looks good to me with a few question below.
> > 
> > Thank you for reviewing.
> > 
> >>
> >>> ---
> >>> v2 -> v3:
> >>> - Fixed typo in patch description,
> >>> - add !gigantic_page_runtime_supported() check instead of removing
> >>>   hstate_is_gigantic() check (suggested by Miaohe and Muchun)
> >>> - add a few more !gigantic_page_runtime_supported() check in
> >>>   set_max_huge_pages() (by Mike).
> >>> ---
> >>>  mm/hugetlb.c | 19 ++++++++++++++++---
> >>>  1 file changed, 16 insertions(+), 3 deletions(-)
> >>>
> >>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> >>> index 2a554f006255..bdc4499f324b 100644
> >>> --- a/mm/hugetlb.c
> >>> +++ b/mm/hugetlb.c
> >>> @@ -2432,8 +2432,7 @@ static void return_unused_surplus_pages(struct hstate *h,
> >>>  	/* Uncommit the reservation */
> >>>  	h->resv_huge_pages -= unused_resv_pages;
> >>>  
> >>> -	/* Cannot return gigantic pages currently */
> >>> -	if (hstate_is_gigantic(h))
> >>> +	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
> >>>  		goto out;
> >>>  
> >>>  	/*
> >>> @@ -3315,7 +3314,8 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> >>>  	 * the user tries to allocate gigantic pages but let the user free the
> >>>  	 * boottime allocated gigantic pages.
> >>>  	 */
> >>> -	if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
> >>> +	if (hstate_is_gigantic(h) && (!IS_ENABLED(CONFIG_CONTIG_ALLOC) ||
> >>> +				      !gigantic_page_runtime_supported())) {
> >>>  		if (count > persistent_huge_pages(h)) {
> >>>  			spin_unlock_irq(&hugetlb_lock);
> >>>  			mutex_unlock(&h->resize_lock);
> >>> @@ -3363,6 +3363,19 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
> >>>  			goto out;
> >>>  	}
> >>>  
> >>> +	/*
> >>> +	 * We can not decrease gigantic pool size if runtime modification
> >>> +	 * is not supported.
> >>> +	 */
> >>> +	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) {
> >>> +		if (count < persistent_huge_pages(h)) {
> >>> +			spin_unlock_irq(&hugetlb_lock);
> >>> +			mutex_unlock(&h->resize_lock);
> >>> +			NODEMASK_FREE(node_alloc_noretry);
> >>> +			return -EINVAL;
> >>> +		}
> >>> +	}
> >>
> >> With above change, we're not allowed to decrease the pool size now. But it was allowed previously
> >> even if !gigantic_page_runtime_supported. Does this will break user?
> > 
> > Yes, it does. I might get the wrong idea about the definition of
> > gigantic_page_runtime_supported(), which shows that runtime pool *extension*
> > is supported or not (implying that pool shrinking is always possible).
> > If this is right, this new if-block is not necessary.
> 
> I tend to remove above new if-block to keep pool shrinking available.
> 

Not sure I am following the questions.

Take a look at __update_and_free_page which will refuse to 'free' a
gigantic page if !gigantic_page_runtime_supported.  I 'think' attempting
to shrink the pool when !gigantic_page_runtime_supported will result in
leaking gigantic pages.  i.e.  Memory will remain allocated for the
gigantic page, but it can not be used.

I can take a closer look during my tomorrow.

IIRC, the only way gigantic_page_runtime_supported is not set to day is
in the case of powerpc using 16GB pages allocated/managed by firmware.
-- 
Mike Kravetz