Re: [External] Re: [PATCH v3 4/6] mm: hugetlb: add return -EAGAIN for dissolve_free_huge_page

[Muchun Song <songmuchun@bytedance.com>]

On Tue, Jan 12, 2021 at 6:06 PM Michal Hocko <mhocko@suse.com> wrote:
>
> On Tue 12-01-21 17:51:05, Muchun Song wrote:
> > On Tue, Jan 12, 2021 at 4:33 PM Michal Hocko <mhocko@suse.com> wrote:
> > >
> > > On Mon 11-01-21 17:20:51, Mike Kravetz wrote:
> > > > On 1/10/21 4:40 AM, Muchun Song wrote:
> > > > > There is a race between dissolve_free_huge_page() and put_page(),
> > > > > and the race window is quite small. Theoretically, we should return
> > > > > -EBUSY when we encounter this race. In fact, we have a chance to
> > > > > successfully dissolve the page if we do a retry. Because the race
> > > > > window is quite small. If we seize this opportunity, it is an
> > > > > optimization for increasing the success rate of dissolving page.
> > > > >
> > > > > If we free a HugeTLB page from a non-task context, it is deferred
> > > > > through a workqueue. In this case, we need to flush the work.
> > > > >
> > > > > The dissolve_free_huge_page() can be called from memory hotplug,
> > > > > the caller aims to free the HugeTLB page to the buddy allocator
> > > > > so that the caller can unplug the page successfully.
> > > > >
> > > > > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > > > > ---
> > > > >  mm/hugetlb.c | 26 +++++++++++++++++++++-----
> > > > >  1 file changed, 21 insertions(+), 5 deletions(-)
> > > >
> > > > I am unsure about the need for this patch.  The code is OK, there are no
> > > > issues with the code.
> > > >
> > > > As mentioned in the commit message, this is an optimization and could
> > > > potentially cause a memory offline operation to succeed instead of fail.
> > > > However, we are very unlikely to ever exercise this code.  Adding an
> > > > optimization that is unlikely to be exercised is certainly questionable.
> > > >
> > > > Memory offline is the only code that could benefit from this optimization.
> > > > As someone with more memory offline user experience, what is your opinion
> > > > Michal?
> > >
> > > I am not a great fun of optimizations without any data to back them up.
> > > I do not see any sign this code has been actually tested and the
> > > condition triggered.
> >
> > This race is quite small. I only trigger this only once on my server.
> > And then the kernel panic. So I sent this patch series to fix some
> > bugs.
>
> Memory hotplug shouldn't panic when this race happens. Are you sure you
> have seen a race that is directly related to this patch?

I mean the panic is fixed by:

  [PATCH v3 3/6] mm: hugetlb: fix a race between freeing and dissolving the page

>
> > > Besides that I have requested to have an explanation of why blocking on
> > > the WQ is safe and that hasn't happened.
> >
> > I have seen all the caller of dissolve_free_huge_page, some caller is under
> > page lock (via lock_page). Others are also under a sleep context.
> >
> > So I think that blocking on the WQ is safe. Right?
>
> I have requested to explicitly write your thinking why this is safe so
> that we can double check it. Dependency on a work queue progress is much
> more complex than any other locks because there is no guarantee that WQ
> will make forward progress (all workers might be stuck, new workers not
> able to be created etc.).

OK. I know about your concern. How about setting the page as temporary
when hitting this race?

 int dissolve_free_huge_page(struct page *page)
 {
@@ -1793,8 +1794,10 @@ int dissolve_free_huge_page(struct page *page)
                 * We should make sure that the page is already on the free list
                 * when it is dissolved.
                 */
-               if (unlikely(!PageHugeFreed(head)))
+               if (unlikely(!PageHugeFreed(head))) {
+                       SetPageHugeTemporary(page)
                        goto out;
+               }

Setting the page as temporary and just return -EBUSY (do not flush
the work). __free_huge_page() will free it to the buddy allocator later.

> --
> Michal Hocko
> SUSE Labs