Re: [PATCH 4/6] hugetlb: avoid allocation failed when page reporting is on going

[Liang Li <liliang324@gmail.com>]

> > > Please don't use this email address for me anymore. Either use
> > > alexander.duyck@gmail.com or alexanderduyck@fb.com. I am getting
> > > bounces when I reply to this thread because of the old address.
> >
> > No problem.
> >
> > > > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > > > index eb533995cb49..0fccd5f96954 100644
> > > > --- a/mm/hugetlb.c
> > > > +++ b/mm/hugetlb.c
> > > > @@ -2320,6 +2320,12 @@ struct page *alloc_huge_page(struct vm_area_struct *vma,
> > > >                 goto out_uncharge_cgroup_reservation;
> > > >
> > > >         spin_lock(&hugetlb_lock);
> > > > +       while (h->free_huge_pages <= 1 && h->isolated_huge_pages) {
> > > > +               spin_unlock(&hugetlb_lock);
> > > > +               mutex_lock(&h->mtx_prezero);
> > > > +               mutex_unlock(&h->mtx_prezero);
> > > > +               spin_lock(&hugetlb_lock);
> > > > +       }
> > >
> > > This seems like a bad idea. It kind of defeats the whole point of
> > > doing the page zeroing outside of the hugetlb_lock. Also it is
> > > operating on the assumption that the only way you might get a page is
> > > from the page zeroing logic.
> > >
> > > With the page reporting code we wouldn't drop the count to zero. We
> > > had checks that were going through and monitoring the watermarks and
> > > if we started to hit the low watermark we would stop page reporting
> > > and just assume there aren't enough pages to report. You might need to
> > > look at doing something similar here so that you can avoid colliding
> > > with the allocator.
> >
> > For hugetlb, things are a little different, Just like Mike points out:
> >      "On some systems, hugetlb pages are a precious resource and
> >       the sysadmin carefully configures the number needed by
> >       applications.  Removing a hugetlb page (even for a very short
> >       period of time) could cause serious application failure."
> >
> > Just keeping some pages in the freelist is not enough to prevent that from
> > happening, because these pages may be allocated while zero out is on
> > going, and application may still run into a situation for not available free
> > pages.
>
> I get what you are saying. However I don't know if it is acceptable
> for the allocating thread to be put to sleep in this situation. There
> are two scenarios where I can see this being problematic.
>
> One is a setup where you put the page allocator to sleep and while it
> is sleeping another thread is then freeing a page and your thread
> cannot respond to that newly freed page and is stuck waiting on the
> zeroed page.
>
> The second issue is that users may want a different option of just
> breaking up the request into smaller pages rather than waiting on the
> page zeroing, or to do something else while waiting on the page. So
> instead of sitting on the request and waiting it might make more sense
> to return an error pointer like EAGAIN or EBUSY to indicate that there
> is a page there, but it is momentarily tied up.

It seems returning EAGAIN or EBUSY will still change the application's
behavior,  I am not sure if it's acceptable.

Thanks
Liang