Re: [PATCH v3 3/3] mm/hwpoison: fix unpoison_memory()

[Yang Shi]

On Mon, Nov 8, 2021 at 4:53 PM Naoya Horiguchi
<naoya.horiguchi@xxxxxxxxx> wrote:
>
> On Mon, Nov 08, 2021 at 03:27:55PM -0800, Yang Shi wrote:
> > On Thu, Nov 4, 2021 at 10:52 PM Naoya Horiguchi
> > <naoya.horiguchi@xxxxxxxxx> wrote:
> ...
> > > @@ -1942,6 +1987,26 @@ core_initcall(memory_failure_init);
> > >                 pr_info(fmt, pfn);                      \
> > >  })
> > >
> > > +static inline int clear_page_hwpoison(struct ratelimit_state *rs, struct page *p)
> > > +{
> > > +       if (TestClearPageHWPoison(p)) {
> > > +               unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
> > > +                                page_to_pfn(p), rs);
> > > +               num_poisoned_pages_dec();
> > > +               return 1;
> > > +       }
> > > +       return 0;
> > > +}
> > > +
> > > +static inline int unpoison_taken_off_page(struct ratelimit_state *rs,
> > > +                                         struct page *p)
> > > +{
> > > +       if (put_page_back_buddy(p) && clear_page_hwpoison(rs, p))
> >
> > It seems there might be race condition between free and allocation to
> > kick the page out of buddy before the hwpoisoned flag is cleared IIUC?
> >
> >         CPU A                     CPU B
> > Free:                            Allocation:
> > acquire zone lock
> > put back to buddy
> > release zone lock
> >                                      acquire zone lock
> >                                      try to allocate this page
> >                                      if it is hwpoisoned then reset
> > PageBuddy (not buddy page anymore)
> >                                      release zone lock
> >
> > Did I miss something?
>
> No, this is a nice catch. CPU B finds hwpoisoned page on free list then
> calling bad_page().  So doing "put back to buddy" and clear_page_hwpoison
> inside a single zone lock should prevent this race.

Yeah, or clear the flag before putting it back to buddy? Anyway
clearing the flag under zone lock is not a big deal either.

>
> Thanks,
> Naoya Horiguchi