Date:   Tue, 20 Jun 2023 11:05:33 -0700
From:   Mike Kravetz <mike.kravetz@...cle.com>
To:     Naoya Horiguchi <naoya.horiguchi@...ux.dev>
Cc:     Jiaqi Yan <jiaqiyan@...gle.com>,
        HORIGUCHI NAOYA(堀口 直也) 
        <naoya.horiguchi@....com>,
        "songmuchun@...edance.com" <songmuchun@...edance.com>,
        "shy828301@...il.com" <shy828301@...il.com>,
        "linmiaohe@...wei.com" <linmiaohe@...wei.com>,
        "akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
        "linux-mm@...ck.org" <linux-mm@...ck.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "duenwen@...gle.com" <duenwen@...gle.com>,
        "axelrasmussen@...gle.com" <axelrasmussen@...gle.com>,
        "jthoughton@...gle.com" <jthoughton@...gle.com>
Subject: Re: [PATCH v1 1/3] mm/hwpoison: find subpage in hugetlb HWPOISON list

On 06/19/23 17:23, Naoya Horiguchi wrote:
> On Sat, Jun 17, 2023 at 03:59:27PM -0700, Mike Kravetz wrote:
> > On 06/16/23 19:18, Jiaqi Yan wrote:
> > > On Fri, Jun 16, 2023 at 4:35 PM Mike Kravetz <mike.kravetz@...cle.com> wrote:
> > > > On 06/16/23 14:19, Jiaqi Yan wrote:
> > > > >
> > > > > Now looking again this, I think concurrent adding and deleting are
> > > > > fine with each other and with themselves, because raw_hwp_list is
> > > > > lock-less llist.
> > > >
> > > > Correct.
> > > >
> > > > > As for synchronizing traversal with adding and deleting, I wonder is
> > > > > it a good idea to make __update_and_free_hugetlb_folio hold
> > > > > hugetlb_lock before it folio_clear_hugetlb_hwpoison(which traverse +
> > > > > delete raw_hwp_list)? In hugetlb, get_huge_page_for_hwpoison already
> > > > > takes hugetlb_lock; it seems to me __update_and_free_hugetlb_folio is
> > > > > missing the lock.
> > > >
> > > > I do not think the lock is needed.  However, while looking more closely
> > > > at this I think I discovered another issue.
> > > > This is VERY subtle.
> > > > Perhaps Naoya can help verify if my reasoning below is correct.
> > > >
> > > > In __update_and_free_hugetlb_folio we are not operating on a hugetlb page.
> > > > Why is this?
> > > > Before calling update_and_free_hugetlb_folio we call remove_hugetlb_folio.
> > > > The purpose of remove_hugetlb_folio is to remove the huge page from the
> > > > list AND compound page destructor indicating this is a hugetlb page is changed.
> > > > This is all done while holding the hugetlb lock.  So, the test for
> > > > folio_test_hugetlb(folio) is false.
> > > >
> > > > We have technically a compound non-hugetlb page with a non-null raw_hwp_list.
> > > >
> > > > Important note: at this time we have not reallocated vmemmap pages if
> > > > hugetlb page was vmemmap optimized.  That is done later in
> > > > __update_and_free_hugetlb_folio.
> > > 
> > > 
> > > >
> > > > The 'good news' is that after this point get_huge_page_for_hwpoison will
> > > > not recognize this as a hugetlb page, so nothing will be added to the
> > > > list.  There is no need to worry about entries being added to the list
> > > > during traversal.
> > > >
> > > > The 'bad news' is that if we get a memory error at this time we will
> > > > treat it as a memory error on a regular compound page.  So,
> > > > TestSetPageHWPoison(p) in memory_failure() may try to write a read only
> > > > struct page. :(
> > > 
> > > At least I think this is an issue.
> > > 
> > > Would it help if dissolve_free_huge_page doesn't unlock hugetlb_lock
> > > until update_and_free_hugetlb_folio is done, or basically until
> > > dissolve_free_huge_page is done?
> > 
> > Unfortunately, update_and_free_hugetlb_folio is designed to be called
> > without locks held.  This is because we can not hold any locks while
> > allocating vmemmap pages.
> > 
> > I'll try to think of some way to restructure the code.  IIUC, this is a
> > potential general issue, not just isolated to memory error handling.
> 
> Considering this issue as one specific to memory error handling, checking
> HPG_vmemmap_optimized in __get_huge_page_for_hwpoison() might be helpful to
> detect the race.  Then, an idea like the below diff (not tested) can make
> try_memory_failure_hugetlb() retry (with retaking hugetlb_lock) to wait
> for complete the allocation of vmemmap pages.
> 
> @@ -1938,8 +1938,11 @@ int __get_huge_page_for_hwpoison(unsigned long pfn, int flags,
>         int ret = 2;    /* fallback to normal page handling */
>         bool count_increased = false;
> 
> -       if (!folio_test_hugetlb(folio))
> +       if (!folio_test_hugetlb(folio)) {
> +               if (folio_test_hugetlb_vmemmap_optimized(folio))
> +                       ret = -EBUSY;

The hugetlb specific page flags (HPG_vmemmap_optimized here) reside in
the folio->private field.

In the case where the folio is a non-hugetlb folio, the folio->private field
could be any arbitrary value.  As such, the test for vmemmap_optimized may
return a false positive.  We could end up retrying for an arbitrarily
long time.

I am looking at how to restructure the code which removes and frees
hugetlb pages so that folio_test_hugetlb() would remain true until
vmemmap pages are allocated.  The easiest way to do this would introduce
another hugetlb lock/unlock cycle in the page freeing path.  This would
undo some of the speedups in the series:
https://lore.kernel.org/all/20210409205254.242291-4-mike.kravetz@oracle.com/T/#m34321fbcbdf8bb35dfe083b05d445e90ecc1efab

-- 
Mike Kravetz

>                 goto out;
> +       }
> 
>         if (flags & MF_COUNT_INCREASED) {
>                 ret = 1;
> 
> 
> Thanks,
> Naoya Horiguchi
> 
> > -- 
> > Mike Kravetz
> > 
> > > 
> > > TestSetPageHWPoison in memory_failure is called after
> > > try_memory_failure_hugetlb, and folio_test_hugetlb is tested within
> > > __get_huge_page_for_hwpoison, which is wrapped by the hugetlb_lock. So
> > > by the time dissolve_free_huge_page returns, subpages already go
> > > through hugetlb_vmemmap_restore and __destroy_compound_gigantic_folio
> > > and become non-compound raw pages (folios). Now
> > > folio_test_hugetlb(p)=false will be correct for memory_failure, and it
> > > can recover p as a dissolved non-hugetlb page.

Powered by blists - more mailing lists