Re: [PATCH v2 4/9] mm, hwpoison, hugetlb: support saving mechanism of raw error pages

[Muchun Song <songmuchun@bytedance.com>]

On Tue, Jun 28, 2022 at 02:41:22AM +0000, HORIGUCHI NAOYA(堀口 直也) wrote:
> On Mon, Jun 27, 2022 at 05:26:01PM +0800, Muchun Song wrote:
> > On Fri, Jun 24, 2022 at 08:51:48AM +0900, Naoya Horiguchi wrote:
> > > From: Naoya Horiguchi <naoya.horiguchi@nec.com>
> > > 
> > > When handling memory error on a hugetlb page, the error handler tries to
> > > dissolve and turn it into 4kB pages.  If it's successfully dissolved,
> > > PageHWPoison flag is moved to the raw error page, so that's all right.
> > > However, dissolve sometimes fails, then the error page is left as
> > > hwpoisoned hugepage. It's useful if we can retry to dissolve it to save
> > > healthy pages, but that's not possible now because the information about
> > > where the raw error pages is lost.
> > > 
> > > Use the private field of a few tail pages to keep that information.  The
> > > code path of shrinking hugepage pool uses this info to try delayed dissolve.
> > > In order to remember multiple errors in a hugepage, a singly-linked list
> > > originated from SUBPAGE_INDEX_HWPOISON-th tail page is constructed.  Only
> > > simple operations (adding an entry or clearing all) are required and the
> > > list is assumed not to be very long, so this simple data structure should
> > > be enough.
> > > 
> > > If we failed to save raw error info, the hwpoison hugepage has errors on
> > > unknown subpage, then this new saving mechanism does not work any more,
> > > so disable saving new raw error info and freeing hwpoison hugepages.
> > > 
> > > Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
> > 
> > Thanks for your work on this. I have several quesions below.
> > 
> ...
> 
> > > @@ -1499,6 +1499,97 @@ static int try_to_split_thp_page(struct page *page, const char *msg)
> > >  }
> > >  
> > >  #ifdef CONFIG_HUGETLB_PAGE
> > > +/*
> > > + * Struct raw_hwp_page represents information about "raw error page",
> > > + * constructing singly linked list originated from ->private field of
> > > + * SUBPAGE_INDEX_HWPOISON-th tail page.
> > > + */
> > > +struct raw_hwp_page {
> > > +	struct llist_node node;
> > > +	struct page *page;
> > > +};
> > > +
> > > +static inline struct llist_head *raw_hwp_list_head(struct page *hpage)
> > > +{
> > > +	return (struct llist_head *)&page_private(hpage + SUBPAGE_INDEX_HWPOISON);
> > > +}
> > > +
> > > +static inline int raw_hwp_unreliable(struct page *hpage)
> > > +{
> > > +	return page_private(hpage + SUBPAGE_INDEX_HWPOISON_UNRELIABLE);
> > > +}
> > > +
> > > +static inline void set_raw_hwp_unreliable(struct page *hpage)
> > > +{
> > > +	set_page_private(hpage + SUBPAGE_INDEX_HWPOISON_UNRELIABLE, 1);
> > > +}
> > 
> > Why not use HPAGEFLAG(HwpoisonUnreliable, hwpoison_unreliable) directly?
> > 
> 
> OK, that sounds better, I'll do it.
> 
> > > +
> > > +/*
> > > + * about race consideration
> > > + */
> > > +static inline int hugetlb_set_page_hwpoison(struct page *hpage,
> > > +					struct page *page)
> > > +{
> > > +	struct llist_head *head;
> > > +	struct raw_hwp_page *raw_hwp;
> > > +	struct llist_node *t, *tnode;
> > > +	int ret;
> > > +
> > > +	/*
> > > +	 * Once the hwpoison hugepage has lost reliable raw error info,
> > > +	 * there is little mean to keep additional error info precisely,
> > > +	 * so skip to add additional raw error info.
> > > +	 */
> > > +	if (raw_hwp_unreliable(hpage))
> > > +		return -EHWPOISON;
> > > +	head = raw_hwp_list_head(hpage);
> > > +	llist_for_each_safe(tnode, t, head->first) {
> > > +		struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node);
> > > +
> > > +		if (p->page == page)
> > > +			return -EHWPOISON;
> > > +	}
> > > +
> > > +	ret = TestSetPageHWPoison(hpage) ? -EHWPOISON : 0;
> > > +	/* the first error event will be counted in action_result(). */
> > > +	if (ret)
> > > +		num_poisoned_pages_inc();
> > > +
> > > +	raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_KERNEL);
> > 
> > This function can be called in atomic context, GFP_ATOMIC should be used
> > here.
> 
> OK, I'll use GFP_ATOMIC.
> 
> > 
> > > +	if (raw_hwp) {
> > > +		raw_hwp->page = page;
> > > +		llist_add(&raw_hwp->node, head);
> > 
> > The maximum amount of items in the list is one, right?
> 
> The maximum is the number of subpages in the hugepage (512 for 2MB hugepage,
> 262144 for 1GB hugepage).
>

You are right. I have missed something yesterday.
 
> > 
> > > +	} else {
> > > +		/*
> > > +		 * Failed to save raw error info.  We no longer trace all
> > > +		 * hwpoisoned subpages, and we need refuse to free/dissolve
> > > +		 * this hwpoisoned hugepage.
> > > +		 */
> > > +		set_raw_hwp_unreliable(hpage);
> > > +		return ret;
> > > +	}
> > > +	return ret;
> > > +}
> > > +
> > > +inline int hugetlb_clear_page_hwpoison(struct page *hpage)
> > > +{
> > > +	struct llist_head *head;
> > > +	struct llist_node *t, *tnode;
> > > +
> > > +	if (raw_hwp_unreliable(hpage))
> > > +		return -EBUSY;
> > 
> > IIUC, we use head page's PageHWPoison to synchronize hugetlb_clear_page_hwpoison()
> > and hugetlb_set_page_hwpoison(), right? If so, who can set hwp_unreliable here?
> 
> Sorry if I might miss your point, but raw_hwp_unreliable is set when
> allocating raw_hwp_page failed.  hugetlb_set_page_hwpoison() can be called

Sorry. I have missed this. Thanks for your clarification.

> multiple times on a hugepage and if one of the calls fails, the hwpoisoned
> hugepage becomes unreliable.
> 
> BTW, as you pointed out above, if we switch to passing GFP_ATOMIC to kmalloc(),
> the kmalloc() never fails, so we no longer have to implement this unreliable

No. kmalloc() with GFP_ATOMIC can fail unless I miss something important.

> flag, so things get simpler.
> 
> > 
> > > +	ClearPageHWPoison(hpage);
> > > +	head = raw_hwp_list_head(hpage);
> > > +	llist_for_each_safe(tnode, t, head->first) {
> > 
> > Is it possible that a new item is added hugetlb_set_page_hwpoison()  and we do not
> > traverse it (we have cleared page's PageHWPoison)? Then we ignored a real hwpoison
> > page, right?
> 
> Maybe you are mentioning the race like below. Yes, that's possible.
>

Sorry, ignore my previous comments, I'm thinking something wrong.

>   CPU 0                            CPU 1
> 
>                                    free_huge_page
>                                      lock hugetlb_lock
>                                      ClearHPageMigratable
				       remove_hugetlb_page()
				       // the page is non-HugeTLB now
>                                      unlock hugetlb_lock
>   get_huge_page_for_hwpoison
>     lock hugetlb_lock
>     __get_huge_page_for_hwpoison

	// cannot reach here since it is not a HugeTLB page now.
	// So this race is impossible. Then we fallback to normal
	// page handling. Seems there is a new issue here.
	//
	// memory_failure()
	//	try_memory_failure_hugetlb()
	//	if (hugetlb)
	//		goto unlock_mutex;
	//	if (TestSetPageHWPoison(p)) {
	//	// This non-HugeTLB page's vmemmap is still optimized.
	
Setting COMPOUND_PAGE_DTOR after hugetlb_vmemmap_restore() might fix this
issue, but we will encounter this race as you mentioned below.

Thanks.
	
>       hugetlb_set_page_hwpoison
>         allocate raw_hwp_page
>         TestSetPageHWPoison
>                                      update_and_free_page
>                                        __update_and_free_page
>                                          if (PageHWPoison)
>                                            hugetlb_clear_page_hwpoison
>                                              TestClearPageHWPoison
>                                              // remove all list items
>         llist_add
>     unlock hugetlb_lock
> 
> 
> The end result seems not critical (leaking raced raw_hwp_page?), but
> we need fix.
> 
> > 
> > > +		struct raw_hwp_page *p = container_of(tnode, struct raw_hwp_page, node);
> > > +
> > > +		SetPageHWPoison(p->page);
> > > +		kfree(p);
> > > +	}
> > > +	llist_del_all(head);
> > 
> > If the above issue exists, moving ClearPageHWPoison(hpage) to here could
> > fix it. We should clear PageHWPoison carefully since the head page is
> > also can be poisoned.
> 
> The reason why I put ClearPageHWPoison(hpage) before llist_for_each_safe()
> was that raw error page can be the head page.  But this can be solved
> with some additional code to remember whether raw_hwp_page list has an item
> associated with the head page.
>
> Or another approach in my mind now is to call hugetlb_clear_page_hwpoison()
> with taking mf_mutex.
> 
> > 
> > Thanks.
> 
> Thank you for valuable feedbacks.
> 
> - Naoya Horiguchi
> 
> > 
> > > +	return 0;
> > > +}
> > > +
> > >  /*
> > >   * Called from hugetlb code with hugetlb_lock held.
> > >   *
> > > @@ -1533,7 +1624,7 @@ int __get_huge_page_for_hwpoison(unsigned long pfn, int flags)
> > >  		goto out;
> > >  	}
> > >  
> > > -	if (TestSetPageHWPoison(head)) {
> > > +	if (hugetlb_set_page_hwpoison(head, page)) {
> > >  		ret = -EHWPOISON;
> > >  		goto out;
> > >  	}
> > > @@ -1585,7 +1676,7 @@ static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb
> > >  	lock_page(head);
> > >  
> > >  	if (hwpoison_filter(p)) {
> > > -		ClearPageHWPoison(head);
> > > +		hugetlb_clear_page_hwpoison(head);
> > >  		res = -EOPNOTSUPP;
> > >  		goto out;
> > >  	}
> > > -- 
> > > 2.25.1
> > > 
> > >