Re: [PATCH 3/3] mm/memory_hotplug: make HWPoisoned dirty swapcache pages unmovable

[Andrew Morton]

On Mon, 23 Aug 2021 17:14:29 +0800 Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:

> On 2021/8/23 16:26, HORIGUCHI NAOYA(堀口 直也) wrote:
> > On Sat, Aug 21, 2021 at 05:42:46PM +0800, Miaohe Lin wrote:
> >> HWPoisoned dirty swapcache pages are kept for killing owner processes.
> >> We should not offline these pages or do_swap_page() would access the
> >> offline pages and lead to bad ending.
> >>
> > 
> > Thank you for the report.  I'm not yet sure of the whole picture of this
> > issue.  do_swap_page() is expected to return with fault VM_FAULT_HWPOISON
> > when called via the access to the error page, so I wonder why this doesn't
> > work for your situation.  And what is the "bad ending" in the description?
> > 
> 
> IMO we might hotremove the page while SwapCache still have ref to it. Thus the page
> struct would be accessed after offlined. The page struct should be invalid in this case
> and this would make do_swap_page fragile. Or am I miss something?
> 
> > I feel that aborting memory hotremove due to a hwpoisoned dirty swapcache
> > might be too hard, so I'd like to find another solution if we have.
> 
> If there is a better way, we can just drop this one.
> 
> Many thanks for your review and reply! :)
> 
> > # You may separate this patch from former two to make them merged to
> > # mainline soon.
>
> ...
>
> >> --- a/mm/memory_hotplug.c
> >> +++ b/mm/memory_hotplug.c
> >> @@ -1664,6 +1664,12 @@ static int scan_movable_pages(unsigned long start, unsigned long end,
> >>  		 */
> >>  		if (PageOffline(page) && page_count(page))
> >>  			return -EBUSY;
> >> +		/*
> >> +		 * HWPoisoned dirty swapcache pages are definitely unmovable
> >> +		 * because they are kept for killing owner processes.
> >> +		 */
> >> +		if (PageHWPoison(page) && PageSwapCache(page))
> >> +			return -EBUSY;
> 

I'll drop this.  Please resend something if you still believe that
changes are desirable.