Re: [PATCH v1 3/7] mm: rename and move page_is_poisoned()

[Aili Yao <yaoaili@kingsoft.com>]

On Thu, 6 May 2021 09:06:14 +0200
Michal Hocko <mhocko@suse.com> wrote:

> On Thu 06-05-21 08:56:11, Aili Yao wrote:
> > On Wed, 5 May 2021 15:27:39 +0200
> > Michal Hocko <mhocko@suse.com> wrote:
> >   
> > > On Wed 05-05-21 15:17:53, David Hildenbrand wrote:  
> > > > On 05.05.21 15:13, Michal Hocko wrote:    
> > > > > On Thu 29-04-21 14:25:15, David Hildenbrand wrote:    
> > > > > > Commit d3378e86d182 ("mm/gup: check page posion status for coredump.")
> > > > > > introduced page_is_poisoned(), however, v5 [1] of the patch used
> > > > > > "page_is_hwpoison()" and something went wrong while upstreaming. Rename the
> > > > > > function and move it to page-flags.h, from where it can be used in other
> > > > > > -- kcore -- context.
> > > > > > 
> > > > > > Move the comment to the place where it belongs and simplify.
> > > > > > 
> > > > > > [1] https://lkml.kernel.org/r/20210322193318.377c9ce9@alex-virtual-machine
> > > > > > 
> > > > > > Signed-off-by: David Hildenbrand <david@redhat.com>    
> > > > > 
> > > > > I do agree that being explicit about hwpoison is much better. Poisoned
> > > > > page can be also an unitialized one and I believe this is the reason why
> > > > > you are bringing that up.    
> > > > 
> > > > I'm bringing it up because I want to reuse that function as state above :)
> > > >     
> > > > > 
> > > > > But you've made me look at d3378e86d182 and I am wondering whether this
> > > > > is really a valid patch. First of all it can leak a reference count
> > > > > AFAICS. Moreover it doesn't really fix anything because the page can be
> > > > > marked hwpoison right after the check is done. I do not think the race
> > > > > is feasible to be closed. So shouldn't we rather revert it?    
> > > > 
> > > > I am not sure if we really care about races here that much here? I mean,
> > > > essentially we are racing with HW breaking asynchronously. Just because we
> > > > would be synchronizing with SetPageHWPoison() wouldn't mean we can stop HW
> > > > from breaking.    
> > > 
> > > Right
> > >   
> > > > Long story short, this should be good enough for the cases we actually can
> > > > handle? What am I missing?    
> > > 
> > > I am not sure I follow. My point is that I fail to see any added value
> > > of the check as it doesn't prevent the race (it fundamentally cannot as
> > > the page can be poisoned at any time) but the failure path doesn't
> > > put_page which is incorrect even for hwpoison pages.  
> > 
> > Sorry, I have something to say:
> > 
> > I have noticed the ref count leak in the previous topic ,but  I don't think
> > it's a really matter. For memory recovery case for user pages, we will keep one
> > reference to the poison page so the error page will not be freed to buddy allocator.
> > which can be checked in memory_faulure() function.  
> 
> So what would happen if those pages are hwpoisoned from userspace rather
> than by HW. And repeatedly so?

Sorry, I may be not totally understand what you mean.

Do you mean hard page offline from mcelog?
If yes, I think it's not for one real UC error but for CE storms.
when we access this page in kernel, the access may success even it was marked hwpoison.

Thanks!
Aili Yao