Re: [PATCH v2 1/2] mm: Uncharge poisoned pages

[Naoya Horiguchi]

On Tue, May 09, 2017 at 11:18:23AM +0200, Michal Hocko wrote:
> On Mon 08-05-17 02:58:36, Naoya Horiguchi wrote:
> > On Tue, May 02, 2017 at 08:55:07PM +0200, Michal Hocko wrote:
> > > On Tue 02-05-17 16:59:30, Laurent Dufour wrote:
> > > > On 28/04/2017 15:48, Michal Hocko wrote:
> > > [...]
> > > > > This is getting quite hairy. What is the expected page count of the
> > > > > hwpoison page?
> > > 
> > > OK, so from the quick check of the hwpoison code it seems that the ref
> > > count will be > 1 (from get_hwpoison_page).
> > > 
> > > > > I guess we would need to update the VM_BUG_ON in the
> > > > > memcg uncharge code to ignore the page count of hwpoison pages if it can
> > > > > be arbitrary.
> > > > 
> > > > Based on the experiment I did, page count == 2 when isolate_lru_page()
> > > > succeeds, even in the case of a poisoned page.
> > > 
> > > that would make some sense to me. The page should have been already
> > > unmapped therefore but memory_failure increases the ref count and 1 is
> > > for isolate_lru_page().
> > 
> > # sorry for late reply, I was on holidays last week...
> > 
> > Right, and the refcount taken for memory_failure is not freed after
> > memory_failure() returns. unpoison_memory() does free the refcount.
> 
> OK, from the charge POV this would be safe because we clear page->memcg
> so it wouldn't get uncharged more times.
> 
> > > > In my case I think this
> > > > is because the page is still used by the process which is calling madvise().
> > > > 
> > > > I'm wondering if I'm looking at the right place. May be the poisoned
> > > > page should remain attach to the memory_cgroup until no one is using it.
> > > > In that case this means that something should be done when the page is
> > > > off-lined... I've to dig further here.
> > > 
> > > No, AFAIU the page will not drop the reference count down to 0 in most
> > > cases. Maybe there are some scenarios where this can happen but I would
> > > expect that the poisoned page will be mapped and in use most of the time
> > > and won't drop down 0. And then we should really uncharge it because it
> > > will pin the memcg and make it unfreeable which doesn't seem to be what
> > > we want.  So does the following work reasonable? Andi, Johannes, what do
> > > you think? I cannot say I would be really comfortable touching hwpoison
> > > code as I really do not understand the workflow. Maybe we want to move
> > > this uncharge down to memory_failure() right before we report success?
> > 
> > memory_failure() can be called for any types of page (including slab or
> > any kernel/driver pages), and the reported problem seems happen only on
> > in-use user pages, so uncharging in delete_from_lru_cache() as done below
> > looks better to me.
> 
> Yeah, we do see problems only for LRU/page cache pages but my
> understanding is that error_states (e.g. me_kernel for the kernel
> memory) might change in the future and then we wouldn't catch the same
> bug, no?

Right about future change, and we will see the same bug. I guess that the
first target of kernel page is slab page, and memcg_kmem_uncharge() will
be used there. Implementors/Reviewers should care about uncharging when the
time comes.

Thanks,
Naoya Horiguchi

> 
> > > ---
> > > From 8bf0791bcf35996a859b6d33fb5494e5b53de49d Mon Sep 17 00:00:00 2001
> > > From: Michal Hocko <mhocko@xxxxxxxx>
> > > Date: Tue, 2 May 2017 20:32:24 +0200
> > > Subject: [PATCH] hwpoison, memcg: forcibly uncharge LRU pages
> > > 
> > > Laurent Dufour has noticed that hwpoinsoned pages are kept charged. In
> > > his particular case he has hit a bad_page("page still charged to cgroup")
> > > when onlining a hwpoison page.
> > 
> > > While this looks like something that shouldn't
> > > happen in the first place because onlining hwpages and returning them to
> > > the page allocator makes only little sense it shows a real problem.
> > > 
> > > hwpoison pages do not get freed usually so we do not uncharge them (at
> > > least not since 0a31bc97c80c ("mm: memcontrol: rewrite uncharge API")).
> > > Each charge pins memcg (since e8ea14cc6ead ("mm: memcontrol: take a css
> > > reference for each charged page")) as well and so the mem_cgroup and the
> > > associated state will never go away. Fix this leak by forcibly
> > > uncharging a LRU hwpoisoned page in delete_from_lru_cache(). We also
> > > have to tweak uncharge_list because it cannot rely on zero ref count
> > > for these pages.
> > > 
> > > Fixes: 0a31bc97c80c ("mm: memcontrol: rewrite uncharge API")
> > > Reported-by: Laurent Dufour <ldufour@xxxxxxxxxxxxxxxxxx>
> > > Signed-off-by: Michal Hocko <mhocko@xxxxxxxx>
> > 
> > Reviewed-by: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>
> 
> Thanks! I will wait a day or two for Johannes and repost the patch.
> Andrew could you drop
> http://www.ozlabs.org/~akpm/mmotm/broken-out/mm-uncharge-poisoned-pages.patch
> in the mean time, please?
> 
> -- 
> Michal Hocko
> SUSE Labs
>