Date:	Wed, 17 Jun 2009 23:08:13 +0900
From:	Minchan Kim <minchan.kim@...il.com>
To:	Wu Fengguang <fengguang.wu@...el.com>
Cc:	Andrew Morton <akpm@...ux-foundation.org>,
	LKML <linux-kernel@...r.kernel.org>,
	Andi Kleen <ak@...ux.intel.com>, Ingo Molnar <mingo@...e.hu>,
	Mel Gorman <mel@....ul.ie>,
	Thomas Gleixner <tglx@...utronix.de>,
	"H. Peter Anvin" <hpa@...or.com>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Nick Piggin <npiggin@...e.de>,
	Hugh Dickins <hugh.dickins@...cali.co.uk>,
	Andi Kleen <andi@...stfloor.org>,
	"riel@...hat.com" <riel@...hat.com>,
	"chris.mason@...cle.com" <chris.mason@...cle.com>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>
Subject: Re: [PATCH 09/22] HWPOISON: Handle hardware poisoned pages in 
	try_to_unmap

On Wed, Jun 17, 2009 at 11:03 PM, Wu Fengguang<fengguang.wu@...el.com> wrote:
> On Wed, Jun 17, 2009 at 09:43:29PM +0800, Minchan Kim wrote:
>> On Wed, Jun 17, 2009 at 10:37 PM, Wu Fengguang<fengguang.wu@...el.com> wrote:
>> > On Wed, Jun 17, 2009 at 09:27:36PM +0800, Minchan Kim wrote:
>> >> On Wed, Jun 17, 2009 at 4:23 PM, Wu Fengguang<fengguang.wu@...el.com> wrote:
>> >> > On Wed, Jun 17, 2009 at 08:28:26AM +0800, Minchan Kim wrote:
>> >> >> On Tue, 16 Jun 2009 21:49:44 +0800
>> >> >> Wu Fengguang <fengguang.wu@...el.com> wrote:
>> >> >>
>> >> >> > On Tue, Jun 16, 2009 at 08:03:08AM +0800, Minchan Kim wrote:
>> >> >> > > On Mon, 15 Jun 2009 23:26:12 +0800
>> >> >> > > Wu Fengguang <fengguang.wu@...el.com> wrote:
>> >> >> > >
>> >> >> > > > On Mon, Jun 15, 2009 at 09:09:03PM +0800, Minchan Kim wrote:
>> >> >> > > > > On Mon, Jun 15, 2009 at 11:45 AM, Wu Fengguang<fengguang.wu@...el.com> wrote:
>> >> >> > > > > > From: Andi Kleen <ak@...ux.intel.com>
>> >> >> > > > > >
>> >> >> > > > > > When a page has the poison bit set replace the PTE with a poison entry.
>> >> >> > > > > > This causes the right error handling to be done later when a process runs
>> >> >> > > > > > into it.
>> >> >> > > > > >
>> >> >> > > > > > Also add a new flag to not do that (needed for the memory-failure handler
>> >> >> > > > > > later)
>> >> >> > > > > >
>> >> >> > > > > > Reviewed-by: Wu Fengguang <fengguang.wu@...el.com>
>> >> >> > > > > > Signed-off-by: Andi Kleen <ak@...ux.intel.com>
>> >> >> > > > > >
>> >> >> > > > > > ---
>> >> >> > > > > >  include/linux/rmap.h |    1 +
>> >> >> > > > > >  mm/rmap.c            |    9 ++++++++-
>> >> >> > > > > >  2 files changed, 9 insertions(+), 1 deletion(-)
>> >> >> > > > > >
>> >> >> > > > > > --- sound-2.6.orig/mm/rmap.c
>> >> >> > > > > > +++ sound-2.6/mm/rmap.c
>> >> >> > > > > > @@ -958,7 +958,14 @@ static int try_to_unmap_one(struct page
>> >> >> > > > > >        /* Update high watermark before we lower rss */
>> >> >> > > > > >        update_hiwater_rss(mm);
>> >> >> > > > > >
>> >> >> > > > > > -       if (PageAnon(page)) {
>> >> >> > > > > > +       if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
>> >> >> > > > > > +               if (PageAnon(page))
>> >> >> > > > > > +                       dec_mm_counter(mm, anon_rss);
>> >> >> > > > > > +               else if (!is_migration_entry(pte_to_swp_entry(*pte)))
>> >> >> > > > >
>> >> >> > > > > Isn't it straightforward to use !is_hwpoison_entry ?
>> >> >> > > >
>> >> >> > > > Good catch!  It looks like a redundant check: the
>> >> >> > > > page_check_address() at the beginning of the function guarantees that
>> >> >> > > > !is_migration_entry() or !is_migration_entry() tests will all be TRUE.
>> >> >> > > > So let's do this?
>> >> >> > > It seems you expand my sight :)
>> >> >> > >
>> >> >> > > I don't know migration well.
>> >> >> > > How page_check_address guarantee it's not migration entry ?
>> >> >> >
>> >> >> > page_check_address() calls pte_present() which returns the
>> >> >> > (_PAGE_PRESENT | _PAGE_PROTNONE) bits. While x86-64 defines
>> >> >> >
>> >> >> > #define __swp_entry(type, offset)       ((swp_entry_t) { \
>> >> >> >                                          ((type) << (_PAGE_BIT_PRESENT + 1)) \
>> >> >> >                                          | ((offset) << SWP_OFFSET_SHIFT) })
>> >> >> >
>> >> >> > where SWP_OFFSET_SHIFT is defined to the bigger one of
>> >> >> > max(_PAGE_BIT_PROTNONE + 1, _PAGE_BIT_FILE + 1) = max(8+1, 6+1) = 9.
>> >> >> >
>> >> >> > So __swp_entry(type, offset) := (type << 1) | (offset << 9)
>> >> >> >
>> >> >> > We know that the swap type is 5 bits. So the bit 0 _PAGE_PRESENT and bit 8
>> >> >> > _PAGE_PROTNONE will all be zero for swap entries.
>> >> >> >
>> >> >>
>> >> >> Thanks for kind explanation :)
>> >> >
>> >> > You are welcome~
>> >> >
>> >> >> >
>> >> >> > > In addtion, If the page is poison while we are going to
>> >> >> > > migration((PAGE_MIGRATION && migration) == TRUE), we should decrease
>> >> >> > > file_rss ?
>> >> >> >
>> >> >> > It will die on trying to migrate the poisoned page so we don't care
>> >> >> > the accounting. But normally the poisoned page shall already be
>> >> >>
>> >> >>
>> >> >> Okay. then, how about this ?
>> >> >> We should not increase file_rss on trying to migrate the poisoned page
>> >> >>
>> >> >> -               else if (!is_migration_entry(pte_to_swp_entry(*pte)))
>> >> >> +               else if (!(PAGE_MIGRATION && migration))
>> >> >
>> >> > This is good if we are going to stop the hwpoison page from being
>> >> > consumed by move_to_new_page(), but I highly doubt we'll ever add
>> >> > PageHWPoison() checks into the migration code.
>> >> >
>> >> > Because this race window is small enough:
>> >> >
>> >> >        TestSetPageHWPoison(p);
>> >> >                                   lock_page(page);
>> >> >                                   try_to_unmap(page, TTU_MIGRATION|...);
>> >> >        lock_page_nosync(p);
>> >> >
>> >> > such small race windows can be found all over the kernel, it's just
>> >> > insane to try to fix any of them.
>> >>
>> >> Sorry for too late response.
>> >>
>> >> I see your point.
>> >> My opinion is that at least we must be notified when such situation happen.
>> >> So I think it would be better to add some warning to fix up it when it
>> >> happen even thought  it is small race window.
>> >
>> > Notification is also pointless here: we'll die hard on
>> > accessing/consuming the poisoned page anyway :(
>>
>> My intention wasn't to recover it.
>
> Yes, that's not the point.
>
>> It just add something like WARN_ON.
>> You said it is small window enough. but I think it can happen more
>> hight probability in migration-workload.(At a moment, I don't know
>> what kinds of app)
>> For such case, If we can hear reporting of warning, at that time we
>> can consider migration handling for HWPoison.
>
> The point is, any page can go corrupted any time. We don't need to add
> 1000 PageHWPoison() tests in the kernel like this. We don't aim for
> 100% protection, that's impossible. I'd be very contented if ever it
> can reach 80% coverage :)

Okay.
If it is your goal, I also think migration portion of all is very small.
Thanks for kind reply for my boring discussion.

Reviewed-by : Minchan Kim <minchan.kim@...il.com>

> Thanks,
> Fengguang
>
>



-- 
Kinds regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists