lkml.org 
[lkml]   [2009]   [Jun]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 09/22] HWPOISON: Handle hardware poisoned pages in try_to_unmap
    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@intel.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@intel.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@intel.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@intel.com> wrote:
    > >> > > > > > From: Andi Kleen <ak@linux.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@intel.com>
    > >> > > > > > Signed-off-by: Andi Kleen <ak@linux.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 :(

    Thanks,
    Fengguang

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

    \
     
     \ /
      Last update: 2009-06-17 15:41    [W:0.031 / U:31.664 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site