On Wed, Sep 12, 2018 at 02:49:21PM +0800, Peter Xu wrote: > Add an extra check on page dirty bit in change_pte_range() since there > might be case where PTE dirty bit is unset but it's actually dirtied. > One example is when a huge PMD is splitted after written: the dirty bit > will be set on the compound page however we won't have the dirty bit set > on each of the small page PTEs. > > I noticed this when debugging with a customized kernel that implemented > userfaultfd write-protect. In that case, the dirty bit will be critical > since that's required for userspace to handle the write protect page > fault (otherwise it'll get a SIGBUS with a loop of page faults). > However it should still be good even for upstream Linux to cover more > scenarios where we shouldn't need to do extra page faults on the small > pages if the previous huge page is already written, so the dirty bit > optimization path underneath can cover more. > > CC: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> > CC: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx> > CC: Khalid Aziz <khalid.aziz@xxxxxxxxxx> > CC: Thomas Gleixner <tglx@xxxxxxxxxxxxx> > CC: "David S. Miller" <davem@xxxxxxxxxxxxx> > CC: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> > CC: Andi Kleen <ak@xxxxxxxxxxxxxxx> > CC: Henry Willard <henry.willard@xxxxxxxxxx> > CC: Anshuman Khandual <khandual@xxxxxxxxxxxxxxxxxx> > CC: Andrea Arcangeli <aarcange@xxxxxxxxxx> > CC: Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> > CC: Jerome Glisse <jglisse@xxxxxxxxxx> > CC: Zi Yan <zi.yan@xxxxxxxxxxxxxx> > CC: linux-mm@xxxxxxxxx > CC: linux-kernel@xxxxxxxxxxxxxxx > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx> > --- > v2: > - checking the dirty bit when changing PTE entries rather than fixing up > the dirty bit when splitting the huge page PMD. > - rebase to 4.19-rc3 > > Instead of keeping this in my local tree, I'm giving it another shot to > see whether this could be acceptable for upstream since IMHO it should > still benefit the upstream. Thanks, > --- > mm/mprotect.c | 11 +++++++++++ > 1 file changed, 11 insertions(+) > > diff --git a/mm/mprotect.c b/mm/mprotect.c > index 6d331620b9e5..5fe752515161 100644 > --- a/mm/mprotect.c > +++ b/mm/mprotect.c > @@ -115,6 +115,17 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd, > if (preserve_write) > ptent = pte_mk_savedwrite(ptent); > > + /* > + * The extra PageDirty() check will make sure > + * we'll capture the dirty page even if the PTE > + * dirty bit is unset. One case is when the > + * PTE is splitted from a huge PMD, in that > + * case the dirty flag might only be set on the > + * compound page instead of this PTE. > + */ > + if (PageDirty(pte_page(ptent))) > + ptent = pte_mkdirty(ptent); > + How do you protect against concurent clearing of PG_dirty? You can end up with unaccounted dirty page. NAK. -- Kirill A. Shutemov
