On 11/29/2016 01:15 AM, Linus Torvalds wrote:

However, I also independently think I found an actual bug while
looking at the code as part of looking at the patch.

This part looks racy:

                /*
                 * We are remapping a dirty PTE, make sure to
                 * flush TLB before we drop the PTL for the
                 * old PTE or we may race with page_mkclean().
                 */
                if (pte_present(*old_pte) && pte_dirty(*old_pte))
                        force_flush = true;
                pte = ptep_get_and_clear(mm, old_addr, old_pte);

where the issue is that another thread might make the pte be dirty (in
the hardware walker, so no locking of ours make any difference)
*after* we checked whether it was dirty, but *before* we removed it
from the page tables.

Ah, very right. Thanks for the catch!

So I think the "check for force-flush" needs to come *after*, and we should do

                pte = ptep_get_and_clear(mm, old_addr, old_pte);
                if (pte_present(pte) && pte_dirty(pte))
                        force_flush = true;

instead.

This happens for the pmd case too.

Here is a fix patch, sorry for the trouble.