On Mon, 23 Jan 2017 15:15:52 -0800 Shaohua Li <shli@xxxxxx> wrote: > The page reclaim has an assumption writting to a page with clean pte > should trigger a page fault, because there is a window between pte zero > and tlb flush where a new write could come. If the new write doesn't > trigger page fault, page reclaim will not notice it and think the page > is clean and reclaim it. The MADV_FREE pages don't comply with the rule > and the pte is just cleaned without writeprotect, so there will be no > pagefault for new write. This will cause data corruption. I'd like to see here a complete description of the bug's effects: waht sort of workload will trigger it, what the end-user visible effects are, etc. > --- a/mm/huge_memory.c > +++ b/mm/huge_memory.c > @@ -1381,6 +1381,7 @@ bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, > tlb->fullmm); > orig_pmd = pmd_mkold(orig_pmd); > orig_pmd = pmd_mkclean(orig_pmd); > + orig_pmd = pmd_wrprotect(orig_pmd); Is this the right way round? There's still a window where we won't get that write fault on the cleaned pte. Should the pmd_wrprotect() happen before the pmd_mkclean()? > set_pmd_at(mm, addr, pmd, orig_pmd); > tlb_remove_pmd_tlb_entry(tlb, pmd, addr); > diff --git a/mm/madvise.c b/mm/madvise.c > index 0e3828e..bfb6800 100644 > --- a/mm/madvise.c > +++ b/mm/madvise.c > @@ -373,6 +373,7 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr, > > ptent = pte_mkold(ptent); > ptent = pte_mkclean(ptent); > + ptent = pte_wrprotect(ptent); > set_pte_at(mm, addr, pte, ptent); > if (PageActive(page)) > deactivate_page(page); -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
