--- a/include/linux/mm.h 2016-08-25 18:07:47.000000000 +0300 +++ b/include/linux/mm.h 2016-10-24 16:19:16.000000000 +0300 @@ -1401,6 +1401,7 @@ #define FOLL_HWPOISON 0x100 /* check page is hwpoisoned */ #define FOLL_NUMA 0x200 /* force NUMA hinting page fault */ #define FOLL_MIGRATION 0x400 /* wait for page to replace migration entry */ +#define FOLL_COW 0x4000 /* internal GUP flag */ typedef int (*pte_fn_t)(pte_t *pte, pgtable_t token, unsigned long addr, void *data); diff -Naur ./mm/memory.c ../linux-2.6.32-642.6.2.el6/mm/memory.c --- a/mm/memory.c 2016-08-25 18:06:57.000000000 +0300 +++ b/mm/memory.c 2016-10-24 16:19:16.000000000 +0300 @@ -1191,6 +1191,24 @@ } EXPORT_SYMBOL_GPL(zap_vma_ptes); +static inline bool can_follow_write_pte(pte_t pte, struct page *page, + unsigned int flags) +{ + if (pte_write(pte)) + return true; + + /* + * Make sure that we are really following CoWed page. We do not really + * have to care about exclusiveness of the page because we only want + * to ensure that once COWed page hasn't disappeared in the meantime + * or it hasn't been merged to a KSM page. + */ + if ((flags & FOLL_FORCE) && (flags & FOLL_COW)) + return page && PageAnon(page) && !PageKsm(page); + + return false; +} + /* * Do a quick page-table lookup for a single page. */ @@ -1280,10 +1298,11 @@ migration_entry_wait(mm, pmd, address); goto split_fallthrough; } - if ((flags & FOLL_WRITE) && !pte_write(pte)) - goto unlock; - page = vm_normal_page(vma, address, pte); + if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, page, flags)) { + pte_unmap_unlock(ptep, ptl); + return NULL; + } if (unlikely(!page)) { if ((flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(pte))) @@ -1304,7 +1323,6 @@ */ mark_page_accessed(page); } -unlock: pte_unmap_unlock(ptep, ptl); out: return page; @@ -1503,17 +1520,13 @@ * The VM_FAULT_WRITE bit tells us that * do_wp_page has broken COW when necessary, * even if maybe_mkwrite decided not to set - * pte_write. We can thus safely do subsequent - * page lookups as if they were reads. But only - * do so when looping for pte_write is futile: - * in some cases userspace may also be wanting - * to write to the gotten user page, which a - * read fault here might prevent (a readonly - * page might get reCOWed by userspace write). + * pte_write. We cannot simply drop FOLL_WRITE + * here because the COWed page might be gone by + * the time we do the subsequent page lookups. */ if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE)) - foll_flags &= ~FOLL_WRITE; + foll_flags |= FOLL_COW; cond_resched(); }