target/linux/generic/pending-3.18/097-mm-remove-gup_flags-FOLL_WRITE-games-from-__get_user.patch

   1 From e45a502bdeae5a075257c4f061d1ff4ff0821354 Mon Sep 17 00:00:00 2001
   2 From: Linus Torvalds <torvalds@linux-foundation.org>
   3 Date: Thu, 13 Oct 2016 13:07:36 -0700
   4 Subject: [PATCH] mm: remove gup_flags FOLL_WRITE games from __get_user_pages()
   5
   6 [ Upstream commit 19be0eaffa3ac7d8eb6784ad9bdbc7d67ed8e619 ]
   7
   8 This is an ancient bug that was actually attempted to be fixed once
   9 (badly) by me eleven years ago in commit 4ceb5db9757a ("Fix
  10 get_user_pages() race for write access") but that was then undone due to
  11 problems on s390 by commit f33ea7f404e5 ("fix get_user_pages bug").
  12
  13 In the meantime, the s390 situation has long been fixed, and we can now
  14 fix it by checking the pte_dirty() bit properly (and do it better).  The
  15 s390 dirty bit was implemented in abf09bed3cce ("s390/mm: implement
  16 software dirty bits") which made it into v3.9.  Earlier kernels will
  17 have to look at the page state itself.
  18
  19 Also, the VM has become more scalable, and what used a purely
  20 theoretical race back then has become easier to trigger.
  21
  22 To fix it, we introduce a new internal FOLL_COW flag to mark the "yes,
  23 we already did a COW" rather than play racy games with FOLL_WRITE that
  24 is very fundamental, and then use the pte dirty flag to validate that
  25 the FOLL_COW flag is still valid.
  26
  27 Reported-and-tested-by: Phil "not Paul" Oester <kernel@linuxace.com>
  28 Acked-by: Hugh Dickins <hughd@google.com>
  29 Reviewed-by: Michal Hocko <mhocko@suse.com>
  30 Cc: Andy Lutomirski <luto@kernel.org>
  31 Cc: Kees Cook <keescook@chromium.org>
  32 Cc: Oleg Nesterov <oleg@redhat.com>
  33 Cc: Willy Tarreau <w@1wt.eu>
  34 Cc: Nick Piggin <npiggin@gmail.com>
  35 Cc: Greg Thelen <gthelen@google.com>
  36 Cc: stable@vger.kernel.org
  37 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
  38 Signed-off-by: Sasha Levin <alexander.levin@verizon.com>
  39 ---
  40  include/linux/mm.h |  1 +
  41  mm/gup.c           | 14 ++++++++++++--
  42  2 files changed, 13 insertions(+), 2 deletions(-)
  43
  44 --- a/include/linux/mm.h
  45 +++ b/include/linux/mm.h
  46 @@ -2029,6 +2029,7 @@ static inline struct page *follow_page(s
  47  #define FOLL_NUMA      0x200   /* force NUMA hinting page fault */
  48  #define FOLL_MIGRATION 0x400   /* wait for page to replace migration entry */
  49  #define FOLL_TRIED     0x800   /* a retry, previous pass started an IO */
  50 +#define FOLL_COW       0x4000  /* internal GUP flag */
  51
  52  typedef int (*pte_fn_t)(pte_t *pte, pgtable_t token, unsigned long addr,
  53                         void *data);
  54 --- a/mm/gup.c
  55 +++ b/mm/gup.c
  56 @@ -32,6 +32,16 @@ static struct page *no_page_table(struct
  57         return NULL;
  58  }
  59
  60 +/*
  61 + * FOLL_FORCE can write to even unwritable pte's, but only
  62 + * after we've gone through a COW cycle and they are dirty.
  63 + */
  64 +static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
  65 +{
  66 +       return pte_write(pte) ||
  67 +               ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
  68 +}
  69 +
  70  static struct page *follow_page_pte(struct vm_area_struct *vma,
  71                 unsigned long address, pmd_t *pmd, unsigned int flags)
  72  {
  73 @@ -66,7 +76,7 @@ retry:
  74         }
  75         if ((flags & FOLL_NUMA) && pte_numa(pte))
  76                 goto no_page;
  77 -       if ((flags & FOLL_WRITE) && !pte_write(pte)) {
  78 +       if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
  79                 pte_unmap_unlock(ptep, ptl);
  80                 return NULL;
  81         }
  82 @@ -315,7 +325,7 @@ static int faultin_page(struct task_stru
  83          * reCOWed by userspace write).
  84          */
  85         if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
  86 -               *flags &= ~FOLL_WRITE;
  87 +               *flags |= FOLL_COW;
  88         return 0;
  89  }
  90