mremap: fix race between mremap() and page cleanning
authorAaron Lu <aaron.lu@intel.com>
Thu, 10 Nov 2016 09:16:33 +0000 (17:16 +0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 17 Nov 2016 17:46:56 +0000 (09:46 -0800)
Prior to 3.15, there was a race between zap_pte_range() and
page_mkclean() where writes to a page could be lost.  Dave Hansen
discovered by inspection that there is a similar race between
move_ptes() and page_mkclean().

We've been able to reproduce the issue by enlarging the race window with
a msleep(), but have not been able to hit it without modifying the code.
So, we think it's a real issue, but is difficult or impossible to hit in
practice.

The zap_pte_range() issue is fixed by commit 1cf35d47712d("mm: split
'tlb_flush_mmu()' into tlb flushing and memory freeing parts").  And
this patch is to fix the race between page_mkclean() and mremap().

Here is one possible way to hit the race: suppose a process mmapped a
file with READ | WRITE and SHARED, it has two threads and they are bound
to 2 different CPUs, e.g.  CPU1 and CPU2.  mmap returned X, then thread
1 did a write to addr X so that CPU1 now has a writable TLB for addr X
on it.  Thread 2 starts mremaping from addr X to Y while thread 1
cleaned the page and then did another write to the old addr X again.
The 2nd write from thread 1 could succeed but the value will get lost.

        thread 1                           thread 2
     (bound to CPU1)                    (bound to CPU2)

  1: write 1 to addr X to get a
     writeable TLB on this CPU

                                        2: mremap starts

                                        3: move_ptes emptied PTE for addr X
                                           and setup new PTE for addr Y and
                                           then dropped PTL for X and Y

  4: page laundering for N by doing
     fadvise FADV_DONTNEED. When done,
     pageframe N is deemed clean.

  5: *write 2 to addr X

                                        6: tlb flush for addr X

  7: munmap (Y, pagesize) to make the
     page unmapped

  8: fadvise with FADV_DONTNEED again
     to kick the page off the pagecache

  9: pread the page from file to verify
     the value. If 1 is there, it means
     we have lost the written 2.

  *the write may or may not cause segmentation fault, it depends on
  if the TLB is still on the CPU.

Please note that this is only one specific way of how the race could
occur, it didn't mean that the race could only occur in exact the above
config, e.g. more than 2 threads could be involved and fadvise() could
be done in another thread, etc.

For anonymous pages, they could race between mremap() and page reclaim:
THP: a huge PMD is moved by mremap to a new huge PMD, then the new huge
PMD gets unmapped/splitted/pagedout before the flush tlb happened for
the old huge PMD in move_page_tables() and we could still write data to
it.  The normal anonymous page has similar situation.

To fix this, check for any dirty PTE in move_ptes()/move_huge_pmd() and
if any, did the flush before dropping the PTL.  If we did the flush for
every move_ptes()/move_huge_pmd() call then we do not need to do the
flush in move_pages_tables() for the whole range.  But if we didn't, we
still need to do the whole range flush.

Alternatively, we can track which part of the range is flushed in
move_ptes()/move_huge_pmd() and which didn't to avoid flushing the whole
range in move_page_tables().  But that would require multiple tlb
flushes for the different sub-ranges and should be less efficient than
the single whole range flush.

KBuild test on my Sandybridge desktop doesn't show any noticeable change.
v4.9-rc4:
  real    5m14.048s
  user    32m19.800s
  sys     4m50.320s

With this commit:
  real    5m13.888s
  user    32m19.330s
  sys     4m51.200s

Reported-by: Dave Hansen <dave.hansen@intel.com>
Signed-off-by: Aaron Lu <aaron.lu@intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/huge_mm.h
mm/huge_memory.c
mm/mremap.c

index 9b9f65d9987393d456911f41eacb4bdfa9fe0284..e35e6de633b9a7bc2a080d6b3596aa16e8c8582f 100644 (file)
@@ -22,7 +22,7 @@ extern int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
                        unsigned char *vec);
 extern bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
                         unsigned long new_addr, unsigned long old_end,
-                        pmd_t *old_pmd, pmd_t *new_pmd);
+                        pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush);
 extern int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
                        unsigned long addr, pgprot_t newprot,
                        int prot_numa);
index cdcd25cb30fea3f2ad2c660e547d014b7378b3dd..eff3de359d50a30588abf70676dc6993c751471d 100644 (file)
@@ -1426,11 +1426,12 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 
 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
                  unsigned long new_addr, unsigned long old_end,
-                 pmd_t *old_pmd, pmd_t *new_pmd)
+                 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
 {
        spinlock_t *old_ptl, *new_ptl;
        pmd_t pmd;
        struct mm_struct *mm = vma->vm_mm;
+       bool force_flush = false;
 
        if ((old_addr & ~HPAGE_PMD_MASK) ||
            (new_addr & ~HPAGE_PMD_MASK) ||
@@ -1455,6 +1456,8 @@ bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
                new_ptl = pmd_lockptr(mm, new_pmd);
                if (new_ptl != old_ptl)
                        spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
+               if (pmd_present(*old_pmd) && pmd_dirty(*old_pmd))
+                       force_flush = true;
                pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
                VM_BUG_ON(!pmd_none(*new_pmd));
 
@@ -1467,6 +1470,10 @@ bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
                set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
                if (new_ptl != old_ptl)
                        spin_unlock(new_ptl);
+               if (force_flush)
+                       flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
+               else
+                       *need_flush = true;
                spin_unlock(old_ptl);
                return true;
        }
index da22ad2a5678265ea9f2d0aa5ece9e14c519a494..6ccecc03f56ad05940484a481c5a092f1ad245ef 100644 (file)
@@ -104,11 +104,13 @@ static pte_t move_soft_dirty_pte(pte_t pte)
 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
                unsigned long old_addr, unsigned long old_end,
                struct vm_area_struct *new_vma, pmd_t *new_pmd,
-               unsigned long new_addr, bool need_rmap_locks)
+               unsigned long new_addr, bool need_rmap_locks, bool *need_flush)
 {
        struct mm_struct *mm = vma->vm_mm;
        pte_t *old_pte, *new_pte, pte;
        spinlock_t *old_ptl, *new_ptl;
+       bool force_flush = false;
+       unsigned long len = old_end - old_addr;
 
        /*
         * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
@@ -146,6 +148,14 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
                                   new_pte++, new_addr += PAGE_SIZE) {
                if (pte_none(*old_pte))
                        continue;
+
+               /*
+                * 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);
                pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
                pte = move_soft_dirty_pte(pte);
@@ -156,6 +166,10 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
        if (new_ptl != old_ptl)
                spin_unlock(new_ptl);
        pte_unmap(new_pte - 1);
+       if (force_flush)
+               flush_tlb_range(vma, old_end - len, old_end);
+       else
+               *need_flush = true;
        pte_unmap_unlock(old_pte - 1, old_ptl);
        if (need_rmap_locks)
                drop_rmap_locks(vma);
@@ -201,13 +215,12 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
                                if (need_rmap_locks)
                                        take_rmap_locks(vma);
                                moved = move_huge_pmd(vma, old_addr, new_addr,
-                                                   old_end, old_pmd, new_pmd);
+                                                   old_end, old_pmd, new_pmd,
+                                                   &need_flush);
                                if (need_rmap_locks)
                                        drop_rmap_locks(vma);
-                               if (moved) {
-                                       need_flush = true;
+                               if (moved)
                                        continue;
-                               }
                        }
                        split_huge_pmd(vma, old_pmd, old_addr);
                        if (pmd_trans_unstable(old_pmd))
@@ -220,11 +233,10 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
                        extent = next - new_addr;
                if (extent > LATENCY_LIMIT)
                        extent = LATENCY_LIMIT;
-               move_ptes(vma, old_pmd, old_addr, old_addr + extent,
-                         new_vma, new_pmd, new_addr, need_rmap_locks);
-               need_flush = true;
+               move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
+                         new_pmd, new_addr, need_rmap_locks, &need_flush);
        }
-       if (likely(need_flush))
+       if (need_flush)
                flush_tlb_range(vma, old_end-len, old_addr);
 
        mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);